2019

• R. Ravichandran, “A mediator system for querying heterogeneous data in robotic applications,” Master Thesis, 41716,Hirschberger Str. 58 – 64, Bonn, 53119., 2019.
  [BibTeX] [Abstract]
  

  In robotic applications, sensor-generated data is often ignored after robots utilize the data for making decisions and sometimes save into persistent storage. Since these data are not adequately modeled and stored, it makes it hard for someone who wants to replay the experiments or to find faults in the sensor data. It is even more difficult to debug this massive amount of data if there are multi-robots involved in a task. It is likely that different vendors/developers develop multi-robots with different database instances and attribute names to save data which introduces heterogeneity in the sensor data. Heterogeneity includes additional problems like interoperability issues when sharing data between other robots and also with fault diagnosis tools. One way to overcome these issues is by employing a mediator component as a middle man for all robots and even for humans. In our approach, we designed a mediator architecture which solves integrating sensor data from different databases which are deployed on different robots. Also, the data modeling issue from the EU ROPOD project’s data logger system is addressed by creating an extendable generic data model for each critical entities in the robot system. Lastly, sensor observations interoperability issue is solved by adding meaningful contexts to all the entities, and it is achieved by using JSON-LD data representation. Overall mediator component is developed with GraphQL as a base framework and JSON-LD to represent the response data. This choice of GraphQL and JSON-LD provides further advantages to the system such as a single query language to fetch sensor data regardless of databases used in the robots and context-based data model.

  @MastersThesis{ 2019ravichandran,
  abstract  = {In robotic applications, sensor-generated data is often
  ignored after robots utilize the data for making decisions
  and sometimes save into persistent storage. Since these
  data are not adequately modeled and stored, it makes it
  hard for someone who wants to replay the experiments or to
  find faults in the sensor data. It is even more difficult
  to debug this massive amount of data if there are
  multi-robots involved in a task. It is likely that
  different vendors/developers develop multi-robots with
  different database instances and attribute names to save
  data which introduces heterogeneity in the sensor data.
  Heterogeneity includes additional problems like
  interoperability issues when sharing data between other
  robots and also with fault diagnosis tools. One way to
  overcome these issues is by employing a mediator component
  as a middle man for all robots and even for humans. In our
  approach, we designed a mediator architecture which solves
  integrating sensor data from different databases which are
  deployed on different robots. Also, the data modeling issue
  from the EU ROPOD project's data logger system is addressed
  by creating an extendable generic data model for each
  critical entities in the robot system. Lastly, sensor
  observations interoperability issue is solved by adding
  meaningful contexts to all the entities, and it is achieved
  by using JSON-LD data representation. Overall mediator
  component is developed with GraphQL as a base framework and
  JSON-LD to represent the response data. This choice of
  GraphQL and JSON-LD provides further advantages to the
  system such as a single query language to fetch sensor data
  regardless of databases used in the robots and
  context-based data model.},
  address  = {41716,Hirschberger Str. 58 - 64, Bonn, 53119. },
  annote  = {WS16/17 FH-BRS - A mediator system for querying
  heterogeneous data in robotic applications - Prof. Dr.
  Prassler, Prof. Dr. Kaul, Huebel and Blumenthal},
  author  = {Rubanraj Ravichandran},
  month = {April},
  school  = {Bonn-Rhein-Sieg University of Applied Sciences},
  title = {A mediator system for querying heterogeneous data in
  robotic applications},
  year = {2019}
  }

• A. Padalkar, “Compliant Manipulation with Reinforcement Learning Guided by Task Specification,” Master Thesis, Grantham-Allee 20, 53757 St. Augustin, Germany, 2019.
  [BibTeX] [Abstract]
  

  For effective integration of robots in the working environments along with humans, they need to possess the ability of compliant manipulation. Compliant manipulation can be achieved by modeling the environment and providing an engineered solution for specific tasks. But the limitations on the accuracy of models limit the ability of such approaches. Alternatively, reinforcement learning enables robots to learn compliant manipulation tasks on their own by interacting with the environment. But reinforcement learning needs a huge number of costly robot-environment interactions before learning any meaningful strategy to perform the task. A task can be learned in a simulated environment and then can be transferred to a real robot. In this case, the accuracy of simulation has a very big impact on the transferability of the solution. Furthermore, reinforcement learning can be aided with the task knowledge to minimize the interactions required for learning a particular task. In this master thesis, we designed, implemented and evaluated a compliant manipulation approach using reinforcement learning guided by task frame formalism, which is a task specification method. Developed solution allows us to model easy to model task knowledge using task frame formalism and then learn the unmodeled components in task specification using reinforcement learning. We evaluated the approach by performing two compliant manipulation tasks namely door opening and vegetable cutting with three different robots. Robots were successful in opening doors using task frame formalism alone. Vegetable cutting task was successfully learned with the help of reinforcement learning aided by task frame formalism. We were able to learn force control policy for vegetable cutting task using two different policy representations directly on the robot without using any simulation.

  @MastersThesis{ 2019padalkar,
  abstract  = {For effective integration of robots in the working
  environments along with humans, they need to possess the
  ability of compliant manipulation. Compliant manipulation
  can be achieved by modeling the environment and providing
  an engineered solution for specific tasks. But the
  limitations on the accuracy of models limit the ability of
  such approaches. Alternatively, reinforcement learning
  enables robots to learn compliant manipulation tasks on
  their own by interacting with the environment. But
  reinforcement learning needs a huge number of costly
  robot-environment interactions before learning any
  meaningful strategy to perform the task. A task can be
  learned in a simulated environment and then can be
  transferred to a real robot. In this case, the accuracy of
  simulation has a very big impact on the transferability of
  the solution. Furthermore, reinforcement learning can be
  aided with the task knowledge to minimize the interactions
  required for learning a particular task. In this master
  thesis, we designed, implemented and evaluated a compliant
  manipulation approach using reinforcement learning guided
  by task frame formalism, which is a task specification
  method. Developed solution allows us to model easy to model
  task knowledge using task frame formalism and then learn
  the unmodeled components in task specification using
  reinforcement learning. We evaluated the approach by
  performing two compliant manipulation tasks namely door
  opening and vegetable cutting with three different robots.
  Robots were successful in opening doors using task frame
  formalism alone. Vegetable cutting task was successfully
  learned with the help of reinforcement learning aided by
  task frame formalism. We were able to learn force control
  policy for vegetable cutting task using two different
  policy representations directly on the robot without using
  any simulation.},
  address  = {Grantham-Allee 20, 53757 St. Augustin, Germany},
  annote  = {SS17 Pl{\"o}ger, Jonas, Schneider supervising},
  author  = {Abhishek Padalkar},
  month = {November},
  school  = {Bonn-Rhein-Sieg University of Applied Sciences},
  title = {Compliant Manipulation with Reinforcement Learning Guided
  by Task Specification},
  year = {2019}
  }

• E. Ovchinnikova, “Automatic Sentence Generation Using Generative Adversarial Neural Networks,” Master Thesis, Grantham-Allee 20, 53757 St. Augustin, Germany, 2019.
  [BibTeX] [Abstract]
  

  This work aims to create a natural language generation (NLG) base for further development of systems for automatic examination questions generation and automatic summarization in Hochschule Bonn-Rhein-Sieg and Fraunhofer IAIS, respectively. Nowadays both tasks are very relevant. The first can significantly simplify the university teachers’ work and the second – to be of assistance for a faster retrieval of knowledge from an excessively large amount of information that people often work with. We focus on the search for an efficient and robust approach to the controlled NLG problem. Therefore, though the initial idea of the project was the usage of the generative adversarial neural networks (GANs), we switched our attention to more robust and easily-controllable autoencoders. Thus, in this work we implement an autoencoder for unsupervised discovery of latent space representations of text, and show the ability of the system to generate new sentences based on this latent space. Apart from that, we apply Gaussian mixture techniques in order to obtain meaningful text clusters and thereby try to create a tool that would allow us to generate sentences relevant to the semantics of the Gaussian clusters, e.g. positive or negative reviews or examination questions on certain topic. The developed system is tested on several datasets and compared to GANs’ performance.

  @MastersThesis{ 2019ovchinnikova,
  abstract  = {This work aims to create a natural language generation
  (NLG) base for further development of systems for automatic
  examination questions generation and automatic
  summarization in Hochschule Bonn-Rhein-Sieg and Fraunhofer
  IAIS, respectively. Nowadays both tasks are very relevant.
  The first can significantly simplify the university
  teachers’ work and the second – to be of assistance for
  a faster retrieval of knowledge from an excessively large
  amount of information that people often work with. We focus
  on the search for an efficient and robust approach to the
  controlled NLG problem. Therefore, though the initial idea
  of the project was the usage of the generative adversarial
  neural networks (GANs), we switched our attention to more
  robust and easily-controllable autoencoders. Thus, in this
  work we implement an autoencoder for unsupervised discovery
  of latent space representations of text, and show the
  ability of the system to generate new sentences based on
  this latent space. Apart from that, we apply Gaussian
  mixture techniques in order to obtain meaningful text
  clusters and thereby try to create a tool that would allow
  us to generate sentences relevant to the semantics of the
  Gaussian clusters, e.g. positive or negative reviews or
  examination questions on certain topic. The developed
  system is tested on several datasets and compared to
  GANs’ performance.},
  address  = {Grantham-Allee 20, 53757 St. Augustin, Germany},
  annote  = {[WS15/16] [HBRS] - [E-assessment] [Fraunhofer IAIS] -
  [Text summarization] [Pl{\"o}ger], [Kraetzschmar], [Ojeda],
  [Brito] supervising},
  author  = {Evgeniya Ovchinnikova},
  month = {March},
  school  = {Bonn-Rhein-Sieg University of Applied Sciences},
  title = {Automatic Sentence Generation Using Generative Adversarial
  Neural Networks},
  year = {2019}
  }

• R. Nalla, “Evaluation of Deep Learning Technique for Impulse Sound Classification in 2D Representation,” Master Thesis, Grantham-Allee 20, 53757 St. Augustin, Germany, 2019.
  [BibTeX] [Abstract]
  

  Impulse sound classification has several civil and security related applications.Convolutional Neural Networks(CNNs) have been proven very effective in image classification and show promise for audio applications as well. So, we would like to apply image classification networks from vision domain to audio domain for impulse sound classification task by treating audio segments as images. We use ballistic sound dataset collected by Fraunhofer Institute researchers for impulse sound classification task. For using audio segments as images we extract spectrograms of them and treat them as images for the training. We design a convolutional 2D network specifically for impulse sound classification task, and then create a convolutional 1D network with similar architecture. We compare the performance of these networks to see how beneficial it is to use spectrograms instead of audio raw data. We observe that the accuracy achieved using spectrogram data representation is 1.3% less than raw data representation but with a 40% gain in the number of parameters used, which can be considered as a good trade-off. Later, we compare the performance of this network with our selected image classifiers. From state-of-the-art image classification networks, VGGNet-16, Inception v3, Inception-ResNet-v2, NASNet, ResNeXt and Efficient Net are chosen as candidate architectures. Appling image data augmentation techniques to spectrograms doesnot make sense, so we apply spectrogram data augmentation techniques: time masking and frequency masking and then evaluate the effect of augmentation techniques on the network performance. The image classification networks selected are trained from scratch on the ballistic sound dataset and their performance is evaluated on different test sets. The same chosen classification networks are trained on ballistic sound dataset with ImageNet pretrained weights to study if the pretrained weights help the training. We compare and evaluate the performance of networks trained from scratch and networks trained with pretrained weights. ResNeXt-50 performs best with an averaged F1-score of 80 with 23.1M parameters. We also study the effect of unsupervised learning by using Autoencoders, on which the data is first trained to learn unsupervised features and use these Autoencoder weights as initial point for the classifier. But, we have seen that classifiers performance is not effected due to autoencoder weights. Index terms: Convolutional Neural Networks, Auto-encoders, data Augmentation, Audio classification.

  @MastersThesis{ 2019nalla,
  abstract  = {Impulse sound classification has several civil and
  security related applications.Convolutional Neural
  Networks(CNNs) have been proven very effective in image
  classification and show promise for audio applications as
  well. So, we would like to apply image classification
  networks from vision domain to audio domain for impulse
  sound classification task by treating audio segments as
  images. We use ballistic sound dataset collected by
  Fraunhofer Institute researchers for impulse sound
  classification task.
  For using audio segments as images we extract spectrograms
  of them and treat them as images for the training. We
  design a convolutional 2D network specifically for impulse
  sound classification task, and then create a convolutional
  1D network with similar architecture. We compare the
  performance of these networks to see how beneficial it is
  to use spectrograms instead of audio raw data. We observe
  that the accuracy achieved using spectrogram data
  representation is 1.3% less than raw data representation
  but with a 40% gain in the number of parameters used, which
  can be considered as a good trade-off. Later, we compare
  the performance of this network with our selected image
  classifiers. From state-of-the-art image classification
  networks, VGGNet-16, Inception v3, Inception-ResNet-v2,
  NASNet, ResNeXt and Efficient Net are chosen as candidate
  architectures. Appling image data augmentation techniques
  to spectrograms doesnot make sense, so we apply spectrogram
  data augmentation techniques: time masking and frequency
  masking and then evaluate the effect of augmentation
  techniques on the network performance. The image
  classification networks selected are trained from scratch
  on the ballistic sound dataset and their performance is
  evaluated on different test sets. The same chosen
  classification networks are trained on ballistic sound
  dataset with ImageNet pretrained weights to study if the
  pretrained weights help the training. We compare and
  evaluate the performance of networks trained from scratch
  and networks trained with pretrained weights. ResNeXt-50
  performs best with an averaged F1-score of 80 with 23.1M
  parameters.
  We also study the effect of unsupervised learning by using
  Autoencoders, on which the data is first trained to learn
  unsupervised features and use these Autoencoder weights as
  initial point for the classifier. But, we have seen that
  classifiers performance is not effected due to autoencoder
  weights.
  Index terms: Convolutional Neural Networks, Auto-encoders,
  data Augmentation, Audio classification.},
  address  = {Grantham-Allee 20, 53757 St. Augustin, Germany},
  annote  = {[WS16/17] [Fraunhofer FKIE] [Thiele], [Koch], [Oispuu]
  supervising},
  author  = {Ravali Nalla},
  month = {December},
  school  = {Bonn-Rhein-Sieg University of Applied Sciences},
  title = {Evaluation of Deep Learning Technique for Impulse Sound
  Classification in 2D Representation},
  year = {2019}
  }

• H. Talaat, “Robust Environmental Sound Classification & Anomaly Detection using Deep Learning,” Master Thesis, Grantham-Allee 20, 53757 St. Augustin, Germany, 2019.
  [BibTeX] [Abstract]
  

  In the world we live today, the robot manufacturers do not focus anymore only for industrial robots, the service robots which aids the human has also become part of their objectives. Having said that, for a robot to aid a human he has to understand the environment he is operating in. The environment’s perception can be performed by multiple sensory to cover various domains such as cameras, microphones, lidar, etc. With that being said, this project focuses on finding a way for the robot to understand the environmental sounds and detect anomaly behaviors through it. Classifying the audio data hash been mainly focused on speech and music tasks and the environmental sounds received the least attention, nevertheless, with the emerging of mobile robots technologies in the last decade has risen the interests of studying the environmental sound classification (ESC) tasks. Therefore, most of the algorithms implemented for speech and music were tested for environmental sounds and could not achieve the task due to the different kind of characteristics the audio data convey. The speech and music audio data is usually conveying stationary features which can be modeled, but the environmental sounds carry non-stationary characteristics. Fortunately, the CNN has proven to be superior with higher dimensional data and achieved great results with vision tasks which also been adopted to classify the environmental sounds. Another aspect in the decision making for anomaly sounds, the robot has to know the sound source location in order to have another parameter in the decision process. In this study a parallel CNN with different filter sizes is implemented and tested on the UrbanSounds 8K dataset to learn the environmental sounds, as well as the Respeaker 4 Mic, microphone array, used for sound source localization will be implemented to detect anomaly behavior in the environment. The system will be composed of a Client-Server modules and can be mounted onto any robot. The overall system achieved an accuracy of 70\% at 1meter distance from the robot, and for future work some improvements are proposed on the existing system to get higher accuracies.

  @MastersThesis{ 2019talaat,
  abstract  = {In the world we live today, the robot manufacturers do not
  focus anymore only for industrial robots, the service
  robots which aids the human has also become part of their
  objectives. Having said that, for a robot to aid a human he
  has to understand the environment he is operating in. The
  environment's perception can be performed by multiple
  sensory to cover various domains such as cameras,
  microphones, lidar, etc. With that being said, this project
  focuses on finding a way for the robot to understand the
  environmental sounds and detect anomaly behaviors through
  it. Classifying the audio data hash been mainly focused on
  speech and music tasks and the environmental sounds
  received the least attention, nevertheless, with the
  emerging of mobile robots technologies in the last decade
  has risen the interests of studying the environmental sound
  classification (ESC) tasks. Therefore, most of the
  algorithms implemented for speech and music were tested for
  environmental sounds and could not achieve the task due to
  the different kind of characteristics the audio data
  convey. The speech and music audio data is usually
  conveying stationary features which can be modeled, but the
  environmental sounds carry non-stationary characteristics.
  Fortunately, the CNN has proven to be superior with higher
  dimensional data and achieved great results with vision
  tasks which also been adopted to classify the environmental
  sounds. Another aspect in the decision making for anomaly
  sounds, the robot has to know the sound source location in
  order to have another parameter in the decision process. In
  this study a parallel CNN with different filter sizes is
  implemented and tested on the UrbanSounds 8K dataset to
  learn the environmental sounds, as well as the Respeaker 4
  Mic, microphone array, used for sound source localization
  will be implemented to detect anomaly behavior in the
  environment. The system will be composed of a Client-Server
  modules and can be mounted onto any robot. The overall
  system achieved an accuracy of 70\% at 1meter distance from
  the robot, and for future work some improvements are
  proposed on the existing system to get higher accuracies.},
  address  = {Grantham-Allee 20, 53757 St. Augustin, Germany},
  annote  = {WS15/16 FH-BRS Pl{\"o}ger, Breuer, Mitrevski supervising},
  author  = {Hashem Talaat},
  month = {March},
  school  = {Bonn-Rhein-Sieg University of Applied Sciences},
  title = {Robust Environmental Sound Classification & Anomaly
  Detection using Deep Learning},
  year = {2019}
  }

• I. Vishniakou, “Machine learning approaches for light scattering control,” Master Thesis, Grantham-Allee 20, 53757 St. Augustin, Germany, 2019.
  [BibTeX] [Abstract]
  

  Several wavefront shaping algorithms for light scattering control were implemented, both in simulations and in experiments. Some of these methods are established, such as the transmission matrix approach or iterative optimization methods. Others, specifically the machine learning approaches using single-layer linear and nonlinear neural networks, as well as convolutional neural networks for light control were novel developments in the context of this thesis. All these methods were evaluated and compared under similar conditions in terms of their ability to focus light through scattering media. For this purpose, software and hardware for automated experimental control was set up and optimized for short run-times and fast data acquisition. A novel approach for backscattering-based focusing through scattering materials based on variational autoencoders was suggested and implemented. It was shown that this model is suitable for finding statistical relationships between binary wavefront corrections linking transmitted and reflected of the sample light.

  @MastersThesis{ 2019vishniakou,
  abstract  = {Several wavefront shaping algorithms for light scattering
  control were implemented, both in simulations and in
  experiments. Some of these methods are established, such as
  the transmission matrix approach or iterative optimization
  methods. Others, specifically the machine learning
  approaches using single-layer linear and nonlinear neural
  networks, as well as convolutional neural networks for
  light control were novel developments in the context of
  this thesis. All these methods were evaluated and compared
  under similar conditions in terms of their ability to focus
  light through scattering media. For this purpose, software
  and hardware for automated experimental control was set up
  and optimized for short run-times and fast data
  acquisition. A novel approach for backscattering-based
  focusing through scattering materials based on variational
  autoencoders was suggested and implemented. It was shown
  that this model is suitable for finding statistical
  relationships between binary wavefront corrections linking
  transmitted and reflected of the sample light.},
  address  = {Grantham-Allee 20, 53757 St. Augustin, Germany},
  annote  = {WS14/15 HBRS, Center of advanced European Studies and
  Research (caesar) - Pl{\"o}ger, Asteroth, Seelig
  supervising},
  author  = {Ivan Vishniakou},
  month = {January},
  school  = {Bonn-Rhein-Sieg University of Applied Sciences},
  title = {Machine learning approaches for light scattering control},
  year = {2019}
  }

• N. Lu, Y. Wu, L. Feng, and J. Song, “Deep Learning for Fall Detection: 3D-CNN Combined with LSTM on Video Kinematic Data,” IEEE Journal of Biomedical and Health Informatics, vol. 23, p. 314 – 323, 2019.
  [BibTeX]

  @Article{ lu2018,
  author  = {Lu, Na and Wu, Yidan and Feng, Li and Song, Jinbo},
  journal  = {IEEE Journal of Biomedical and Health Informatics},
  title = {{Deep Learning for Fall Detection: 3D-CNN Combined with
  LSTM on Video Kinematic Data}},
  year = {2019},
  volume  = {23},
  pages = {314 -- 323}
  }

• M. Wasil, “Point Cloud Classification of Real Objects Using CNNs,” Master Thesis, Grantham-Allee 20, 53757 St. Augustin, Germany, 2019.
  [BibTeX] [Abstract]
  

  In autonomous systems, the ability to understand and interpret environment is essential. For instance, in robot manipulation, it is necessary for robots to understand which objects are present, known as object classification. Although image classification using convolutional neural networks (CNNs) has shown remarkable results, they have been mostly successful on Euclidean data structure or a grid-like structure. However, robots cannot rely only on images which lack geometrical structures. The geometrical information helps the robots not only to classify the objects having a similar shape, but also to differentiate between an object and an image of objects. This geometrical information can only be extracted from the 3D representation of objects. There have been many CNN-based 3D object classification methods using point clouds, but they were primarily evaluated on synthetic datasets. Therefore, in this work, we address the problem of point cloud classification using CNNs on real dataset. We use a modified Fisher vector (3DmFV) as a point cloud feature extraction and build deep neural networks using optimized inception modules. The main contribution of this work is a deep learning architecture for point cloud classification (3DmFV-Inception). Furthermore, an automatic point cloud dataset collection is also developed. We evaluated our method on three different datasets: our in-house PMD multiview, Washington RGB-D, and JHUIT-50 dataset, and compared the result against 3DmFV-Net and dynamic graph convolution (DGCNN). Our method achieved 93.21\%, 97.05\% and 94.39\% accuracy on PMD multiview, Washington RGB-D and JHUIT-50 dataset respectively. Moreover, our method outperformed 3DmFV-Net and DGCNN on the aforementioned dataset and had a 70\% smaller model size compared to that of 3DmFV-Net. Finally, the PCA alignment significantly improved the accuracy by 3.87\%.

  @MastersThesis{ 2019wasil,
  abstract  = {In autonomous systems, the ability to understand and
  interpret environment is essential. For instance, in robot
  manipulation, it is necessary for robots to understand
  which objects are present, known as object classification.
  Although image classification using convolutional neural
  networks (CNNs) has shown remarkable results, they have
  been mostly successful on Euclidean data structure or a
  grid-like structure. However, robots cannot rely only on
  images which lack geometrical structures. The geometrical
  information helps the robots not only to classify the
  objects having a similar shape, but also to differentiate
  between an object and an image of objects. This geometrical
  information can only be extracted from the 3D
  representation of objects.
  There have been many CNN-based 3D object classification
  methods using point clouds, but they were primarily
  evaluated on synthetic datasets. Therefore, in this work,
  we address the problem of point cloud classification using
  CNNs on real dataset. We use a modified Fisher vector
  (3DmFV) as a point cloud feature extraction and build deep
  neural networks using optimized inception modules. The main
  contribution of this work is a deep learning architecture
  for point cloud classification (3DmFV-Inception).
  Furthermore, an automatic point cloud dataset collection is
  also developed.
  We evaluated our method on three different datasets: our
  in-house PMD multiview, Washington RGB-D, and JHUIT-50
  dataset, and compared the result against 3DmFV-Net and
  dynamic graph convolution (DGCNN). Our method achieved
  93.21\%, 97.05\% and 94.39\% accuracy on PMD multiview,
  Washington RGB-D and JHUIT-50 dataset respectively.
  Moreover, our method outperformed 3DmFV-Net and DGCNN on
  the aforementioned dataset and had a 70\% smaller model
  size compared to that of 3DmFV-Net. Finally, the PCA
  alignment significantly improved the accuracy by 3.87\%.},
  address  = {Grantham-Allee 20, 53757 St. Augustin, Germany},
  annote  = {WS16/17 H-BRS - Pl{\"o}ger, Hinkenjann, Thoduka
  supervising},
  author  = {Mohammad Wasil},
  month = {July},
  school  = {Bonn-Rhein-Sieg University of Applied Sciences},
  title = {Point Cloud Classification of Real Objects Using CNNs},
  year = {2019}
  }

• D. Vukcevic, “Lazy Robot Control by Relaxation of Motion and Force Constraints,” Master Thesis, Grantham-Allee 21, 53757 St. Augustin, Germany, 2019.
  [BibTeX] [Abstract]
  

  Human and robot tasks in household environments include actions such as carrying an object, cleaning a surface, etc. These tasks are performed by means of dexterous manipulation, and for humans, they are straightforward to accomplish. Moreover, humans perform these actions with reasonable accuracy and precision but with much less energy and stress on the actuators (muscles) than the robots do. The high agility in controlling their forces and motions is actually due to “laziness”, i.e. humans exploit the existing natural forces and constraints to execute the tasks. The above-mentioned properties of the human lazy strategy motivate us to relax the problem of controlling robot motions and forces, and solve it with the help of the environment. Therefore, in this work, we developed a lazy control strategy, i.e. task specification models and control architectures that relax several aspects of robot control by exploiting prior knowledge about the task and environment. The developed control strategy is realized in four different robotics use cases. In this work, the Popov-Vereshchagin hybrid dynamics solver is used as one of the building blocks in the proposed control architectures. An extension of the solver’s interface with the virtual Cartesian force and feed-forward joint torque task-drivers is proposed in this thesis. To validate the proposed lazy control approach, an experimental evaluation was performed in a simulation environment and on a real robot platform.

  @MastersThesis{ 2019vukcevic,
  abstract  = {Human and robot tasks in household environments include
  actions such as carrying an object, cleaning a surface,
  etc. These tasks are performed by means of dexterous
  manipulation, and for humans, they are straightforward to
  accomplish. Moreover, humans perform these actions with
  reasonable accuracy and precision but with much less energy
  and stress on the actuators (muscles) than the robots do.
  The high agility in controlling their forces and motions is
  actually due to ``laziness'', i.e. humans exploit the
  existing natural forces and constraints to execute the
  tasks. The above-mentioned properties of the human lazy
  strategy motivate us to relax the problem of controlling
  robot motions and forces, and solve it with the help of the
  environment. Therefore, in this work, we developed a lazy
  control strategy, i.e. task specification models and
  control architectures that relax several aspects of robot
  control by exploiting prior knowledge about the task and
  environment. The developed control strategy is realized in
  four different robotics use cases. In this work, the
  Popov-Vereshchagin hybrid dynamics solver is used as one of
  the building blocks in the proposed control architectures.
  An extension of the solver's interface with the virtual
  Cartesian force and feed-forward joint torque task-drivers
  is proposed in this thesis. To validate the proposed lazy
  control approach, an experimental evaluation was performed
  in a simulation environment and on a real robot platform.},
  address  = {Grantham-Allee 21, 53757 St. Augustin, Germany},
  annote  = {WS16/17 Pl{\"o}ger, Bruyninckx, Schneider supervising},
  author  = {Djordje Vukcevic},
  month = {October},
  school  = {Bonn-Rhein-Sieg University of Applied Sciences},
  title = {Lazy Robot Control by Relaxation of Motion and Force
  Constraints},
  year = {2019}
  }

• L. Naik, “Semantic localization and navigation for indoor robots using OpenStreetMap,” Master Thesis, Grantham-Allee 20, 53757 St. Augustin, Germany, 2019.
  [BibTeX] [Abstract]
  

  A significant difference exists between how humans and robots navigate in indoor environments. Indoor robot navigation works at a geometric level in the form of geometric way-points to follow, while human navigation works at a much higher level of abstraction such as go straight for 100 m then turn left at the next junction and so on without relying on the geometric localization estimate. The goal of this research work is to enable a robot to navigate from one location to another location purely using semantic features present in the indoor environment, without depending on its geometric localization estimate. The robot navigation problem in the indoor environment is divided into environment-specific local navigation tasks. Thus the navigation problem is solved with the execution of these local navigation tasks and switching between them. One of the main reasons for the significant difference in human and robot navigation is the excellent perception capabilities of humans, which allow understanding the semantics of the environment. However, robots have limited perception capabilities, which makes it challenging for them to understand the environment the same way as humans. This work uses a prior map with rich semantic details to improve robots understanding of the environment. The main contribution of this work is a proof of concept of semantic localization and navigation for indoor robots using OpenStreetMap as a semantic map. The indoor environment is divided into different types, and semantic localization and local navigation tasks are defined for these environments. All the local navigation tasks use perception-based control, and semantic features are used for localization. Robot perception, the information provided by the route plan and semantic map are used to execute and switch between these local navigation tasks. Individual local navigation tasks were successfully tested on the ROPOD platform while the overall proposed navigation approach was tested in a Gazebo simulator.

  @MastersThesis{ 2019naik,
  abstract  = {A significant difference exists between how humans and
  robots navigate in indoor environments. Indoor robot
  navigation works at a geometric level in the form of
  geometric way-points to follow, while human navigation
  works at a much higher level of abstraction such as go
  straight for 100 m then turn left at the next junction and
  so on without relying on the geometric localization
  estimate. The goal of this research work is to enable a
  robot to navigate from one location to another location
  purely using semantic features present in the indoor
  environment, without depending on its geometric
  localization estimate. The robot navigation problem in the
  indoor environment is divided into environment-specific
  local navigation tasks. Thus the navigation problem is
  solved with the execution of these local navigation tasks
  and switching between them. One of the main reasons for the
  significant difference in human and robot navigation is the
  excellent perception capabilities of humans, which allow
  understanding the semantics of the environment. However,
  robots have limited perception capabilities, which makes it
  challenging for them to understand the environment the same
  way as humans. This work uses a prior map with rich
  semantic details to improve robots understanding of the
  environment. The main contribution of this work is a proof
  of concept of semantic localization and navigation for
  indoor robots using OpenStreetMap as a semantic map. The
  indoor environment is divided into different types, and
  semantic localization and local navigation tasks are
  defined for these environments. All the local navigation
  tasks use perception-based control, and semantic features
  are used for localization. Robot perception, the
  information provided by the route plan and semantic map are
  used to execute and switch between these local navigation
  tasks. Individual local navigation tasks were successfully
  tested on the ROPOD platform while the overall proposed
  navigation approach was tested in a Gazebo simulator.},
  address  = {Grantham-Allee 20, 53757 St. Augustin, Germany},
  annote  = {SS17 HBRS - ROPOD Prassler, Bruyninckx, Huebel
  supervising},
  author  = {Lakshadeep Naik},
  month = {June},
  school  = {Bonn-Rhein-Sieg University of Applied Sciences},
  title = {Semantic localization and navigation for indoor robots
  using OpenStreetMap},
  year = {2019}
  }

• P. Nagel, “A Framework for Robot Development using Computer-Aided Test Design and Reporting,” Master Thesis, 2019.
  [BibTeX] [Abstract]
  

  With the EU programme, Horizon 2020, the European Commission introduces a metric to asses the level of a technology for robot development projects. The scale is used as a control mechanism that allows to follow the progress and helps to make decisions about further investments. Project partners are obligated to verify the quality and functionalities of their robotic systems. With this in mind, quality assurance techniques, testing and reporting become even more important for the development process. Although tools exist that address the aspects, no single solution can be found that incorporates all the techniques. The consequences are, that coherent elements are maintained decentralized resulting in complexity and additional labor for the developers. This master thesis aims to elaborate a framework for robot development with computer-aided test design and reporting. It includes existing approaches, tools and requirements of ongoing projects into the definition process. As a second step the framework is implemented in an application and finally it is applied to three use cases and the utility of the framework analyzed. The preliminary results of using the application in the three different use cases indicate its usefulness for documenting the development and testing process, thereby improving the communication and allowing a smooth knowledge exchange process between different parties involved in the projects.

  @MastersThesis{ 2019nagelpatrick,
  abstract  = {With the EU programme, Horizon 2020, the European
  Commission introduces a metric to asses the level of a
  technology for robot development projects. The scale is
  used as a control mechanism that allows to follow the
  progress and helps to make decisions about further
  investments. Project partners are obligated to verify the
  quality and functionalities of their robotic systems. With
  this in mind, quality assurance techniques, testing and
  reporting become even more important for the development
  process. Although tools exist that address the aspects, no
  single solution can be found that incorporates all the
  techniques. The consequences are, that coherent elements
  are maintained decentralized resulting in complexity and
  additional labor for the developers.
  This master thesis aims to elaborate a framework for robot
  development with computer-aided test design and reporting.
  It includes existing approaches, tools and requirements of
  ongoing projects into the definition process. As a second
  step the framework is implemented in an application and
  finally it is applied to three use cases and the utility of
  the framework analyzed.
  The preliminary results of using the application in the
  three different use cases indicate its usefulness for
  documenting the development and testing process, thereby
  improving the communication and allowing a smooth knowledge
  exchange process between different parties involved in the
  projects. },
  annote  = {SS2016 Prassler, Pl{\"o}ger, Blumenthal supervising},
  author  = {Patrick Nagel},
  month = {March},
  school  = {Bonn-Rhein-Sieg University of Applied Sciences},
  title = {A Framework for Robot Development using Computer-Aided
  Test Design and Reporting},
  year = {2019}
  }

• W. Hsu, “Automated gait and pose analysis for lameness detection in cows,” Master Thesis, Grantham-Allee 20, 53757 St. Augustin, Germany, 2019.
  [BibTeX] [Abstract]
  

  Lameness is a serious disorder in dairy farms that increases the risk of culling of cattle as well as economic losses. Locomotion scoring is an assessment tool for monitoring and managing herd lameness levels. While manual scoring is a common and easy way, many researchers proposed various methods of automatic lameness detection using computer vision techniques. With the popularity and power of deep learning, this thesis aims to evaluate the efficacy of deep neural network for lameness detection. To analyze videos that consist of spatio-temporal information, recurrent neural networks were used. More specifically, a hierarchical recurrent neural network was used to classify the lameness of cows. The network is based on the idea that behaviors are dependent on the movements of individual body parts and their combinations. The network architecture had nine layers. The first layer was composed of five subnets, each accepted a body part as input. Pose estimation was first applied to extract sequences of keypoints from videos of walking cows. The keypoints were arranged into five body parts and fed into separate subnets of the model. By hierarchically fusing the features of body parts, the model could analyze the cow’s behaviors. The last layer predicted the most likely class of lameness level. The approach achieved an accuracy of 75\% in classifying non-lame and lame cows under 5-fold cross validation. The network was slightly adjusted to predict lameness as a four-class classification task, and it showed the correct trend of classifying the level of lameness. Several variables were investigated, and it turned out that the sequence length is trivial, as long as it has at least 50 frames. The body part of trunk may played a relatively important role in lameness detection, but just to a slight extent. The method and data quality and other elements of the approach can be refined to improve the overall performance.

  @MastersThesis{ 2019hsu,
  abstract  = {Lameness is a serious disorder in dairy farms that
  increases the risk of culling of cattle as well as economic
  losses. Locomotion scoring is an assessment tool for
  monitoring and managing herd lameness levels. While manual
  scoring is a common and easy way, many researchers proposed
  various methods of automatic lameness detection using
  computer vision techniques. With the popularity and power
  of deep learning, this thesis aims to evaluate the efficacy
  of deep neural network for lameness detection. To analyze
  videos that consist of spatio-temporal information,
  recurrent neural networks were used. More specifically, a
  hierarchical recurrent neural network was used to classify
  the lameness of cows. The network is based on the idea that
  behaviors are dependent on the movements of individual body
  parts and their combinations. The network architecture had
  nine layers. The first layer was composed of five subnets,
  each accepted a body part as input. Pose estimation was
  first applied to extract sequences of keypoints from videos
  of walking cows. The keypoints were arranged into five body
  parts and fed into separate subnets of the model. By
  hierarchically fusing the features of body parts, the model
  could analyze the cow's behaviors. The last layer predicted
  the most likely class of lameness level. The approach
  achieved an accuracy of 75\% in classifying non-lame and
  lame cows under 5-fold cross validation. The network was
  slightly adjusted to predict lameness as a four-class
  classification task, and it showed the correct trend of
  classifying the level of lameness. Several variables were
  investigated, and it turned out that the sequence length is
  trivial, as long as it has at least 50 frames. The body
  part of trunk may played a relatively important role in
  lameness detection, but just to a slight extent. The method
  and data quality and other elements of the approach can be
  refined to improve the overall performance.},
  address  = {Grantham-Allee 20, 53757 St. Augustin, Germany},
  annote  = {WS16/17 FH-BRS - [Project Name] Herpers, Koch, Schleiss},
  author  = {Wei-Chan Hsu},
  month = {October},
  school  = {Bonn-Rhein-Sieg University of Applied Sciences},
  title = {Automated gait and pose analysis for lameness detection in
  cows},
  year = {2019}
  }

• B. Ha, “Generalization of SELU to CNN,” Master Thesis, Grantham-Allee 20, 53757 St. Augustin, Germany, 2019.
  [BibTeX] [Abstract]
  

  Neural network based object detectors are able to automatize many difficult, tedious tasks. However, they are usually slow and/or require powerful hardware. One main reason is called Batch Normalization (BN) [1], which is an important method for building these detectors. Recent studies present a potential replacement called Self-normalizing Neural Network (SNN) [2], which at its core is a special activation function named Scaled Exponential Linear Unit (SELU). This replacement seems to have most of BN’s benefits while requiring less computational power. Nonetheless, it is uncertain that SELU and neural network based detectors are compatible with one another. An evaluation of SELU incorporated networks would help clarify that uncertainty. Such evaluation is performed through series of tests on different neural networks. After the evaluation, it is concluded that, while indeed faster, SELU is still not as good as BN for building complex object detector networks.

  @MastersThesis{ 2019ha,
  abstract  = {Neural network based object detectors are able to
  automatize many difficult, tedious tasks. However, they are
  usually slow and/or require powerful hardware. One main
  reason is called Batch Normalization (BN) [1], which is an
  important method for building these detectors. Recent
  studies present a potential replacement called
  Self-normalizing Neural Network (SNN) [2], which at its
  core is a special activation function named Scaled
  Exponential Linear Unit (SELU). This replacement seems to
  have most of BN’s benefits while requiring less
  computational power. Nonetheless, it is uncertain that SELU
  and neural network based detectors are compatible with one
  another. An evaluation of SELU incorporated networks would
  help clarify that uncertainty. Such evaluation is performed
  through series of tests on different neural networks. After
  the evaluation, it is concluded that, while indeed faster,
  SELU is still not as good as BN for building complex object
  detector networks. },
  address  = {Grantham-Allee 20, 53757 St. Augustin, Germany},
  annote  = {WS15/16 FH-BRS Pl{\"o}ger, Kraetzschmar, Zimmermann
  supervising},
  author  = {Bach Ha},
  month = {January},
  school  = {Bonn-Rhein-Sieg University of Applied Sciences},
  title = {Generalization of SELU to CNN},
  year = {2019}
  }

• T. Bontzorlos, “An Evaluation of Deep Learning Object Detection Pipelines for Maritime Application Purposes,” Master Thesis, Grantham-Allee 20, 53757 St. Augustin, Germany, 2019.
  [BibTeX] [Abstract]
  

  Maritime is an important sector of modern life and many goods are transported through sea. As a result, there is intense research on Unmanned Surface Vehicle (USV). Autonomous or semi-autonomous ships will decrease the need for personnel on-board, reducing the cost and freeing living space for more cargo. Detection of nearby objects is an essential part of the required autonomy. In more detail, object detection using an optical sensor is a necessary step, since Radar systems can only detect metal objects and at a distance of minimum 1 kilometer. Therefore, an optical sensor can provide information on non-metal maritime objects and objects closer than 1 kilometer. Furthermore, a usual — and expensive — threat in maritime is piracy. Pirates attack vessels using plastic boats, which renders their detection through Radar impossible. An on-board object detection system using optical sensors can provide the crew with early warning of potential threats. Moreover, surveillance of sea borders using USV could benefit of improvements in object detection. There has been intense research on object detection for maritime purposes. However, in the past years, there has been a breakthrough in object detection through Deep Learning (DL) techniques. These techniques provide accurate and fast detection of objects and have been widely employed in autonomous driving projects. Although there are proposals of object detection using DL in the maritime domain, there is currently limited work on comparison of how different DL architectures perform in a maritime environment. In this work, different DL architectures using various feature extractors have been compared on a novel maritime dataset, the Singapore Maritime Dataset (SMD). The metrics used for evaluation are the mean Average Precision (mAP), the speed of inference and the maximum usage of the Graphic Processing Unit (GPU) memory. The results have shown that the Faster R-CNN architecture using Inception v2 for feature extraction has achieved the highest mAP of 76%, although it is not able to run in real-time. On the other hand, SSD architecture has maintained a good ratio of speed and mAP, with up to 40 Frames per Second (FPS) inference speed and up to 65% mAP depending on the feature extractor used. Furthermore, its GPU memory usage has been significantly lower than the other architectures and as low as just 2GB. The results acquired using the SMD have been compared and qualitatively validated with the results of a similar work on the — also novel — SeaShips dataset.

  @MastersThesis{ 2019bontzorlos,
  abstract  = {Maritime is an important sector of modern life and many
  goods are transported through sea. As a result, there is
  intense research on Unmanned Surface Vehicle (USV).
  Autonomous or semi-autonomous ships will decrease the need
  for personnel on-board, reducing the cost and freeing
  living space for more cargo. Detection of nearby objects is
  an essential part of the required autonomy. In more detail,
  object detection using an optical sensor is a necessary
  step, since Radar systems can only detect metal objects and
  at a distance of minimum 1 kilometer. Therefore, an optical
  sensor can provide information on non-metal maritime
  objects and objects closer than 1 kilometer. Furthermore, a
  usual — and expensive — threat in maritime is piracy.
  Pirates attack vessels using plastic boats, which renders
  their detection through Radar impossible. An on-board
  object detection system using optical sensors can provide
  the crew with early warning of potential threats. Moreover,
  surveillance of sea borders using USV could benefit of
  improvements in object detection. There has been intense
  research on object detection for maritime purposes.
  However, in the past years, there has been a breakthrough
  in object detection through Deep Learning (DL) techniques.
  These techniques provide accurate and fast detection of
  objects and have been widely employed in autonomous driving
  projects. Although there are proposals of object detection
  using DL in the maritime domain, there is currently limited
  work on comparison of how different DL architectures
  perform in a maritime environment. In this work, different
  DL architectures using various feature extractors have been
  compared on a novel maritime dataset, the Singapore
  Maritime Dataset (SMD). The metrics used for evaluation are
  the mean Average Precision (mAP), the speed of inference
  and the maximum usage of the Graphic Processing Unit (GPU)
  memory. The results have shown that the Faster R-CNN
  architecture using Inception v2 for feature extraction has
  achieved the highest mAP of 76%, although it is not able to
  run in real-time. On the other hand, SSD architecture has
  maintained a good ratio of speed and mAP, with up to 40
  Frames per Second (FPS) inference speed and up to 65% mAP
  depending on the feature extractor used. Furthermore, its
  GPU memory usage has been significantly lower than the
  other architectures and as low as just 2GB. The results
  acquired using the SMD have been compared and qualitatively
  validated with the results of a similar work on the —
  also novel — SeaShips dataset.},
  address  = {Grantham-Allee 20, 53757 St. Augustin, Germany},
  annote  = {SS15 H-BRS Ploeger, Indiveri, Valdenegro supervising},
  author  = {Tilemachos Bontzorlos},
  month = {May},
  school  = {Bonn-Rhein-Sieg University of Applied Sciences},
  title = {An Evaluation of Deep Learning Object Detection Pipelines
  for Maritime Application Purposes},
  year = {2019}
  }

• R. Bhat, “Application of Model Interpretability Techniques for Short Answer Grading,” Master Thesis, Grantham-Allee 20, 53757 St. Augustin, Germany, 2019.
  [BibTeX] [Abstract]
  

  Automated scoring of short answer type questions via Natural Language Processing (NLP) is a field that has a substantial body of associated research in order to develop methods to reduce workload of manual graders. However, procedures that associates a reasoning behind the automated score to enhance the reliability of the assigned grade is a task that has received little attention in literature. This work focuses on applying and evaluating different strategies of model interpretability to the task of automatic short answer grading as a means of rationalizing the predicted score to the human graders. Pre investigations were carried out on two state-of-the-art model-agnostic interpretation techniques, namely Local interpretable model-agnostic explanations (LIME) and SHapley Additive exPlanations (SHAP). The influence of change in feature extractors and machine learning models on LIME was analyzed, and a quantitative comparison of the interpretation levels from LIME and SHAP was done for different case studies. The results showed that LIME and SHAP tend to give better explanations when answers are restricted to one sentence length, and do not have a large linguistic diversity. Furthermore, LIME’s restriction of highlighting only unigram keywords as part of the explanation, and SHAP’s exponential time requirement to compute feature attributions, made them unsuitable for our application. As a means of overcoming these challenges, we develop a solution that is model agnostic, is capable of highlights phrases (bigrams and trigrams) in the explanation, and improves the autograding and generation of explanations for new data by involving a human oracle both during initial annotation of data, and later in providing feedback for the predicted scores and explanations. A prototype web based GUI tool was developed to integrate the proposed solution. Tests run with the proposed solution showed a significant improvement in the level of interpretation of the highlighted explanation when compared against the pre investigation results for LIME and SHAP.

  @MastersThesis{ 2019bhat,
  abstract  = {Automated scoring of short answer type questions via
  Natural Language Processing (NLP) is a field that has a
  substantial body of associated research in order to develop
  methods to reduce workload of manual graders. However,
  procedures that associates a reasoning behind the automated
  score to enhance the reliability of the assigned grade is a
  task that has received little attention in literature. This
  work focuses on applying and evaluating different
  strategies of model interpretability to the task of
  automatic short answer grading as a means of rationalizing
  the predicted score to the human graders. Pre
  investigations were carried out on two state-of-the-art
  model-agnostic interpretation techniques, namely Local
  interpretable model-agnostic explanations (LIME) and
  SHapley Additive exPlanations (SHAP). The influence of
  change in feature extractors and machine learning models on
  LIME was analyzed, and a quantitative comparison of the
  interpretation levels from LIME and SHAP was done for
  different case studies. The results showed that LIME and
  SHAP tend to give better explanations when answers are
  restricted to one sentence length, and do not have a large
  linguistic diversity. Furthermore, LIME's restriction of
  highlighting only unigram keywords as part of the
  explanation, and SHAP's exponential time requirement to
  compute feature attributions, made them unsuitable for our
  application. As a means of overcoming these challenges, we
  develop a solution that is model agnostic, is capable of
  highlights phrases (bigrams and trigrams) in the
  explanation, and improves the autograding and generation of
  explanations for new data by involving a human oracle both
  during initial annotation of data, and later in providing
  feedback for the predicted scores and explanations. A
  prototype web based GUI tool was developed to integrate the
  proposed solution. Tests run with the proposed solution
  showed a significant improvement in the level of
  interpretation of the highlighted explanation when compared
  against the pre investigation results for LIME and SHAP.},
  address  = {Grantham-Allee 20, 53757 St. Augustin, Germany},
  annote  = {WS16/17 HBRS - ASAG Pl{\"o}ger, Becker, Nair supervising},
  author  = {Ravikiran Bhat},
  month = {March},
  school  = {Bonn-Rhein-Sieg University of Applied Sciences},
  title = {Application of Model Interpretability Techniques for Short
  Answer Grading},
  year = {2019}
  }

• T. Jandt, “Multiagent System for Warehouse Material Flow Control,” Master Thesis, Grantham-Allee 20, 53757 St. Augustin, Germany, 2019.
  [BibTeX] [Abstract]
  

  Increasing sales volumes and customers demanding just-in-time and flexible deliveries result in higher requirements and complexity for logistics systems. Additional challenges are the continually changing conditions in supply chain management. Innovative warehouse management is an important part and is responsible for coordinating and synchronizing activities to improve performance. In this work, the authors propose an approach for warehouse material flow control based on a multiagent system (MAS). A MAS consists of several intelligent agents. The main idea behind the method described here is that the individual agents act autonomously and self-interested. Such an agent would interact with other agents (e.g. conveyor belt, intersection, vehicles) to achieve its desired goal. The developed architecture relies on every single transport unit (TPU) inside the warehouse being an agent. Furthermore, additional agents are responsible for locations, to which TPUs can move to. Using the contract net protocol (CNP), the two types of agents negotiate the transport between locations and vehicles. The TPU agents follow their own objectives, depending on the overall goal of the warehouse. Examples are the delivery of orders or storing of incoming goods. There is no overall control. Each agent acts on its own, given an objective. The whole approach was implemented using UniWare provided by the Unitechnik Systems GmbH. Interfaces for all main entities present in the warehouse of the Netherland scenario were developed, and agents with appropriate behaviors created. The scenario features the delivery of orders for outbound TPUs and storing of inbound goods in two different classes of storage. Various vehicles are included as well. Evaluation is done by comparing the performance of the MAS approach to UniWare using four different benchmarks in simulation. In total, the simulated runtime of the warehouse easily exceeds half a year of operation per technique. Unfortunately, the MAS style performs worse than UniWare in all benchmarks and needs further optimizations regarding the individual strategies. However, the method presented here still managed to fully operate the warehouse without any failures and features essential properties, such as flexibility, dynamic event handling, and scalability. By using the MAS approach, the logic of material flow is encapsulated within every agent’s behavior. Thus allowing for better maintainability and portability.

  @MastersThesis{ 2019jandt,
  author  = {Torsten Jandt},
  title = {Multiagent System for Warehouse Material Flow Control},
  school  = {Bonn-Rhein-Sieg University of Applied Sciences},
  year = {2019},
  address  = {Grantham-Allee 20, 53757 St. Augustin, Germany},
  month = {July},
  abstract  = {Increasing sales volumes and customers demanding
  just-in-time and flexible deliveries result in higher
  requirements and complexity for logistics systems.
  Additional challenges are the continually changing
  conditions in supply chain management. Innovative warehouse
  management is an important part and is responsible for
  coordinating and synchronizing activities to improve
  performance.
  In this work, the authors propose an approach for warehouse
  material flow control based on a multiagent system (MAS). A
  MAS consists of several intelligent agents. The main idea
  behind the method described here is that the individual
  agents act autonomously and self-interested. Such an agent
  would interact with other agents (e.g. conveyor belt,
  intersection, vehicles) to achieve its desired goal.
  The developed architecture relies on every single transport
  unit (TPU) inside the warehouse being an agent.
  Furthermore, additional agents are responsible for
  locations, to which TPUs can move to. Using the contract
  net protocol (CNP), the two types of agents negotiate the
  transport between locations and vehicles. The TPU agents
  follow their own objectives, depending on the overall goal
  of the warehouse. Examples are the delivery of orders or
  storing of incoming goods. There is no overall control.
  Each agent acts on its own, given an objective.
  The whole approach was implemented using UniWare provided
  by the Unitechnik Systems GmbH. Interfaces for all main
  entities present in the warehouse of the Netherland
  scenario were developed, and agents with appropriate
  behaviors created. The scenario features the delivery of
  orders for outbound TPUs and storing of inbound goods in
  two different classes of storage. Various vehicles are
  included as well.
  Evaluation is done by comparing the performance of the MAS
  approach to UniWare using four different benchmarks in
  simulation. In total, the simulated runtime of the
  warehouse easily exceeds half a year of operation per
  technique. Unfortunately, the MAS style performs worse than
  UniWare in all benchmarks and needs further optimizations
  regarding the individual strategies. However, the method
  presented here still managed to fully operate the warehouse
  without any failures and features essential properties,
  such as flexibility, dynamic event handling, and
  scalability. By using the MAS approach, the logic of
  material flow is encapsulated within every agent's
  behavior. Thus allowing for better maintainability and
  portability.},
  annote  = {WS16/17 Unitechnik Systems GmbH Kraetzschmar, Asteroth,
  Blauel supervising}
  }

• R. Kumar, “Evidence Extraction for Fact Validation using Neural Network Architectures,” Master Thesis, Grantham-Allee 20, 53757 St. Augustin, Germany, 2019.
  [BibTeX] [Abstract]
  

  With the increase of misleading information on the internet, fact validation is the prominent and difficult task that helps the people to determine the veracity of the facts by verifying through the reliable knowledge sources such as Wikipedia. To accomplish this, generally, three steps are considered: Document retrieval, Evidence retrieval, and Claim classification. Recently largest fact validation datasets known as Fact Extraction and VERification (FEVER) original and FEVER simple claims were introduced to benchmark the state of the art approaches, in which system has to assign the label as well as retrieve evidences related to the claim/fact. In this work, we explored the classical and deep learning sentence retrieval and claim classification approaches to enhance the performance of DefactoNLP \cite{DeFactoNLP} (proposed to solve this task) for FEVER original dataset. Currently, DefactoNLP employs Term Frequency – Inverse Document Frequency (TF-IDF) for evidence retrieval and hand-crafted features along with the textual-entailment model to assign the final label to the claim. The main drawback of TF-IDF is, it relies on string similarity and does not account for synonyms. Furthermore, we also evaluate different claim classification approaches on FEVER simple claims dataset and compare the results to the state of the art. We evaluated sentence retrieval and claim classical approaches on FEVER original and FEVER simple claims dataset. As a result, we observed that simple LSTM using BERT outperforms state of the art on FEVER simple claim dataset. While, on FEVER original dataset, after integration with DefactoNLP, we obtained comparable performance to the state of the art.

  @MastersThesis{ 2019kumar,
  abstract  = {With the increase of misleading information on the
  internet, fact validation is the prominent and difficult
  task that helps the people to determine the veracity of the
  facts by verifying through the reliable knowledge sources
  such as Wikipedia. To accomplish this, generally, three
  steps are considered: Document retrieval, Evidence
  retrieval, and Claim classification. Recently largest fact
  validation datasets known as Fact Extraction and
  VERification (FEVER) original and FEVER simple claims were
  introduced to benchmark the state of the art approaches, in
  which system has to assign the label as well as retrieve
  evidences related to the claim/fact.
  In this work, we explored the classical and deep learning
  sentence retrieval and claim classification approaches to
  enhance the performance of DefactoNLP \cite{DeFactoNLP}
  (proposed to solve this task) for FEVER original dataset.
  Currently, DefactoNLP employs Term Frequency - Inverse
  Document Frequency (TF-IDF) for evidence retrieval and
  hand-crafted features along with the textual-entailment
  model to assign the final label to the claim. The main
  drawback of TF-IDF is, it relies on string similarity and
  does not account for synonyms. Furthermore, we also
  evaluate different claim classification approaches on FEVER
  simple claims dataset and compare the results to the state
  of the art.
  We evaluated sentence retrieval and claim classical
  approaches on FEVER original and FEVER simple claims
  dataset. As a result, we observed that simple LSTM using
  BERT outperforms state of the art on FEVER simple claim
  dataset. While, on FEVER original dataset, after
  integration with DefactoNLP, we obtained comparable
  performance to the state of the art. },
  address  = {Grantham-Allee 20, 53757 St. Augustin, Germany},
  annote  = {[WS/2016] [Project Affiliation] - [DEFACTO] [Ploeger],
  [Lehmann], [Esteves] supervising},
  author  = {Ramesh Kumar},
  month = {09},
  school  = {Bonn-Rhein-Sieg University of Applied Sciences},
  title = {Evidence Extraction for Fact Validation using Neural
  Network Architectures},
  year = {2019}
  }

• T. Metzler, “Computer-assisted grading of short answers using word embeddings and keyphrase extraction,” Master Thesis, Grantham-Allee 20, 53757 St. Augustin, Germany, 2019.
  [BibTeX] [Abstract]
  

  This thesis report investigates how an automatic short answer grading system (ASAG) can be improved by calculating the similarity on a phrase level.\\ A modular grading pipeline consisting of processing of the answers, extracting phrases built around verbs, generating feature vectors based on word embeddings and producing an output on the grading scale using ordinal regression is presented. The effect of processing and using different word embeddings, such as GloVe, FastText or Word2Vec, is evaluated on two datasets from the domain of computer science and machine learning using two common similarity metrics, the sum of word embeddings (SOWE) and SIM. In SIM the answers are compared by finding the most similar word in the student answer for each word in the reference answer, calculating the sum and normalizing over the length of the reference answer. It is shown that finding the best processing pipeline can boost the performance of the system across all models, similarity metrics and datasets. Across all models and datasets the SIM metric performs best.

  @MastersThesis{ 2019metzler,
  abstract  = {This thesis report investigates how an automatic short
  answer grading system (ASAG) can be improved by calculating
  the similarity on a phrase level.\\ A modular grading
  pipeline consisting of processing of the answers,
  extracting phrases built around verbs, generating feature
  vectors based on word embeddings and producing an output on
  the grading scale using ordinal regression is presented.
  The effect of processing and using different word
  embeddings, such as GloVe, FastText or Word2Vec, is
  evaluated on two datasets from the domain of computer
  science and machine learning using two common similarity
  metrics, the sum of word embeddings (SOWE) and SIM. In SIM
  the answers are compared by finding the most similar word
  in the student answer for each word in the reference
  answer, calculating the sum and normalizing over the length
  of the reference answer.
  It is shown that finding the best processing pipeline can
  boost the performance of the system across all models,
  similarity metrics and datasets. Across all models and
  datasets the SIM metric performs best.},
  address  = {Grantham-Allee 20, 53757 St. Augustin, Germany},
  annote  = {SS16 HBRS - E-Grading Pl{\"o}ger, Kraetzschmar
  supervising},
  author  = {Tim Metzler},
  month = {April},
  school  = {Bonn-Rhein-Sieg University of Applied Sciences},
  title = {Computer-assisted grading of short answers using word
  embeddings and keyphrase extraction},
  year = {2019}
  }

• A. Mallick, “Sonar Patch Matching via Deep Learning,” Master Thesis, Grantham-Allee 20, 53757 St. Augustin, Germany, 2019.
  [BibTeX] [Abstract]
  

  Application of underwater robots are on the rise, most of them are dependent on sonar for underwater vision. Though sonar is more effective than optical vision in difficult underwater scenarios, acoustic images have a low signal-to-noise ratio, low resolution and therefore harder to model accurately. Recently in Valdenegro-Toro [52], instead of modeling features manually, a Convolutional Neural Network (CNN) is encoded to learn general similarity function and predict whether two input sonar images are similar or not. This work presents a method that gives a prediction accuracy for sonar image matching functionality, which is instrumental in aquatic applications like localization, mapping, even in some object detection/recognition. With the objective of improving the sonar image matching problem further, three state-of-the-art CNN architectures are evaluated on the dataset from [52], containing about 39K pairs of training data, and another 8K for testing. To ensure a fair evaluation of each network, thorough hyperparameter optimization is executed. Using DenseNet Two-Channel network the obtained average prediction accuracy is 0.955 area under ROC curve (AUC). VGG-Siamese (with Contrastive loss function) and DenseNet-Siamese perform the prediction with an average AUC of 0.949 and 0.921 respectively. All these results are an improvement over the result (0.894 AUC) from Valdenegro-Toro [52]. By encoding an ensemble of DenseNet two-channel and DenseNet-Siamese models, with the respective highest AUC scores, the obtained prediction accuracy is 0.978 AUC, which is the overall highest AUC in the scope of this thesis.

  @MastersThesis{ 2019mallick,
  abstract  = {Application of underwater robots are on the rise, most of
  them are dependent on sonar for underwater vision. Though
  sonar is more effective than optical vision in difficult
  underwater scenarios, acoustic images have a low
  signal-to-noise ratio, low resolution and therefore harder
  to model accurately. Recently in Valdenegro-Toro [52],
  instead of modeling features manually, a Convolutional
  Neural Network (CNN) is encoded to learn general similarity
  function and predict whether two input sonar images are
  similar or not. This work presents a method that gives a
  prediction accuracy for sonar image matching functionality,
  which is instrumental in aquatic applications like
  localization, mapping, even in some object
  detection/recognition. With the objective of improving the
  sonar image matching problem further, three
  state-of-the-art CNN architectures are evaluated on the
  dataset from [52], containing about 39K pairs of training
  data, and another 8K for testing. To ensure a fair
  evaluation of each network, thorough hyperparameter
  optimization is executed. Using DenseNet Two-Channel
  network the obtained average prediction accuracy is 0.955
  area under ROC curve (AUC). VGG-Siamese (with Contrastive
  loss function) and DenseNet-Siamese perform the prediction
  with an average AUC of 0.949 and 0.921 respectively. All
  these results are an improvement over the result (0.894
  AUC) from Valdenegro-Toro [52]. By encoding an ensemble of
  DenseNet two-channel and DenseNet-Siamese models, with the
  respective highest AUC scores, the obtained prediction
  accuracy is 0.978 AUC, which is the overall highest AUC in
  the scope of this thesis.},
  address  = {Grantham-Allee 20, 53757 St. Augustin, Germany},
  annote  = {SS14/15 Pl{\"o}ger, Kraetzschmar, Vandenegro-Toro
  supervising},
  author  = {Arka Mallick},
  month = {January},
  school  = {Bonn-Rhein-Sieg University of Applied Sciences},
  title = {Sonar Patch Matching via Deep Learning},
  year = {2019}
  }

• S. Mahajan, “Realtime Deep Learning for Multispectral Human Detection,” Master Thesis, Grantham-Allee 20, 53757 St. Augustin, Germany, 2019.
  [BibTeX] [Abstract]
  

  Multispectral human detection is the detection of humans using colour and thermal images i.e. sensor fusion. The complementary information provided by the two make the detector more robust to environmental factors. This work aims to leverage sensor fusion to detect humans from an Unmanned Aerial Vehicle (UAV) mounted with a visible light and thermal camera. With this in view, we define a multispectral dataset called the Fraunhofer dataset consisting of two sequences captured with our hardware setup. Fraunhofer dataset’s small size and incompatibility with a deep learning-based detector make it less than ideal for evaluation. Instead, we use the KAIST Multispectral pedestrian detection dataset to train and evaluate our proposed detectors. From state-of-the-art object detection networks, Faster R-CNN with Resnet101 and SSD with MobilenetV2 are chosen as candidate architectures. Additionally, we evaluate weighted addition, LAB and stacking (RGB-T) image fusion techniques to achieve sensor fusion. Since no fusion technique performs significantly better than others, stacking is chosen based on the simplicity of computations. We compare the RGB-T input based Faster R-CNN Resnet101 and SSD MobilenetV2 with ACF+T+THOG and other state-of-the-art detectors Halfway Fusion, Fusion RPN + BDT, and MSDS-RCNN. MSDS-RCNN performs best with a miss rate of 18.50\% but at the cost of computational complexity which is \~7x compared to Faster R-CNN Resnet101 and \~46x compared to SSD MobilenetV2. Faster R-CNN Resnet101 and SSD MobilenetV2 achieve miss rates of 38.19\% and 41.28\% which are both better than the baseline ACF+T+THOG. Better accuracy of Faster R-CNN compared to SSD is offset by their performance in terms of speed. At \~81 FPS, SSD is \~8x faster than Faster R-CNN on an NVIDIA GeForce GTX 1080 Ti GPU. Finally, we curate the Fraunhofer dataset by applying image processing techniques to the colour and thermal images, obtaining synchronised, aligned images of the same size which are compatible with proposed detectors. To deploy the SSD MobilenetV2 on the UAV, the network is 8-bit quantised as required by Google’s Coral USB Accelerator; a USB drive-sized portable processor. On the Fraunhofer test set, the final SSD MobilnetV2 model achieves a miss rate of 89.04\% and mAP of 16.78\% while running at realtime inference speeds of \~14 FPS on the Coral TPU.

  @MastersThesis{ 2019mahajan,
  abstract  = {Multispectral human detection is the detection of humans
  using colour and thermal images i.e. sensor fusion. The
  complementary information provided by the two make the
  detector more robust to environmental factors. This work
  aims to leverage sensor fusion to detect humans from an
  Unmanned Aerial Vehicle (UAV) mounted with a visible light
  and thermal camera. With this in view, we define a
  multispectral dataset called the Fraunhofer dataset
  consisting of two sequences captured with our hardware
  setup.
  Fraunhofer dataset's small size and incompatibility with a
  deep learning-based detector make it less than ideal for
  evaluation. Instead, we use the KAIST Multispectral
  pedestrian detection dataset to train and evaluate our
  proposed detectors. From state-of-the-art object detection
  networks, Faster R-CNN with Resnet101 and SSD with
  MobilenetV2 are chosen as candidate architectures.
  Additionally, we evaluate weighted addition, LAB and
  stacking (RGB-T) image fusion techniques to achieve sensor
  fusion. Since no fusion technique performs significantly
  better than others, stacking is chosen based on the
  simplicity of computations. We compare the RGB-T input
  based Faster R-CNN Resnet101 and SSD MobilenetV2 with
  ACF+T+THOG and other state-of-the-art detectors Halfway
  Fusion, Fusion RPN + BDT, and MSDS-RCNN. MSDS-RCNN performs
  best with a miss rate of 18.50\% but at the cost of
  computational complexity which is \~7x compared to Faster
  R-CNN Resnet101 and \~46x compared to SSD MobilenetV2.
  Faster R-CNN Resnet101 and SSD MobilenetV2 achieve miss
  rates of 38.19\% and 41.28\% which are both better than the
  baseline ACF+T+THOG. Better accuracy of Faster R-CNN
  compared to SSD is offset by their performance in terms of
  speed. At \~81 FPS, SSD is \~8x faster than Faster R-CNN on
  an NVIDIA GeForce GTX 1080 Ti GPU.
  Finally, we curate the Fraunhofer dataset by applying image
  processing techniques to the colour and thermal images,
  obtaining synchronised, aligned images of the same size
  which are compatible with proposed detectors. To deploy the
  SSD MobilenetV2 on the UAV, the network is 8-bit quantised
  as required by Google's Coral USB Accelerator; a USB
  drive-sized portable processor. On the Fraunhofer test set,
  the final SSD MobilnetV2 model achieves a miss rate of
  89.04\% and mAP of 16.78\% while running at realtime
  inference speeds of \~14 FPS on the Coral TPU.},
  address  = {Grantham-Allee 20, 53757 St. Augustin, Germany},
  annote  = {[WS16/17] [Fraunhofer FKIE] [Thiele], [Koch], [Schleiss]
  supervising},
  author  = {Shweta Mahajan},
  month = {September},
  school  = {Bonn-Rhein-Sieg University of Applied Sciences},
  title = {Realtime Deep Learning for Multispectral Human Detection},
  year = {2019}
  }

• S. ul hassan Saad, “Decision Making for an Automated Vehicle Using Deep Reinforcement Learning,” Master Thesis, Grantham-Allee 20, 53757 St. Augustin, Germany, 2019.
  [BibTeX] [Abstract]
  

  At DLR’s Institute of Transportation Systems, research is being carried out on cooperative automated driving, lane following and lane changing. Lane changing in an automated vehicle depends on a large number of influencing factors such as whether there is a car at the front, left or right and the speeds with which the neighboring cars are moving with respect to the ego vehicle. The main focus of this master thesis is the design and implementation of a decision unit responsible for navigation of ego vehicle in a multi-lane and multi agent environment. Decision unit would decide five different actions for the ego vehicle i.e. change the lane either left or right, keep on following the same lane, increase or decrease the speed. In order to design the decision unit, neural networks based reinforcement learning is used. Reinforcement learning offers several advantages over other machine learning approaches. One main advantage is that the decision unit tries to learn the best possible strategy instead of just imitating human expert. Another advantage is that the decision unit can be trained without going into the hassle of preparation of input dataset which sometimes can be time consuming and costly as well. For training and testing, Simulation of Urban Mobility-SUMO environment is used. Two different Q learning based architectures are used to implement the decision unit. One architecture is solely based on fully connected layers and the other architecture uses convolutional layers along with fully connected layers. After training and evaluation of these architectures, it is observed that architecture with fully connected layers performs far better as compared to the other architecture.

  @MastersThesis{ 2015saad,
  abstract  = {At DLR’s Institute of Transportation Systems, research
  is being carried out on cooperative automated driving, lane
  following and lane changing. Lane changing in an automated
  vehicle depends on a large number of influencing factors
  such as whether there is a car at the front, left or right
  and the speeds with which the neighboring cars are moving
  with respect to the ego vehicle.
  The main focus of this master thesis is the design and
  implementation of a decision unit responsible for
  navigation of ego vehicle in a multi-lane and multi agent
  environment. Decision unit would decide five different
  actions for the ego vehicle i.e. change the lane either
  left or right, keep on following the same lane, increase or
  decrease the speed. In order to design the decision unit,
  neural networks based reinforcement learning is used.
  Reinforcement learning offers several advantages over other
  machine learning approaches. One main advantage is that the
  decision unit tries to learn the best possible strategy
  instead of just imitating human expert. Another advantage
  is that the decision unit can be trained without going into
  the hassle of preparation of input dataset which sometimes
  can be time consuming and costly as well.
  For training and testing, Simulation of Urban Mobility-SUMO
  environment is used. Two different Q learning based
  architectures are used to implement the decision unit. One
  architecture is solely based on fully connected layers and
  the other architecture uses convolutional layers along with
  fully connected layers. After training and evaluation of
  these architectures, it is observed that architecture with
  fully connected layers performs far better as compared to
  the other architecture.},
  address  = {Grantham-Allee 20, 53757 St. Augustin, Germany},
  annote  = {WS15/16 Deutschen Zentrums fur Luft- und Raumfahrt (DLR) -
  DLR-Institut fur Verkehrssystemtechnik Kraetzschmar,
  Pl{\"o}ger, Nicthing supervising},
  author  = {Sabeeh ul hassan Saad},
  month = {April},
  school  = {Bonn-Rhein-Sieg University of Applied Sciences},
  title = {Decision Making for an Automated Vehicle Using Deep
  Reinforcement Learning},
  year = {2019}
  }

2018

• J. Rajagopal, “Accurate trajectory tracking on the KUKA youBot manipulator: Computed-torque control and Friction compensation,” Master Thesis, Grantham-Allee 20, 53757 St. Augustin, Germany, 2018.
  [BibTeX] [Abstract]
  

  Accurate trajectory tracking control is a desirable quality for the robotic manipulators. Since the manipulators are highly non-linear due to the presence of structured (i.e. model errors), and unstructured uncertainties (i.e. friction), it becomes very difficult to achieve high-precision tracking in the manipulator joints. Most common solution to this problem is the computed-torque control scheme where the inclusion of the dynamic model linearizes the non-linear system. Disadvantage of this method is that the controller’s performance in high-speed operations relies heavily on the accuracy of the dynamic model parameters. Model parameters provided by the manufacturers are generally not accurate, and thus the identification of link parameters becomes necessary. Additionally, inclusion of friction compensation terms in the dynamic model improves the controller’s performance, and helps us in achieving better dynamics control of the manipulator. This work considers the implementation of both identification of the dynamic model parameters and computed- torque controller. The identification procedure is the continuation of the previous research where the geometric relation semantics and the dynamics of the system needed correction. Real concern in any of the robotics based applications is safety of the hardwares used (i.e. motors, sensors) because they are generally quite expensive. So, this work gives highest priority regarding safety of the real system and the safety control layer monitors the state of the joints with the help of the encoder data such as joint positions, velocities and torques. The model-based controller uses the dynamic model of the youBot manipulator and feed- forward torques are computed by using the inverse dynamics solver based on a library. There are two kinds of control schemes considered in this work such as the basic and the alternate control. The alternate control method is chosen over the basic method, because the basic approach suffers in predicting the model torques due to the presence of inaccurate dynamic model parameters. Both of the schemes implement the same cascade PI controller which is the combination of both the position and velocity controllers for the purpose of better disturbance rejection. Controller gains are tuned empirically to the optimum for most controlwise challenging joints on the end of the kinematic chain. This work reports analysis on the semantics used by the existing rigid-body algorithms and these findings are useful in constructing the geometrical relation semantics between rigid bodies without introducing the logical errors. Then, a safety control check is performed in the real system that handles the breaches in the safety limit of the manipulator joints effectively with an understandable latency issues due to the use of non real-time operating system. The correctness of the dynamic model is tested with the gravity compensation task, and then the pure controller without the dynamic model is validated with the analytical trajectories. The friction modelling and compensation experiments are conducted in the basic control scheme implemented in this work. The computed-torque control scheme is evaluated on the farthest joints of the base with the help of the analytically formulated trajectories. In spite of using the inaccurate model parameters in the dynamic model, the controller tracks of the trajectory accurately with an acceptable tracking error on the manipulator joints.

  @MastersThesis{ 2018rajagopal,
  abstract  = {Accurate trajectory tracking control is a desirable
  quality for the robotic manipulators. Since the
  manipulators are highly non-linear due to the presence of
  structured (i.e. model errors), and unstructured
  uncertainties (i.e. friction), it becomes very difficult to
  achieve high-precision tracking in the manipulator joints.
  Most common solution to this problem is the computed-torque
  control scheme where the inclusion of the dynamic model
  linearizes the non-linear system. Disadvantage of this
  method is that the controller’s performance in high-speed
  operations relies heavily on the accuracy of the dynamic
  model parameters. Model parameters provided by the
  manufacturers are generally not accurate, and thus the
  identification of link parameters becomes necessary.
  Additionally, inclusion of friction compensation terms in
  the dynamic model improves the controller’s performance,
  and helps us in achieving better dynamics control of the
  manipulator. This work considers the implementation of both
  identification of the dynamic model parameters and
  computed- torque controller. The identification procedure
  is the continuation of the previous research where the
  geometric relation semantics and the dynamics of the system
  needed correction. Real concern in any of the robotics
  based applications is safety of the hardwares used (i.e.
  motors, sensors) because they are generally quite
  expensive. So, this work gives highest priority regarding
  safety of the real system and the safety control layer
  monitors the state of the joints with the help of the
  encoder data such as joint positions, velocities and
  torques. The model-based controller uses the dynamic model
  of the youBot manipulator and feed- forward torques are
  computed by using the inverse dynamics solver based on a
  library. There are two kinds of control schemes considered
  in this work such as the basic and the alternate control.
  The alternate control method is chosen over the basic
  method, because the basic approach suffers in predicting
  the model torques due to the presence of inaccurate dynamic
  model parameters. Both of the schemes implement the same
  cascade PI controller which is the combination of both the
  position and velocity controllers for the purpose of better
  disturbance rejection. Controller gains are tuned
  empirically to the optimum for most controlwise challenging
  joints on the end of the kinematic chain. This work reports
  analysis on the semantics used by the existing rigid-body
  algorithms and these findings are useful in constructing
  the geometrical relation semantics between rigid bodies
  without introducing the logical errors. Then, a safety
  control check is performed in the real system that handles
  the breaches in the safety limit of the manipulator joints
  effectively with an understandable latency issues due to
  the use of non real-time operating system. The correctness
  of the dynamic model is tested with the gravity
  compensation task, and then the pure controller without the
  dynamic model is validated with the analytical
  trajectories. The friction modelling and compensation
  experiments are conducted in the basic control scheme
  implemented in this work. The computed-torque control
  scheme is evaluated on the farthest joints of the base with
  the help of the analytically formulated trajectories. In
  spite of using the inaccurate model parameters in the
  dynamic model, the controller tracks of the trajectory
  accurately with an acceptable tracking error on the
  manipulator joints.},
  address  = {Grantham-Allee 20, 53757 St. Augustin, Germany},
  annote  = {SS12 FH-BRS - Plöger, Chakirov, Schneider supervising},
  author  = {Jeyaprakash Rajagopal},
  month = {August},
  school  = {Bonn-Rhein-Sieg University of Applied Sciences},
  title = {Accurate trajectory tracking on the KUKA youBot
  manipulator: Computed-torque control and Friction
  compensation},
  year = {2018}
  }

• S. N. Boris, “An Affordable, Integrated and Digitized System for Sound Insulation Tests in Buildings,” Master Thesis, Grantham-Allee 20, 53757 St. Augustin, Germany, 2018.
  [BibTeX] [Abstract]
  

  In the field of building acoustics, sound insulation is the capacity of buildings not to let external noises in or sounds from the inside out. This property of buildings is important for the well-being, comfort and privacy of building occupants, and therefore regulations on acceptable sound insulation levels are defined by each country depending on the building type and usage. It is then customary to conduct sound insulation tests on buildings in order to determine their sound insulation capability and verify whether they conform to regulations. In this work, an affordable, integrated, and digitized system for sound insulation tests in buildings is proposed. Affordable, because low-cost alternatives to the conventional measurement equipment and strategies used for sound insulation tests are proposed. Integrated, because all sound insulation test procedures are integrated into a single system that generates test signals, acquires and processes acoustic data, and detects whether or not building acoustics regulations are met. Digitized, because the system proposes to automate the interpretation of sound insulation test results, which is usually done by building acoustics experts, by diagnosing and identifying the potential defects that may cause poor sound insulation and suggesting potential remedies to improve the sound insulation. To achieve an affordable system, this work proposes alternative strategies in order to use a directional loudspeaker as well as a low precision and non-calibrated microphone as alternatives to the conventional omnidirectional and high precision and calibrated microphones used in sound insulation tests. Part of the proposed system is implemented, tested and evaluated. The evaluation is done qualitatively by investigating the precision and accuracy of the obtained results in varying environments and conditions, and by analysing how the implemented system compares to other solutions for the calculation of acoustic parameters namely the Sound Level Meter free mobile phone application for Sound Pressure Level calculation, and Norsonic’s building acoustics proprietary software for Sound Pressure Level and reverberation time calculation. Generally the results of the implemented system show closeness to those of the reference solutions, and the main differences and discrepancies are shown to be caused by the choice of the equipment. To digitize the interpretation of sound insulation tests results, a proposition is made to make use of the mass-law which states that the sound reduction of a building element increases by $6 dB$ as its mass density or the sound frequency is doubled. It is shown that a number of features like the Schroeder frequency, the coincidence frequency, the mass density, and the damping level of the building element under test can be extracted from the mass-law and subsequently used to categorize and diagnose which defects are present. Two commonly encountered defects in building acoustics, the weak wall and the air-leakage defects, are particularly investigated and discussed in the light of experiments, and an agreement with the mass-law is observed hence supporting the proposition that it can be used to digitize the interpretation of results.

  @MastersThesis{ 2018ndimubanzi,
  abstract  = { In the field of building acoustics, sound insulation is
  the capacity of buildings not to let external noises in or
  sounds from the inside out. This property of buildings is
  important for the well-being, comfort and privacy of
  building occupants, and therefore regulations on acceptable
  sound insulation levels are defined by each country
  depending on the building type and usage. It is then
  customary to conduct sound insulation tests on buildings in
  order to determine their sound insulation capability and
  verify whether they conform to regulations.
  In this work, an affordable, integrated, and digitized
  system for sound insulation tests in buildings is proposed.
  Affordable, because low-cost alternatives to the
  conventional measurement equipment and strategies used for
  sound insulation tests are proposed. Integrated, because
  all sound insulation test procedures are integrated into a
  single system that generates test signals, acquires and
  processes acoustic data, and detects whether or not
  building acoustics regulations are met. Digitized, because
  the system proposes to automate the interpretation of sound
  insulation test results, which is usually done by building
  acoustics experts, by diagnosing and identifying the
  potential defects that may cause poor sound insulation and
  suggesting potential remedies to improve the sound
  insulation.
  To achieve an affordable system, this work proposes
  alternative strategies in order to use a directional
  loudspeaker as well as a low precision and non-calibrated
  microphone as alternatives to the conventional
  omnidirectional and high precision and calibrated
  microphones used in sound insulation tests. Part of the
  proposed system is implemented, tested and evaluated. The
  evaluation is done qualitatively by investigating the
  precision and accuracy of the obtained results in varying
  environments and conditions, and by analysing how the
  implemented system compares to other solutions for the
  calculation of acoustic parameters namely the Sound Level
  Meter free mobile phone application for Sound Pressure
  Level calculation, and Norsonic's building acoustics
  proprietary software for Sound Pressure Level and
  reverberation time calculation. Generally the results of
  the implemented system show closeness to those of the
  reference solutions, and the main differences and
  discrepancies are shown to be caused by the choice of the
  equipment.
  To digitize the interpretation of sound insulation tests
  results, a proposition is made to make use of the mass-law
  which states that the sound reduction of a building element
  increases by $6 dB$ as its mass density or the sound
  frequency is doubled. It is shown that a number of features
  like the Schroeder frequency, the coincidence frequency,
  the mass density, and the damping level of the building
  element under test can be extracted from the mass-law and
  subsequently used to categorize and diagnose which defects
  are present. Two commonly encountered defects in building
  acoustics, the weak wall and the air-leakage defects, are
  particularly investigated and discussed in the light of
  experiments, and an agreement with the mass-law is observed
  hence supporting the proposition that it can be used to
  digitize the interpretation of results. },
  address  = {Grantham-Allee 20, 53757 St. Augustin, Germany},
  annote  = {SS15 measX GmbH Pl{\"o}ger, Kraetzschmar, Hilsmann
  supervising},
  author  = {Senga Ndimubanzi Boris},
  month = {August},
  school  = {Bonn-Rhein-Sieg University of Applied Sciences},
  title = {An Affordable, Integrated and Digitized System for Sound
  Insulation Tests in Buildings},
  year = {2018}
  }

• A. Vinokurov, “Towards improvements on RoboCup @home robots architecture, capabilities and development process,” Master Thesis, Grantham-Allee 20, 53757 St. Augustin, Germany, 2018.
  [BibTeX] [Abstract]
  

  Domestic robotics is a vast field where a lot of knowledge and team effort is required to build quality software for daily use and for RoboCup competitions. Current robot that the research is carried on is the service robot Care-O-bot 3, developed by Fraunhofer IPA and that is used for research at Hochshule Bonn-Rhein-Sieg. It is a big machine containing 4 different computers and many peripheral devices includ- ing laser scanners, audio system, a manipulator (that has a dedicated computer) and other peripheral devices. It is fairly complicated to come up with a suitable archi- tecture for such a robot, so it would be most efficient in @home scenarios. Partially this is due to the overall complexity of the robot, partially due to the deficiencies in current development process. For example the lack of testing procedures. The goal of this project is to analyse current robot architecture and modify it in a way so it would be more reliable and easy to maintain by the team. Another goal is to separate ROS framework dependencies and the algorithmic part of the software developed as well as to provide a template for the package structure that would facilitate writing software in such a way, so frameworks will be easily inter- changeable. Good documentation is vital for the new members of the team, who are not familiar with the development process and tooling, so to ease the on-boarding process everything will be well documented. After the start of this work a new Toyota HSR robot was acquired, featuring its own architecture and Python API abstraction from ROS framework. This robot will be looked at in detail form the software point of view to study the way of integrating both robots in one common software space.

  @MastersThesis{ 2018vinokurov,
  abstract  = {Domestic robotics is a vast field where a lot of knowledge
  and team effort is required to build quality software for
  daily use and for RoboCup competitions. Current robot that
  the research is carried on is the service robot Care-O-bot
  3, developed by Fraunhofer IPA and that is used for
  research at Hochshule Bonn-Rhein-Sieg. It is a big machine
  containing 4 different computers and many peripheral
  devices includ- ing laser scanners, audio system, a
  manipulator (that has a dedicated computer) and other
  peripheral devices. It is fairly complicated to come up
  with a suitable archi- tecture for such a robot, so it
  would be most efficient in @home scenarios. Partially this
  is due to the overall complexity of the robot, partially
  due to the deficiencies in current development process. For
  example the lack of testing procedures. The goal of this
  project is to analyse current robot architecture and modify
  it in a way so it would be more reliable and easy to
  maintain by the team. Another goal is to separate ROS
  framework dependencies and the algorithmic part of the
  software developed as well as to provide a template for the
  package structure that would facilitate writing software in
  such a way, so frameworks will be easily inter- changeable.
  Good documentation is vital for the new members of the
  team, who are not familiar with the development process and
  tooling, so to ease the on-boarding process everything will
  be well documented. After the start of this work a new
  Toyota HSR robot was acquired, featuring its own
  architecture and Python API abstraction from ROS framework.
  This robot will be looked at in detail form the software
  point of view to study the way of integrating both robots
  in one common software space. },
  address  = {Grantham-Allee 20, 53757 St. Augustin, Germany},
  annote  = {WS14 Pl{\"o}ger, Kraetzschmar, Mitrevski supervising},
  author  = {Artem Vinokurov},
  month = {July},
  school  = {Bonn-Rhein-Sieg University of Applied Sciences},
  title = {Towards improvements on RoboCup @home robots architecture,
  capabilities and development process},
  year = {2018}
  }

• P. S. Vokuda, “Interactive Object Detection,” Master Thesis, Grantham-Allee 20, 53757 St. Augustin, Germany, 2018.
  [BibTeX] [Abstract]
  

  The success of state-of-the-art object detection methods depend heavily on the availability of a large amount of annotated image data. The raw image data available from various sources are abundant but non-annotated. Annotating image data is often costly, time-consuming or needs expert help. In this work, a new paradigm of learning called Active Learning is explored which uses user interaction to obtain annotations for a sub- set of the dataset. The goal of active learning is to achieve superior object detection performance with images that are annotated on demand. To realize active learning method, the trade-off between the effort to annotate (annotation cost) unlabelled data and the performance of object detection model is minimised. Random Forests based method called Hough Forest is chosen as the object detection model and the annotation cost is calculated as the predicted false positive and false negative rate. The framework is successfully evaluated on two Computer Vision benchmark and two Carl Zeiss custom datasets. Also, an evaluation of RGB, HoG and Deep features for the task is presented. Experimental results show that using Deep features with Hough Forest achieves the maximum performance. By employing Active Learning, it is demonstrated that performance comparable to the fully supervised setting can be achieved by annotating just 2.5\% of the dataset. To this end, an annotation tool is developed for user interaction during Active Learning.

  @MastersThesis{ 2018vokuda,
  abstract  = {The success of state-of-the-art object detection methods
  depend heavily on the availability of a large amount of
  annotated image data. The raw image data available from
  various sources are abundant but non-annotated. Annotating
  image data is often costly, time-consuming or needs expert
  help. In this work, a new paradigm of learning called
  Active Learning is explored which uses user interaction to
  obtain annotations for a sub- set of the dataset. The goal
  of active learning is to achieve superior object detection
  performance with images that are annotated on demand. To
  realize active learning method, the trade-off between the
  effort to annotate (annotation cost) unlabelled data and
  the performance of object detection model is minimised.
  Random Forests based method called Hough Forest is chosen
  as the object detection model and the annotation cost is
  calculated as the predicted false positive and false
  negative rate. The framework is successfully evaluated on
  two Computer Vision benchmark and two Carl Zeiss custom
  datasets. Also, an evaluation of RGB, HoG and Deep features
  for the task is presented.
  Experimental results show that using Deep features with
  Hough Forest achieves the maximum performance. By employing
  Active Learning, it is demonstrated that performance
  comparable to the fully supervised setting can be achieved
  by annotating just 2.5\% of the dataset. To this end, an
  annotation tool is developed for user interaction during
  Active Learning.},
  address  = {Grantham-Allee 20, 53757 St. Augustin, Germany},
  annote  = {WS15 FH-BRS - Pl{\"o}ger, Thiele supervising},
  author  = {Priyanka Subramanya Vokuda},
  month = {April},
  school  = {Bonn-Rhein-Sieg University of Applied Sciences},
  title = {Interactive Object Detection},
  year = {2018}
  }

• L. Wang, Y. Xiong, Z. Wang, Y. Qiao, D. Lin, X. Tang, and L. {Van Gool}, “Temporal Segment Networks for Action Recognition in Videos,” , 2018.
  [BibTeX]

  @Article{ wang2018,
  author  = {Wang, Limin and Xiong, Yuanjun and Wang, Zhe and Qiao, Yu
  and Lin, Dahua and Tang, Xiaoou and {Van Gool}, Luc},
  booktitle  = {IEEE Transactions on Pattern Analysis and Machine
  Intelligence},
  publisher  = {IEEE Computer Society},
  title = {{Temporal Segment Networks for Action Recognition in
  Videos}},
  year = {2018}
  }

• D. Wang, “Autonomous Navigation in Multi-floor Buildings,” Master Thesis, Grantham-Allee 20, 53757 St. Augustin, Germany, 2018.
  [BibTeX] [Abstract]
  

  Service robots are developing considerably these years due to their potential applications, especially in complex large-scale multi-floor buildings. The service robots need to autonomously, accurately and robustly navigate in the whole building for whatever task they execute. The previous work realizes multi-floor navigation with a lot of human efforts and external devices installment. This Master thesis aims to examine the possibility and localization accuracy of multi-floor navigation with less human efforts and without external devices installment. The proposed autonomous multi-floor navigation system is divided into four parts: multi-floor mapping, floor level determination, elevator operation, and multi-floor navigation. For each part, the state of the art is surveyed, as well as the algorithm is implemented and tested on a real robot. The multi-floor mapping includes three floor maps generation using Cartographer package, a transition file creation, as well as the multi-floor map server implementation that automatically provides the respective floor map and publishes the corresponding transition pose when the floor level is changed. The floor level determination algorithm outputs the current floor level based on acceleration features during the elevator movement. The elevator button recognition algorithm using KNN learning algorithm and Hog features determines the coordinates of the required button when it is given RGB-D images. The multi-floor navigation based on smach library integrates all subtasks and executes it step by step such that the robot can successfully reach the goal. All experiments are performed on the robot HSR in the main building of our university. The generated three floor maps are clear and close to the floor plan. The floor level determination algorithm is tested for each of three elevators and the accuracy reaches up to 100\%. The button recognition algorithm reaches up to 99.84\% accuracy under the normal light condition, while the accuracy reduces to 80.3\% when the environment is dim. The multi-floor navigation using amcl and move base package in ROS can navigate the robot across floors with up to 100\% success rate, however, the localization accuracy is approximately 50 cm for short path. In some long path cases, the robot can not successfully reach the goal because it can not accurately localize itself due to the featureless environment, inaccurate odometry data, and inaccurate map. In the future, the localization accuracy of the robot can be improved in combination with other localization methods, then the robot will successfully, accurately reach the goal in multi-floor buildings.

  @MastersThesis{ 2018wang,
  abstract  = {Service robots are developing considerably these years due
  to their potential applications, especially in complex
  large-scale multi-floor buildings. The service robots need
  to autonomously, accurately and robustly navigate in the
  whole building for whatever task they execute. The previous
  work realizes multi-floor navigation with a lot of human
  efforts and external devices installment. This Master
  thesis aims to examine the possibility and localization
  accuracy of multi-floor navigation with less human efforts
  and without external devices installment.
  The proposed autonomous multi-floor navigation system is
  divided into four parts: multi-floor mapping, floor level
  determination, elevator operation, and multi-floor
  navigation. For each part, the state of the art is
  surveyed, as well as the algorithm is implemented and
  tested on a real robot. The multi-floor mapping includes
  three floor maps generation using Cartographer package, a
  transition file creation, as well as the multi-floor map
  server implementation that automatically provides the
  respective floor map and publishes the corresponding
  transition pose when the floor level is changed. The floor
  level determination algorithm outputs the current floor
  level based on acceleration features during the elevator
  movement. The elevator button recognition algorithm using
  KNN learning algorithm and Hog features determines the
  coordinates of the required button when it is given RGB-D
  images. The multi-floor navigation based on smach library
  integrates all subtasks and executes it step by step such
  that the robot can successfully reach the goal.
  All experiments are performed on the robot HSR in the main
  building of our university. The generated three floor maps
  are clear and close to the floor plan. The floor level
  determination algorithm is tested for each of three
  elevators and the accuracy reaches up to 100\%. The button
  recognition algorithm reaches up to 99.84\% accuracy under
  the normal light condition, while the accuracy reduces to
  80.3\% when the environment is dim. The multi-floor
  navigation using amcl and move base package in ROS can
  navigate the robot across floors with up to 100\% success
  rate, however, the localization accuracy is approximately
  50 cm for short path. In some long path cases, the robot
  can not successfully reach the goal because it can not
  accurately localize itself due to the featureless
  environment, inaccurate odometry data, and inaccurate map.
  In the future, the localization accuracy of the robot can
  be improved in combination with other localization methods,
  then the robot will successfully, accurately reach the goal
  in multi-floor buildings.},
  address  = {Grantham-Allee 20, 53757 St. Augustin, Germany},
  annote  = {WS16 FH-BRS - ROPOD Pl{\"o}ger, Prassler, Ortega
  supervising},
  author  = {Danning Wang},
  month = {December},
  school  = {Bonn-Rhein-Sieg University of Applied Sciences},
  title = {Autonomous Navigation in Multi-floor Buildings},
  year = {2018}
  }

• M. Naazare, “Simultaneous Exploration and Information Delivery Using Multiple Heterogeneous Robots Under Limited Communication,” Master Thesis, Grantham-Allee 20, 53757 St. Augustin, Germany, 2018.
  [BibTeX] [Abstract]
  

  After a disaster, there is an urgent need to deliver humanitarian aid such as emergency medical treatment, supply of food and medicines to the affected regions. Since disasters can cause existing maps to change, reaching the affected regions immediately can be extremely challenging. In this work, the problem of exploring an unknown environment using a team of Unmanned Aerial Vehicles (UAVs) and Unmanned Ground Vehicles (UGV) to discover traversable regions for emergency response vehicles is investigated. The goal of the exploration is to ensure that the explored information reaches the base station despite its strict limited communication range which allows no inter-robot communication outside the range. For the given exploration scenario, the key issue addressed is the coordination of multiple-robots The proposed Simultaneous Exploration and Information Delivery (SEAID) planner operates as a centralised structure and assigns target positions to the Multi-robot System (MRS) to explore, and report the gathered information periodically. It deploys UAVs to perform Boustrophedon motions while UGVs perform frontier-based exploration. As UAVs return periodically, more information about the environment is revealed and the planner switches from frontier-based planning to grid-based coverage for the UGVs where it assigns grid-like pattern of target positions. To ensure coordination, the planner predicts the influence of other robots due to their presence and their respective goals on the new information that can be gained for a robot. Additionally, it employs a divide and rule strategy and assigns robots their respective regions to carry out their tasks. To evaluate the performance of different coordination strategies and observe the behaviour of the MRS, a Robot Operating System-based benchmark called SEAID-Sim was developed to simulate the given scenario. The proposed SEAID planner was tested against existing techniques for multi-robot exploration and multi-robot coverage. The results demonstrate that SEAID planner explores the complete environment, discovers all the obstacles, traversable and non-traversable regions for the emergency response vehicles faster than existing techniques.

  @MastersThesis{ 2018naazare,
  abstract  = {After a disaster, there is an urgent need to deliver
  humanitarian aid such as emergency medical treatment,
  supply of food and medicines to the affected regions. Since
  disasters can cause existing maps to change, reaching the
  affected regions immediately can be extremely challenging.
  In this work, the problem of exploring an unknown
  environment using a team of Unmanned Aerial Vehicles (UAVs)
  and Unmanned Ground Vehicles (UGV) to discover traversable
  regions for emergency response vehicles is investigated.
  The goal of the exploration is to ensure that the explored
  information reaches the base station despite its strict
  limited communication range which allows no inter-robot
  communication outside the range. For the given exploration
  scenario, the key issue addressed is the coordination of
  multiple-robots
  The proposed Simultaneous Exploration and Information
  Delivery (SEAID) planner operates as a centralised
  structure and assigns target positions to the Multi-robot
  System (MRS) to explore, and report the gathered
  information periodically. It deploys UAVs to perform
  Boustrophedon motions while UGVs perform frontier-based
  exploration. As UAVs return periodically, more information
  about the environment is revealed and the planner switches
  from frontier-based planning to grid-based coverage for the
  UGVs where it assigns grid-like pattern of target
  positions. To ensure coordination, the planner predicts the
  influence of other robots due to their presence and their
  respective goals on the new information that can be gained
  for a robot. Additionally, it employs a divide and rule
  strategy and assigns robots their respective regions to
  carry out their tasks.
  To evaluate the performance of different coordination
  strategies and observe the behaviour of the MRS, a Robot
  Operating System-based benchmark called SEAID-Sim was
  developed to simulate the given scenario. The proposed
  SEAID planner was tested against existing techniques for
  multi-robot exploration and multi-robot coverage. The
  results demonstrate that SEAID planner explores the
  complete environment, discovers all the obstacles,
  traversable and non-traversable regions for the emergency
  response vehicles faster than existing techniques.},
  address  = {Grantham-Allee 20, 53757 St. Augustin, Germany},
  annote  = {[SS 2014] [Becker], [Alda], [Br{\"u}ggemann]},
  author  = {Menaka Naazare},
  month = {June},
  school  = {Bonn-Rhein-Sieg University of Applied Sciences},
  title = {Simultaneous Exploration and Information Delivery Using
  Multiple Heterogeneous Robots Under Limited Communication},
  year = {2018}
  }

• R. Mendieta, “Intelligent Dynamic Occupancy Grids,” Master Thesis, Grantham-Allee 20, 53757 St. Augustin, Germany, 2018.
  [BibTeX] [Abstract]
  

  Occupancy grid maps for autonomous driving applications are built with a time-of-flight sensor to obtain information on the spatial 3D geometry. Motivated with providing redundancy and safety when building maps, an approach using a monocular camera is proposed to obtain depth information and semantic segmentation. Deep learning approaches for monocular depth estimation were trained with the KITTI dataset and tested with different network architectures (residual, convolutional with recurrency, residual with recurrency) and loss function variations. Additionally, the benefits of adding semantic segmentation and vehicle odometry estimation were investigated. Furthermore, skip connections, image input at different resolutions and a second refinement post-processing network were considered in the experiments. The experiments showed that a residual architecture with semantic segmentation input and a refinement post-processing, trained with the reverse Huber loss yielded the best results of the tested networks. This approach showed improved performance when compared against other monocular camera state of the art approaches and even comparable to approaches using two cameras, reaching 5.031 on RMSE linear error using the \Eigen” split of the KITTI dataset. This work was carried out as part of a Master Thesis job for the Advance Software Development team in the Autonomous Driving Group at Intel in Karlsruhe. Keywords: depth estimation, semantic segmentation, convolutional neural network, residual neural network, autonomous driving.

  @MastersThesis{ 2018mendieta,
  abstract  = {Occupancy grid maps for autonomous driving applications
  are built with a time-of-flight sensor to obtain
  information on the spatial 3D geometry. Motivated with
  providing redundancy and safety when building maps, an
  approach using a monocular camera is proposed to obtain
  depth information and semantic segmentation. Deep learning
  approaches for monocular depth estimation were trained with
  the KITTI dataset and tested with different network
  architectures (residual, convolutional with recurrency,
  residual with recurrency) and loss function variations.
  Additionally, the benefits of adding semantic segmentation
  and vehicle odometry estimation were investigated.
  Furthermore, skip connections, image input at different
  resolutions and a second refinement post-processing network
  were considered in the experiments. The experiments showed
  that a residual architecture with semantic segmentation
  input and a refinement post-processing, trained with the
  reverse Huber loss yielded the best results of the tested
  networks. This approach showed improved performance when
  compared against other monocular camera state of the art
  approaches and even comparable to approaches using two
  cameras, reaching 5.031 on RMSE linear error using the
  \Eigen" split of the KITTI dataset. This work was carried
  out as part of a Master Thesis job for the Advance Software
  Development team in the Autonomous Driving Group at Intel
  in Karlsruhe. Keywords: depth estimation, semantic
  segmentation, convolutional neural network, residual neural
  network, autonomous driving.},
  address  = {Grantham-Allee 20, 53757 St. Augustin, Germany},
  annote  = {SS2015 Intel - Intelligent Dynamic Occupancy Grids
  Plöger, Thiele supervising},
  author  = {Roberto Mendieta},
  month = {September},
  school  = {Bonn-Rhein-Sieg University of Applied Sciences},
  title = {Intelligent Dynamic Occupancy Grids},
  year = {2018}
  }

• L. O. A. Camargo, “Artificial Transfer Learning in Convolutional Neural Networks,” Master Thesis, Grantham-Allee 20, 53757 St. Augustin, Germany, 2018.
  [BibTeX] [Abstract]
  

  Recent studies in transfer learning with synthetic data have shown significant results in the context of deep neural networks. However, most of these developments have not been thoroughly studied for the tasks of object detection and image classification. To research the learning transfer ability of artificial neural networks for these specific tasks, we generated over 1M synthetic images. These images were rendered from 35K 3D models extracted from the ShapeNet dataset. We trained several convolutional neural network architectures to perform classification and object detection using our synthetic datasets. We then performed transfer learning by fine-tuning the networks with real images extracted from the COCO and VOC2007/2012 datasets. In order to provide statistically significant results, we trained over 1K CNNs and we show that networks pre-trained with synthetic data are able to obtain higher accuracies and mAPs than the networks trained from scratch. Thus, we show empirical evidence that inexpensive synthetic data can be used to improve general classification and detection tasks in deep learning models. Furthermore, we are able to leverage from this result in order to train a real-time CPU object detector under the single-shot multibox detector scheme. Finally we also provide a self-contained deep learning review that allows us to revisit several hypothesis made in modern CNN architectures.

  @MastersThesis{ 2018arriaga,
  abstract  = {Recent studies in transfer learning with synthetic data
  have shown significant results in the context of deep
  neural networks. However, most of these developments have
  not been thoroughly studied for the tasks of object
  detection and image classification. To research the
  learning transfer ability of artificial neural networks for
  these specific tasks, we generated over 1M synthetic
  images. These images were rendered from 35K 3D models
  extracted from the ShapeNet dataset. We trained several
  convolutional neural network architectures to perform
  classification and object detection using our synthetic
  datasets. We then performed transfer learning by
  fine-tuning the networks with real images extracted from
  the COCO and VOC2007/2012 datasets. In order to provide
  statistically significant results, we trained over 1K CNNs
  and we show that networks pre-trained with synthetic data
  are able to obtain higher accuracies and mAPs than the
  networks trained from scratch. Thus, we show empirical
  evidence that inexpensive synthetic data can be used to
  improve general classification and detection tasks in deep
  learning models. Furthermore, we are able to leverage from
  this result in order to train a real-time CPU object
  detector under the single-shot multibox detector scheme.
  Finally we also provide a self-contained deep learning
  review that allows us to revisit several hypothesis made in
  modern CNN architectures.},
  address  = {Grantham-Allee 20, 53757 St. Augustin, Germany},
  annote  = {WS15/16 Pl{\"o}ger, Asteroth, Valdenegro-Toro supervising},
  author  = {Luis Octavio Arriaga Camargo},
  month = {February},
  school  = {Bonn-Rhein-Sieg University of Applied Sciences},
  title = {Artificial Transfer Learning in Convolutional Neural
  Networks},
  year = {2018}
  }

• D. Ali, “Actor Pre-Training: Application to Plant Control,” Master Thesis, Grantham-Allee 20, 53757 St. Augustin, Germany, 2018.
  [BibTeX] [Abstract]
  

  Reinforcement Learning (RL) methods take millions of interactions with the system to find a reasonable control policy, which in some applications is expensive and time consuming. Also, these policies start from a random control, which can be catastrophic to the system. Hence, to mitigate these problems we pre-train the policy with expert control behavior, such that it starts from a reasonable control policy. Inverse Reinforcement Learning (IRL) methods attempt to recover reward function, give expert demonstrations. They attempt to find an expert policy by RL combined with the recovered reward. Therefore, we employ IRL methods to pre-train policy with expert data and try to compare the performance of pre-trained policy against simple RL based policies. To the best of our knowledge this is the first attempt to use IRL methods for policy pre-training purposes. We showed that good policies can be trained by Imitation Learning methods, given exact system model. We presented IRL methods which can completely work offline during pre-training phase and were able save large training interactions during online learning phase.

  @MastersThesis{ 2018ali,
  abstract  = {Reinforcement Learning (RL) methods take millions of
  interactions with the system to find a reasonable control
  policy, which in some applications is expensive and time
  consuming. Also, these policies start from a random
  control, which can be catastrophic to the system. Hence, to
  mitigate these problems we pre-train the policy with expert
  control behavior, such that it starts from a reasonable
  control policy. Inverse Reinforcement Learning (IRL)
  methods attempt to recover reward function, give expert
  demonstrations. They attempt to find an expert policy by RL
  combined with the recovered reward. Therefore, we employ
  IRL methods to pre-train policy with expert data and try to
  compare the performance of pre-trained policy against
  simple RL based policies. To the best of our knowledge this
  is the first attempt to use IRL methods for policy
  pre-training purposes. We showed that good policies can be
  trained by Imitation Learning methods, given exact system
  model. We presented IRL methods which can completely work
  offline during pre-training phase and were able save large
  training interactions during online learning phase.},
  address  = {Grantham-Allee 20, 53757 St. Augustin, Germany},
  annote  = { SS2014 Recogizer Group GmbH - Actor Pre-Training:
  Application to Plant Control Plöger, Asteroth, Zimmerling
  supervising},
  author  = {Daiem Ali},
  month = {November},
  school  = {Bonn-Rhein-Sieg University of Applied Sciences},
  title = {Actor Pre-Training: Application to Plant Control},
  year = {2018}
  }

• S. Kannaiah, “Weakly Supervised Object Detection,” Master Thesis, Grantham-Allee 20, 53757 St. Augustin, Germany, 2018.
  [BibTeX] [Abstract]
  

  Creating object detection datasets for supervised training is very arduous and time consuming since they require class level labels as well as bounding box annotations. Thus to mitigate the problem weakly supervised object detection is necessary where there are only class level labels and no bounding box annotations. Most of the existing methods for weakly supervised object detection concentrate on just post-processing the output of object proposal algorithms without really understanding the merits or demerits of the object proposal algorithms used. Thus the existing methods fail to detect multiple objects in the scene. Our method takes a holistic approach, by first analyzing the factors that improve object localization using CNN’s which is one of primary contributions of our work. With these findings a new CNN based object proposal algorithm is proposed and has the highest recall rates compared to other object proposal algorithms on pascal VOC dataset with 0.8, 0.95 and 0.96 for 100, 1000 and 1805 number of object proposals respectively for an IoU threshold of 50\% . This new CNN based object proposal algorithm was our second contribution. Finally a pipeline was built around this object proposal algorithm to do object detection in a weakly supervised manner. Our proposed method excels where other methods perform poorly, especially when there are multiple objects in the scene. Our method for weakly supervised object detection achieves a second place in non end-to-end methods and third amongst all the methods doing weakly supervised object detection

  @MastersThesis{ 2018kannaiah,
  abstract  = {Creating object detection datasets for supervised training
  is very arduous and time consuming since they require class
  level labels as well as bounding box annotations. Thus to
  mitigate the problem weakly supervised object detection is
  necessary where there are only class level labels and no
  bounding box annotations. Most of the existing methods for
  weakly supervised object detection concentrate on just
  post-processing the output of object proposal algorithms
  without really understanding the merits or demerits of the
  object proposal algorithms used. Thus the existing methods
  fail to detect multiple objects in the scene. Our method
  takes a holistic approach, by first analyzing the factors
  that improve object localization using CNN's which is one
  of primary contributions of our work. With these findings a
  new CNN based object proposal algorithm is proposed and has
  the highest recall rates compared to other object proposal
  algorithms on pascal VOC dataset with 0.8, 0.95 and 0.96
  for 100, 1000 and 1805 number of object proposals
  respectively for an IoU threshold of 50\% . This new CNN
  based object proposal algorithm was our second
  contribution. Finally a pipeline was built around this
  object proposal algorithm to do object detection in a
  weakly supervised manner. Our proposed method excels where
  other methods perform poorly, especially when there are
  multiple objects in the scene. Our method for weakly
  supervised object detection achieves a second place in non
  end-to-end methods and third amongst all the methods doing
  weakly supervised object detection},
  address  = {Grantham-Allee 20, 53757 St. Augustin, Germany},
  annote  = {WS2015, Towards Weakly Supervised Object Detection,
  Ploeger, Breuer, Valdenegro supervising},
  author  = {Saikiran Kannaiah},
  month = {July},
  school  = {Bonn-Rhein-Sieg University of Applied Sciences},
  title = {Weakly Supervised Object Detection},
  year = {2018}
  }

• S. A. N. Khan, “Path optimization for a VR application using local search,” Master Thesis, Grantham-Allee 20, 53757 St. Augustin, Germany, 2018.
  [BibTeX] [Abstract]
  

  In this thesis, an algorithm is proposed and implemented to solve Time Dependent Shortest Path Problems (TDSPP) within an NP-hard Traveling Salesman Problem (TSP). Local search algorithms are useful in solving such problems. Various local search algorithms are tested in this thesis. Particle Swarm Optimization (PSO) is chosen as the most suitable method to solve TSPs. PSO is adapted to solve TDSPP within a TSP. To do this with the least changes in the original path, various methods are introduced. At the end, it is concluded, that in a TDSPP within a TSP of 30 cities, the implemented algorithm finds an optimal solution 99% of the times with minimal variation from the original path within a 1200 milliseconds processing time limit if a solution exists. A space ride simulation synchronized with a massage program is used as an example scenario.

  @MastersThesis{ 2018khan,
  abstract  = {In this thesis, an algorithm is proposed and implemented
  to solve Time Dependent Shortest Path Problems (TDSPP)
  within an NP-hard Traveling Salesman Problem (TSP). Local
  search algorithms are useful in solving such problems.
  Various local search algorithms are tested in this thesis.
  Particle Swarm Optimization (PSO) is chosen as the most
  suitable method to solve TSPs. PSO is adapted to solve
  TDSPP within a TSP. To do this with the least changes in
  the original path, various methods are introduced. At the
  end, it is concluded, that in a TDSPP within a TSP of 30
  cities, the implemented algorithm finds an optimal solution
  99% of the times with minimal variation from the original
  path within a 1200 milliseconds processing time limit if a
  solution exists. A space ride simulation synchronized with
  a massage program is used as an example scenario.},
  address  = {Grantham-Allee 20, 53757 St. Augustin, Germany},
  annote  = {SS13 N/A - N/A Heiden, Kraetzschmar, AwaadLeon
  supervising},
  author  = {Sardar Adnan Nawaz Khan},
  month = {July},
  school  = {Bonn-Rhein-Sieg University of Applied Sciences},
  title = {Path optimization for a VR application using local
  search},
  year = {2018}
  }

• P. Kulkarni, “nthropomorphic motion mapping under constraints for direct and intuitive control of a dual-arm manipulator,” Master Thesis, Wilhelm-Levison-Str. 22, Bonn, 53115, Germany, 2018.
  [BibTeX] [Abstract]
  

  Robots can accomplish tasks in environments that are hazardous and pose a threat to human safety. However, dynamic, unknown, or unstructured environments still require a degree of human decision making which autonomous robots cannot provide. Teleoperation of robots with a human making decisions offers a solution. Dual-arm robots provide an added advantage in these situations as they offer the possibility to accomplish a task in a manner similar to humans. Additionally, teleoperation with direct transfer of human motion to robot motion is more intuitive and offers easy and direct control of the dual-arm robot. However, dual-arm manipulators introduce additional task constraints and self- collision avoidance constraints (especially with coordinated manipulation tasks), making motion mapping from human to robot more complex. In this thesis, we design, implement, integrate, and evaluate a control architecture for the teleoperation of a dual-arm robot with focus on human-robot motion mapping. This architecture allows for direct and intuitive control of the dual-arm robot. Moreover, it allows flexibility to transfer a degree of control between the robot and the user. The developed architecture can easily integrate task, environmental, and robot hardware constraints while mapping motion between the human and the robot. The use of a hybrid dynamics solver for motion mapping allows these constraints to be imposed on the level of position, velocity, acceleration, as well as force. We present experiments to successfully demonstrate that this architecture can be used to perform single-arm tasks like pressing a button or dual-arm tasks like pouring sand with teleoperation.

  @MastersThesis{ 2018kulkarni,
  abstract  = {Robots can accomplish tasks in environments that are
  hazardous and pose a threat to human safety. However,
  dynamic, unknown, or unstructured environments still
  require a degree of human decision making which autonomous
  robots cannot provide. Teleoperation of robots with a human
  making decisions offers a solution. Dual-arm robots provide
  an added advantage in these situations as they offer the
  possibility to accomplish a task in a manner similar to
  humans. Additionally, teleoperation with direct transfer of
  human motion to robot motion is more intuitive and offers
  easy and direct control of the dual-arm robot. However,
  dual-arm manipulators introduce additional task constraints
  and self- collision avoidance constraints (especially with
  coordinated manipulation tasks), making motion mapping from
  human to robot more complex. In this thesis, we design,
  implement, integrate, and evaluate a control architecture
  for the teleoperation of a dual-arm robot with focus on
  human-robot motion mapping. This architecture allows for
  direct and intuitive control of the dual-arm robot.
  Moreover, it allows flexibility to transfer a degree of
  control between the robot and the user. The developed
  architecture can easily integrate task, environmental, and
  robot hardware constraints while mapping motion between the
  human and the robot. The use of a hybrid dynamics solver
  for motion mapping allows these constraints to be imposed
  on the level of position, velocity, acceleration, as well
  as force. We present experiments to successfully
  demonstrate that this architecture can be used to perform
  single-arm tasks like pressing a button or dual-arm tasks
  like pouring sand with teleoperation.},
  address  = {Wilhelm-Levison-Str. 22, Bonn, 53115, Germany},
  annote  = {WS14/15 Pl{\"o}ger, Bennewitz, Schulz, Schneider},
  author  = {Padmaja Kulkarni},
  month = {October},
  school  = {Hochschule Bonn-Rhein-Sieg},
  title = {nthropomorphic motion mapping under constraints for direct
  and intuitive control of a dual-arm manipulator},
  year = {2018}
  }

• N. R. Koripalli, “Strategies to pre-train deep models for efficient multi-task fine-tuning,” Master Thesis, Grantham-Allee 20, 53757 St. Augustin, Germany, 2018.
  [BibTeX] [Abstract]
  

  Deep Convolutional Neural Networks have made incredible leaps in the realm of image classification but they do not lend themselves to multi-task scenarios. Although they generalize quite well for the dataset that they are trained on, they do not generalize well for other image classification tasks where the dataset is not in the same domain as the dataset that it has been trained on. Transfer learning merely shifts the starting weights of the DCNN to a location that is biased towards the information within one dataset. The reason transfer learning works is because most datasets have similar information content as the original dataset that the DCNN was trained on. The problem is not with the DCNNs itself but with the optimizers that have fixed hyper-parameters. We apply principles of meta-learning, where the gradients of gradients are learned in the optimization process and sometimes the parameters of the optimizers are learned giving the DCNN the freedom to learn weights that lend themselves to being adapted to new tasks much more readily. We show that such a process can lead to reduced fine-tuning times for unseen tasks and datasets and the model can handle scenarios where the dataset is quite small.

  @MastersThesis{ 2018koripalli,
  abstract  = {Deep Convolutional Neural Networks have made incredible
  leaps in the realm of image classification but they do not
  lend themselves to multi-task scenarios. Although they
  generalize quite well for the dataset that they are trained
  on, they do not generalize well for other image
  classification tasks where the dataset is not in the same
  domain as the dataset that it has been trained on. Transfer
  learning merely shifts the starting weights of the DCNN to
  a location that is biased towards the information within
  one dataset. The reason transfer learning works is because
  most datasets have similar information content as the
  original dataset that the DCNN was trained on. The problem
  is not with the DCNNs itself but with the optimizers that
  have fixed hyper-parameters. We apply principles of
  meta-learning, where the gradients of gradients are learned
  in the optimization process and sometimes the parameters of
  the optimizers are learned giving the DCNN the freedom to
  learn weights that lend themselves to being adapted to new
  tasks much more readily. We show that such a process can
  lead to reduced fine-tuning times for unseen tasks and
  datasets and the model can handle scenarios where the
  dataset is quite small.},
  address  = {Grantham-Allee 20, 53757 St. Augustin, Germany},
  annote  = {SS15 BRSU - Recogizer GmbH Plöger, Hinkenjann, Zimmerling
  supervising},
  author  = {Nitish Reddy Koripalli},
  month = {July},
  school  = {Bonn-Rhein-Sieg University of Applied Sciences},
  title = {Strategies to pre-train deep models for efficient
  multi-task fine-tuning},
  year = {2018}
  }

• A. Ajmera, “Estimation of Prediction Uncertainty for Semantic Scene Labeling Using Bayesian Approximation,” Master Thesis, Grantham-Allee 20, 53757 St. Augustin, Germany, 2018.
  [BibTeX] [Abstract]
  

  With the advancement in technology, autonomous and assisted driving are close to being reality. A key component of such systems is the understanding of the surrounding environment. This understanding about the environment can be attained by performing semantic labeling of the driving scenes. Existing deep learning based models have been developed over the years that outperform classical image processing algorithms for the task of semantic labeling. However, the existing models only produce semantic predictions and do not provide a measure of uncertainty about the predictions. Hence, this work focuses on developing a deep learning based semantic labeling model that can produce semantic predictions and their corresponding uncertainties. Autonomous driving needs a real-time operating model, however the Full Resolution Residual Network (FRRN) [4] architecture, which is found as the best performing architecture during literature search, is not able to satisfy this condition. Hence, a small network, similar to FRRN, has been developed and used in this work. Based on the work of [13], the developed network is then extended by adding dropout layers and the dropouts are used during testing to perform approximate Bayesian inference. The existing works on uncertainties, do not have quantitative metrics to evaluate the quality of uncertainties estimated by a model. Hence, the area under curve (AUC) of the receiver operating characteristic (ROC) curves is proposed and used as an evaluation metric in this work. Further, a comparative analysis about the influence of dropout layer position, drop probability and the number of samples, on the quality of uncertainty estimation is performed. Finally, based on the insights gained from the analysis, a model with optimal configuration of dropout is developed. It is then evaluated on the Cityscape dataset and shown to be outperforming the baseline model with an AUC-ROC of about 90%, while the latter having AUC-ROC of about 80%.

  @MastersThesis{ 2018ajmera,
  abstract  = {With the advancement in technology, autonomous and
  assisted driving are close to being reality. A key
  component of such systems is the understanding of the
  surrounding environment. This understanding about the
  environment can be attained by performing semantic labeling
  of the driving scenes. Existing deep learning based models
  have been developed over the years that outperform
  classical image processing algorithms for the task of
  semantic labeling. However, the existing models only
  produce semantic predictions and do not provide a measure
  of uncertainty about the predictions. Hence, this work
  focuses on developing a deep learning based semantic
  labeling model that can produce semantic predictions and
  their corresponding uncertainties.
  Autonomous driving needs a real-time operating model,
  however the Full Resolution Residual Network (FRRN) [4]
  architecture, which is found as the best performing
  architecture during literature search, is not able to
  satisfy this condition. Hence, a small network, similar to
  FRRN, has been developed and used in this work. Based on
  the work of [13], the developed network is then extended by
  adding dropout layers and the dropouts are used during
  testing to perform approximate Bayesian inference. The
  existing works on uncertainties, do not have quantitative
  metrics to evaluate the quality of uncertainties estimated
  by a model. Hence, the area under curve (AUC) of the
  receiver operating characteristic (ROC) curves is proposed
  and used as an evaluation metric in this work. Further, a
  comparative analysis about the influence of dropout layer
  position, drop probability and the number of samples, on
  the quality of uncertainty estimation is performed.
  Finally, based on the insights gained from the analysis, a
  model with optimal configuration of dropout is developed.
  It is then evaluated on the Cityscape dataset and shown to
  be outperforming the baseline model with an AUC-ROC of
  about 90%, while the latter having AUC-ROC of about 80%.},
  address  = {Grantham-Allee 20, 53757 St. Augustin, Germany},
  annote  = {WS15/16, Fraunhofer IAIS Ploeger, Herpers, Eickeler
  supervising},
  author  = {Anand Ajmera},
  month = {January},
  school  = {Bonn-Rhein-Sieg University of Applied Sciences},
  title = {Estimation of Prediction Uncertainty for Semantic Scene
  Labeling Using Bayesian Approximation},
  year = {2018}
  }

2017

• S. S. Khan and J. Hoey, “Review of Fall Detection Techniques: A Data Availability Perspective,” Medical Engineering and Physics, vol. 39, p. 12 – 22, 2017.
  [BibTeX]

  @Article{ khan2017,
  archiveprefix  = {arXiv},
  arxivid  = {1605.09351},
  author  = {Khan, Shehroz S. and Hoey, Jesse},
  title = {{Review of Fall Detection Techniques: A Data Availability
  Perspective}},
  journal  = {Medical Engineering and Physics},
  volume  = {39},
  pages = {12 -- 22},
  year = {2017},
  issn = {1350-4533}
  }

• D. Tran, J. Ray, Z. Shou, S. Chang, and M. Paluri, “ConvNet Architecture Search for Spatiotemporal Feature Learning,” CoRR, 2017.
  [BibTeX]

  @Article{ du2017,
  author  = {Tran, Du and Ray, Jamie and Shou, Zheng and Chang,
  Shih{-}Fu and Paluri, Manohar},
  title = {{ConvNet Architecture Search for Spatiotemporal Feature
  Learning}},
  journal  = {CoRR},
  year = {2017}
  }

• A. Núñez-Marcos, G. Azkune, and I. Arganda-Carreras, “Vision-Based Fall Detection with Convolutional Neural Networks,” Wireless Communications and Mobile Computing, vol. 2017, p. 1–16, 2017.
  [BibTeX]

  @Article{ nunez2017,
  author  = {N{\'{u}}{\~{n}}ez-Marcos, Adri{\'{a}}n and Azkune, Gorka
  and Arganda-Carreras, Ignacio},
  title = {{Vision-Based Fall Detection with Convolutional Neural
  Networks}},
  journal  = {Wireless Communications and Mobile Computing},
  volume  = {2017},
  year = {2017},
  pages = {1--16},
  issn = {1530-8669}
  }

• L. Perez and J. Wang, “The Effectiveness of Data Augmentation in Image Classification using Deep Learning,” CoRR, 2017.
  [BibTeX]

  @Article{ perez2017,
  author  = {Perez, Luis and Wang, Jason},
  year = {2017},
  title = {{The Effectiveness of Data Augmentation in Image
  Classification using Deep Learning}},
  journal  = {CoRR}
  }

• H. Idrees, A. R. Zamir, Y. Jiang, A. Gorban, I. Laptev, R. Sukthankar, and M. Shah, “The THUMOS Challenge on Action Recognition for Videos “in the Wild”,” Computer Vision and Image Understanding, 2017.
  [BibTeX]

  @Article{ zamir2019,
  author  = {Idrees, Haroon and Zamir, Amir R. and Jiang, Yu-Gang and
  Gorban, Alex and Laptev, Ivan and Sukthankar, Rahul and
  Shah, Mubarak},
  year = {2017},
  title = {{The {THUMOS} Challenge on Action Recognition for Videos
  "in the Wild"}},
  journal  = {Computer Vision and Image Understanding}
  }

• M. Schoebel, “Deep Learning for Recognition of Daily-Living Actions,” Master Thesis, Rheinaustraße 32, 53225 Bonn, Germany, 2017.
  [BibTeX] [Abstract]
  

  With the world’s population growing older and caring facilities having reached their capacities already today, the need for assistive (robotic) systems is specifically critical in elderly care. In order to aid, an autonomous assistive system first needs to know in what form and when help is required. One way to achieve this is to recognize the currently performed actions by a human user and offer assistance accordingly. This work studies \textit{temporal order verification}, a quasi unsupervised pre-training method for improving the recognition of daily living actions from video data, using the \textit{C3D} deep convolutional neural network. In particular, we evaluate \textit{temporal order verification} as a means to incorporate motion sensitivity in 3D convolutional neural networks, which otherwise treat the temporal evolution of a video and the spatial dimension of video frames equally. Since pre-training a network in an unsupervised way does not require labelled data, this method can be specifically beneficial for recognition tasks where large amounts of labelled data are sparse. We focus on the Charades dataset, which features a unique yet small amount of mundane daily-living action videos and therefore enables the design of vision-based systems, which can be deployed in real-world applications such as assistive robotics. Our \textit{C3D} implementation yields an accuracy of $44.57\%$ on the UCF101 standard action recognition benchmark and a mean average precision of $9.01\%$ for recognizing daily-living actions of the Charades dataset, when trained from randomly initialized weights. By pre-training the model with \textit{temporal order verification}, we were not able to improve classification results, in fact the pre-trained weights turned out to impair the network’s performance.

  @MastersThesis{ yyyylastname,
  abstract  = { With the world's population growing older and caring
  facilities having reached their capacities already today,
  the need for assistive (robotic) systems is specifically
  critical in elderly care. In order to aid, an autonomous
  assistive system first needs to know in what form and when
  help is required. One way to achieve this is to recognize
  the currently performed actions by a human user and offer
  assistance accordingly. This work studies \textit{temporal
  order verification}, a quasi unsupervised pre-training
  method for improving the recognition of daily living
  actions from video data, using the \textit{C3D} deep
  convolutional neural network. In particular, we evaluate
  \textit{temporal order verification} as a means to
  incorporate motion sensitivity in 3D convolutional neural
  networks, which otherwise treat the temporal evolution of a
  video and the spatial dimension of video frames equally.
  Since pre-training a network in an unsupervised way does
  not require labelled data, this method can be specifically
  beneficial for recognition tasks where large amounts of
  labelled data are sparse. We focus on the Charades dataset,
  which features a unique yet small amount of mundane
  daily-living action videos and therefore enables the design
  of vision-based systems, which can be deployed in
  real-world applications such as assistive robotics. Our
  \textit{C3D} implementation yields an accuracy of $44.57\%$
  on the UCF101 standard action recognition benchmark and a
  mean average precision of $9.01\%$ for recognizing
  daily-living actions of the Charades dataset, when trained
  from randomly initialized weights. By pre-training the
  model with \textit{temporal order verification}, we were
  not able to improve classification results, in fact the
  pre-trained weights turned out to impair the network's
  performance.},
  address  = {Rheinaustraße 32, 53225 Bonn, Germany},
  annote  = {WS 2015/16, RoboLand Project, Prassler, Plöger
  supervising},
  author  = {Maximilian Schoebel},
  month = {October},
  school  = {Bonn-Rhein-Sieg University of Applied Sciences},
  title = {Deep Learning for Recognition of Daily-Living Actions},
  year = {2017}
  }

• W. Kay, J. Carreira, K. Simonyan, B. Zhang, C. Hillier, S. Vijayanarasimhan, F. Viola, T. Green, T. Back, P. Natsev, M. Suleyman, and A. Zisserman, “The Kinetics Human Action Video Dataset,” CoRR, 2017.
  [BibTeX]

  @Article{ carreira2017,
  author  = {Kay,Will and Carreira, Jo{\~{a}}o and Simonyan, Karen and
  Zhang, Brian and Hillier, Chloe and Vijayanarasimhan,
  Sudheendra and Viola, Fabio and Green, Tim and Back, Trevor
  and Natsev, Paul and Suleyman, Mustafa and Zisserman,
  Andrew},
  title = {{The Kinetics Human Action Video Dataset}},
  year = {2017},
  journal  = {CoRR}
  }

• C. Quignon, “Simultaneous Estimation of Rewards and Dynamics in Continuous Environments,” Master Thesis, Grantham-Allee 20, 53757 St. Augustin, Germany, 2017.
  [BibTeX] [Abstract]
  

  The field of robotics is steadily progressing into more diverse settings and increasingly complex tasks. Often, the exact environment that a robot will be applied in is not known nor is it possible to accurately specify solutions to the desired tasks. Instead of defining a rule based solution, Inverse Reinforcement Learning (IRL) provides an approach to learn a task from expert demonstrations by estimating the reward function of a Markov Decision Process. This reward function can be used to infer a policy solving the task even in new environments. The current state of the art in research often assumes the dynamics of the environment to be given. This assumption can hardly be satisfied in real world problems. This thesis introduces a novel approach to simultaneously estimate the reward function and the dynamics from a limited set of expert demonstrations in continuous state and action spaces. The approach is compared to Relative Entropy IRL (REIRL) with a naive dynamics estimate. The experimental evaluation showed that the simultaneous estimation of the dynamics is more accurate than the naive, fixed estimate. Since both the rewards and the dynamics influence the policy, it is plausible that a more expressive reward function can counteract inaccuracies of the dynamics estimate. To test this, the reward function is modified with additional features. Although both approaches can benefit from additional features, a careful choice of the features and initial values is required. The additional degree of freedom from the added features may result in an a worse dynamics estimate, overfitting or an ambiguous reward function. A theoretic explanation on the latter is also provided.

  @MastersThesis{ 2017quignon,
  abstract  = {The field of robotics is steadily progressing into more
  diverse settings and increasingly complex tasks. Often, the
  exact environment that a robot will be applied in is not
  known nor is it possible to accurately specify solutions to
  the desired tasks. Instead of defining a rule based
  solution, Inverse Reinforcement Learning (IRL) provides an
  approach to learn a task from expert demonstrations by
  estimating the reward function of a Markov Decision
  Process. This reward function can be used to infer a policy
  solving the task even in new environments. The current
  state of the art in research often assumes the dynamics of
  the environment to be given. This assumption can hardly be
  satisfied in real world problems.
  This thesis introduces a novel approach to simultaneously
  estimate the reward function and the dynamics from a
  limited set of expert demonstrations in continuous state
  and action spaces. The approach is compared to Relative
  Entropy IRL (REIRL) with a naive dynamics estimate. The
  experimental evaluation showed that the simultaneous
  estimation of the dynamics is more accurate than the naive,
  fixed estimate. Since both the rewards and the dynamics
  influence the policy, it is plausible that a more
  expressive reward function can counteract inaccuracies of
  the dynamics estimate. To test this, the reward function is
  modified with additional features. Although both approaches
  can benefit from additional features, a careful choice of
  the features and initial values is required. The additional
  degree of freedom from the added features may result in an
  a worse dynamics estimate, overfitting or an ambiguous
  reward function. A theoretic explanation on the latter is
  also provided.},
  address  = {Grantham-Allee 20, 53757 St. Augustin, Germany},
  annote  = {SS14 Robert Bosch GmbH Herman, Kraetzschmar, M{\"u}ller,
  supervising},
  author  = {Christophe Quignon},
  month = {February},
  school  = {Bonn-Rhein-Sieg University of Applied Sciences},
  title = {Simultaneous Estimation of Rewards and Dynamics in
  Continuous Environments},
  year = {2017}
  }

• D. Arya, “Complete Path Coverage while Exploring Unknown Environment for Radiation Mapping using Unmanned Ground Systems,” Master Thesis, Grantham-Allee 20, 53757 St. Augustin, Germany, 2017.
  [BibTeX] [Abstract]
  

  In the recent past, several events have highlighted the need for unmanned autonomous systems in disaster situations caused by natural hazards or due to human-driven actions. The use of autonomous systems for the first analysis of the environment and the provision of the initial report can limit hazards to the first responders. This can increase their efficiency in planning response operations and reduce the reaction time. In case of a nuclear disaster, an autonomous system that could gather radiation measurements of the site and provide a radiation map of the area, would be of great help for the response team. Such radiation maps would assist the response team to plan their actions before beginning recovery activities, and thus reduce the risk of putting the life of the recovery team in danger. These radiation maps will also allow them to understand the extent of damage and identify regions where it will be dangerous for the response team to go. In this work, we have developed an online exploration method, Simultaneous Meandering and Mapping (SMAM), for unknown environments which enables a ground robot to cover every free space in the region and gather radiation information. The robot covers this free space effectively without revisiting a space unless it is necessary or unavoidable. Such kind of exploration is called Complete Coverage Path Planning (CCPP). In the proposed algorithm, the robot tries to sweep the region in zig-zag motion and optimizes its path to have maximal straight trajectories. The algorithm stores the environment information regarding obstacles, free spaces and radiation measurement, which is later used to generate a radiation and environment map. We visualized this radiation information as a heat-map laid over the generated environment map and physical layout of the area. We tested SMAM against two existing online CCPP exploration techniques. We chose scenarios with varying complexity from widely open to cluttered environments and analyzed the performance of the three methods in terms of total number of revisits, turns and path length. SMAM has performed better and more efficiently than the other two for complex environments. SMAM can also generate continuous straight trajectories preventing the robot to stop at every position to plan for new goal, and hence save power consumption.

  @MastersThesis{ 2017arya,
  abstract  = {In the recent past, several events have highlighted the
  need for unmanned autonomous systems in disaster situations
  caused by natural hazards or due to human-driven actions.
  The use of autonomous systems for the first analysis of the
  environment and the provision of the initial report can
  limit hazards to the first responders. This can increase
  their efficiency in planning response operations and reduce
  the reaction time. In case of a nuclear disaster, an
  autonomous system that could gather radiation measurements
  of the site and provide a radiation map of the area, would
  be of great help for the response team. Such radiation maps
  would assist the response team to plan their actions before
  beginning recovery activities, and thus reduce the risk of
  putting the life of the recovery team in danger. These
  radiation maps will also allow them to understand the
  extent of damage and identify regions where it will be
  dangerous for the response team to go.
  In this work, we have developed an online exploration
  method, Simultaneous Meandering and Mapping (SMAM), for
  unknown environments which enables a ground robot to cover
  every free space in the region and gather radiation
  information. The robot covers this free space effectively
  without revisiting a space unless it is necessary or
  unavoidable. Such kind of exploration is called Complete
  Coverage Path Planning (CCPP). In the proposed algorithm,
  the robot tries to sweep the region in zig-zag motion and
  optimizes its path to have maximal straight trajectories.
  The algorithm stores the environment information regarding
  obstacles, free spaces and radiation measurement, which is
  later used to generate a radiation and environment map. We
  visualized this radiation information as a heat-map laid
  over the generated environment map and physical layout of
  the area.
  We tested SMAM against two existing online CCPP exploration
  techniques. We chose scenarios with varying complexity from
  widely open to cluttered environments and analyzed the
  performance of the three methods in terms of total number
  of revisits, turns and path length. SMAM has performed
  better and more efficiently than the other two for complex
  environments. SMAM can also generate continuous straight
  trajectories preventing the robot to stop at every position
  to plan for new goal, and hence save power consumption.},
  address  = {Grantham-Allee 20, 53757 St. Augustin, Germany},
  annote  = {WS13/14 Asteroth, M\"uller, Schneider supervising},
  author  = {Devvrat Arya},
  month = {June},
  school  = {Bonn-Rhein-Sieg University of Applied Sciences},
  title = {Complete Path Coverage while Exploring Unknown Environment
  for Radiation Mapping using Unmanned Ground Systems},
  year = {2017}
  }

• M. S. Abubucker, “Humanoid Whole-Body Motion Planning for Locomotion Synchronized with Manipulation Task,” Master Thesis, Grantham-Allee 20, 53757 St. Augustin, Germany, 2017.
  [BibTeX] [Abstract]
  

  In this work, we consider the problem of whole-body motion planning for humanoid that should perform a task by using both the locomotion task and the manipulation task in a synchronized manner. This would allow us to perform smooth reaching motions. The task assigned to the robot would be to reach a goal position with its hand which is quite far away from the robot. In order to reach the goal position, the robot must execute a manipulation task while performing a locomotion task implicitly. The presented whole-body motion planner builds solution by merging CoM movement primitives in the configuration space. The CoM movement primitives are typical humanoid actions like walking with forward step, lateral steps and curved steps. To enable smooth reaching motions, three zones are defined: locomotion zone, loco-manipulation zone and manipulation zone. The planner has been tested for the HRP-4 robot in V-REP simulator. Different types of task were performed with different levels of difficulties to evaluate the planner. The simulation results show that the planner was able to find the solution regardless of the obstacles in the scene and the performance of the planner was evaluated using performance metrics like time taken to plan, tree size and motion duration.

  @MastersThesis{ 2017abubucker,
  abstract  = { In this work, we consider the problem of whole-body
  motion planning for humanoid that should perform a task by
  using both the locomotion task and the manipulation task in
  a synchronized manner. This would allow us to perform
  smooth reaching motions. The task assigned to the robot
  would be to reach a goal position with its hand which is
  quite far away from the robot. In order to reach the goal
  position, the robot must execute a manipulation task while
  performing a locomotion task implicitly. The presented
  whole-body motion planner builds solution by merging CoM
  movement primitives in the configuration space.
  The CoM movement primitives are typical humanoid actions
  like walking with forward step, lateral steps and curved
  steps. To enable smooth reaching motions, three zones are
  defined: locomotion zone, loco-manipulation zone and
  manipulation zone. The planner has been tested for the
  HRP-4 robot in V-REP simulator. Different types of task
  were performed with different levels of difficulties to
  evaluate the planner. The simulation results show that the
  planner was able to find the solution regardless of the
  obstacles in the scene and the performance of the planner
  was evaluated using performance metrics like time taken to
  plan, tree size and motion duration. },
  address  = {Grantham-Allee 20, 53757 St. Augustin, Germany},
  annote  = {SS13 H-BRS , DIAG Sapienza University of Rome - COMANOID
  Pl{\"o}ger, Oriolo, Cognetti supervising},
  author  = {Mohammed Shameer Abubucker},
  month = {December},
  school  = {Bonn-Rhein-Sieg University of Applied Sciences},
  title = {Humanoid Whole-Body Motion Planning for Locomotion
  Synchronized with Manipulation Task},
  year = {2017}
  }

• A. Drak, “DoveCopter: A Tracking and Following Aerial Platform for Aerodynamics Analysis of a Velomobile,” Master Thesis, Grantham-Allee 20, 53757 St. Augustin, Germany, 2017.
  [BibTeX] [Abstract]
  

  Efficient control and design of electrically-assisted vehicles is a well established branch in the field of intelligent transportation which remains under active and constant development. A testbed for such a field is the electrically assisted vehicle “velomobile”. The velomobile has an aerodynamic body fabricated from light weight carbon fiber, two steered front wheels, and an electric motor to assist the rider. Optimization of the body is actively sought and can be achieved through extracting airflow models from tufts attached to the surface of the vehicle. The aim of this project is to provide high quality (HQ) video footage of the tufts under realistic outdoor conditions in order to extract airflow models. Conventionally, airflow models are extracted by means of a wind tunnel or computational fluid dynamics simulation. Woefully however, the former lacks a realistic test environment and the latter comes at a higher computational cost and reduced accuracy. The proposed method to acquire HQ video footage of tufts attached to the velomobile is by means of a drone. The drone will autonomously perform a side-by-side tracking and following of the vehicle while recording HQ video of the tufts. This method has the added benefit of utilizing all the surrounding space to traverse with no limits, thus eliminating the limited recording space restriction caused by side-byside tracking

  @MastersThesis{ 2017drak,
  abstract  = {Efficient control and design of electrically-assisted
  vehicles is a well established branch in the field of
  intelligent transportation which remains under active and
  constant development. A testbed for such a field is the
  electrically assisted vehicle "velomobile". The velomobile
  has an aerodynamic body fabricated from light weight carbon
  fiber, two steered front wheels, and an electric motor to
  assist the rider. Optimization of the body is actively
  sought and can be achieved through extracting airflow
  models from tufts attached to the surface of the vehicle.
  The aim of this project is to provide high quality (HQ)
  video footage of the tufts under realistic outdoor
  conditions in order to extract airflow models.
  Conventionally, airflow models are extracted by means of a
  wind tunnel or computational fluid dynamics simulation.
  Woefully however, the former lacks a realistic test
  environment and the latter comes at a higher computational
  cost and reduced accuracy. The proposed method to acquire
  HQ video footage of tufts attached to the velomobile is by
  means of a drone. The drone will autonomously perform a
  side-by-side tracking and following of the vehicle while
  recording HQ video of the tufts. This method has the added
  benefit of utilizing all the surrounding space to traverse
  with no limits, thus eliminating the limited recording
  space restriction caused by side-byside tracking},
  address  = {Grantham-Allee 20, 53757 St. Augustin, Germany},
  annote  = {[WS2014] [HBRS] [Asteroth], [Julier], [Kruijff]
  supervising},
  author  = {Ahmad Drak},
  month = {July},
  school  = {Bonn-Rhein-Sieg University of Applied Sciences},
  title = {DoveCopter: A Tracking and Following Aerial Platform for
  Aerodynamics Analysis of a Velomobile},
  year = {2017}
  }

• C. Hebbal, “Developemnt of learning lidar sensor model,” Master Thesis, Grantham-Allee 20, 53757 St. Augustin, Germany, 2017.
  [BibTeX] [Abstract]
  

  In recent times, academia and industry alike are working in a big way on autonomous vehicles, with the intention of making the future of transportation safe and efficient. Some more time is needed, before we see fully autonomous vehicles on road, but the journey towards it has already begun with the introduction of ADAS and HAD functions in vehicles that enable them to drive around autonomously in selected scenarios like parking, highway driving etc. The key requirement of these systems is comprehensive and accurate information about the surrounding environment of the vehicle. This information is usually represented by maps. There exists numerous approaches to represent environment in literature, among them occupancy grid representation is the popular and most widely used. Occupancy grids tessellate the area to be mapped by means of finite evenly spaced grid of cells. Each cell in the grid is associated with a binary variable that specifies the occupancy value of the cell. Usually to construct maps, occupancy grids make use of inverse sensor models to transform sensor measurements to occupancy values. Most of the existing grid approaches in literature make use of simplified inverse sensor models, which results in generating maps of lower quality. This is due to the fact that most of these models do not take into account the uncertainty in measurement and environmental factors influencing measurement. Thus, in this thesis we set out to develop a better stochastic inverse sensor model by using classical machine learning technique. We developed the stochastic inverse sensor model for the LIDAR sensor by using regression trees. The developed sensor model was evaluated against existing models in simulation and was found to generate better quality maps as compared to existing models. The model was also integrated into EB robinos framework and was evaluated against existing highly tuned static model on measurement data. The quality of map generated by developed model was found to be better than the static model.

  @MastersThesis{ 2017hebbal,
  abstract  = {In recent times, academia and industry alike are working
  in a big way on autonomous vehicles, with the intention of
  making the future of transportation safe and efficient.
  Some more time is needed, before we see fully autonomous
  vehicles on road, but the journey towards it has already
  begun with the introduction of ADAS and HAD functions in
  vehicles that enable them to drive around autonomously in
  selected scenarios like parking, highway driving etc. The
  key requirement of these systems is comprehensive and
  accurate information about the surrounding environment of
  the vehicle. This information is usually represented by
  maps.
  There exists numerous approaches to represent environment
  in literature, among them occupancy grid representation is
  the popular and most widely used. Occupancy grids
  tessellate the area to be mapped by means of finite evenly
  spaced grid of cells. Each cell in the grid is associated
  with a binary variable that specifies the occupancy value
  of the cell. Usually to construct maps, occupancy grids
  make use of inverse sensor models to transform sensor
  measurements to occupancy values. Most of the existing grid
  approaches in literature make use of simplified inverse
  sensor models, which results in generating maps of lower
  quality. This is due to the fact that most of these models
  do not take into account the uncertainty in measurement and
  environmental factors influencing measurement.
  Thus, in this thesis we set out to develop a better
  stochastic inverse sensor model by using classical machine
  learning technique. We developed the stochastic inverse
  sensor model for the LIDAR sensor by using regression
  trees. The developed sensor model was evaluated against
  existing models in simulation and was found to generate
  better quality maps as compared to existing models. The
  model was also integrated into EB robinos framework and was
  evaluated against existing highly tuned static model on
  measurement data. The quality of map generated by developed
  model was found to be better than the static model.},
  address  = {Grantham-Allee 20, 53757 St. Augustin, Germany},
  annote  = {WS 15/16 Elektrobit Automotive Gmbh – EB Robinos
  Pl{\”o}ger, Asterorth, Tollk{\”u}hn supervising},
  author  = {Chaitanya Hebbal},
  month = {December},
  school  = {Bonn-Rhein-Sieg University of Applied Sciences},
  title = {Developemnt of learning lidar sensor model},
  year = {2017}
  }

• P. Lukin, “Comparison of controller auto-tuning methods for manipulator axis,” mathesis Master Thesis, Grantham-Allee 20, 53757 St. Augustin, Germany, 2017.
  [BibTeX] [Abstract]
  

  Robotic arms have a variety of applications in industry and can be programmed to perform specific tasks. Control process of robotic arms is usually reduced to end-effector trajectory following. This task requires position control of all manipulator joints. The current standard control algorithm is PID controller and various modifications of it. Performance of the controller is governed by a set of parameters that are specified during the tuning process. These parameters can drastically change behavior of the system and must be chosen correctly in order to meet demanded controller performance. There are numerous tuning methods that were designed to achieve different control-loop qualities. Additionally, a control engineer must be involved every time a robot is deployed. Thus, it is important to investigate which methods can be successfully applied to tune robotic arm joints in an automatic manner. The main goal of this thesis is to investigate what controller auto-tuning method is the most efficient for manipulator axis position control. The object of the research is the robotic arm axis that includes a joint with an attached load and a control unit. Efficiency of the controller is based on closed-loop control specifications: transient response, robustness and disturbance rejection. The approach is to design a mathematical model of the manipulator axis and obtain model parameters using system identification methods. Next, linearization and model order reduction are applied to the obtained model that is used for comparative analysis of existing PID-based controllers and tuning methods. The most theoretically promising algorithms are applied for controller design given control-loop specification. MATLAB Simulink batch experiments with pseudo-random manipulator parameters performed to evaluate tuning algorithms. Finally, the most effective auto-tuning method is validated on board of BLDC motor yielding good results and concurrence with theoretical expectations.

  @MastersThesis{ 2017lukin,
  author  = {Petr Lukin},
  title = {Comparison of controller auto-tuning methods for
  manipulator axis},
  school  = {Hochschule Bonn-Rhein-Sieg},
  year = {2017},
  type = {mathesis},
  address  = {Grantham-Allee 20, 53757 St. Augustin, Germany},
  month = nov,
  abstract  = {Robotic arms have a variety of applications in industry
  and can be programmed to perform specific tasks. Control
  process of robotic arms is usually reduced to end-effector
  trajectory following. This task requires position control
  of all manipulator joints. The current standard control
  algorithm is PID controller and various modifications of
  it. Performance of the controller is governed by a set of
  parameters that are specified during the tuning process.
  These parameters can drastically change behavior of the
  system and must be chosen correctly in order to meet
  demanded controller performance. There are numerous tuning
  methods that were designed to achieve different
  control-loop qualities. Additionally, a control engineer
  must be involved every time a robot is deployed. Thus, it
  is important to investigate which methods can be
  successfully applied to tune robotic arm joints in an
  automatic manner.
  The main goal of this thesis is to investigate what
  controller auto-tuning method is the most efficient for
  manipulator axis position control. The object of the
  research is the robotic arm axis that includes a joint with
  an attached load and a control unit. Efficiency of the
  controller is based on closed-loop control specifications:
  transient response, robustness and disturbance rejection.
  The approach is to design a mathematical model of the
  manipulator axis and obtain model parameters using system
  identification methods. Next, linearization and model order
  reduction are applied to the obtained model that is used
  for comparative analysis of existing PID-based controllers
  and tuning methods. The most theoretically promising
  algorithms are applied for controller design given
  control-loop specification. MATLAB Simulink batch
  experiments with pseudo-random manipulator parameters
  performed to evaluate tuning algorithms. Finally, the most
  effective auto-tuning method is validated on board of BLDC
  motor yielding good results and concurrence with
  theoretical expectations. },
  annote  = {WS15/16 Synapticon GmbH, Pl{\"o}ger, Chakirov supervising}
  }

• S. Biswas, “Deployment of a Correlation Based Algorithm for a Robotic Black Box,” Master Thesis, Grantham-Allee 20, 53757 St. Augustin, Germany, 2017.
  [BibTeX] [Abstract]
  

  Robotic systems are vulnerable to faults. If faults in robot systems are not handled quickly enough it might lead to a catastrophe. Hence real-time fault detection and diagnosis on robots are important. Since robots are desired to work autonomously there is usually no human to maintain the robot instantly. This necessitates a fault detection and diagno- sis module on the robot. Again this module itself should not induce fault in the robot system and kept separate from the robot system as much as possible. Hence the idea of a black box is conceived, which would act as an external fault observer to the robot system. Considering the requirement of real-time operation, computationally light ap- proaches should be implemented. Since correlation measure gives hint about how system components might be interdependent on the same system variables, an approach based on correlation is investigated in this work. In order to detect fault the observable sensors and actuators values along with structural model of the system are available to the detection module. This works tries to keep the correlation based approach as general as possible. Finally after investigating deeply, an approach based on correlation change is proposed which is able to detect more faults. Upon detection of fault a simple diagnosis is done based on the knowledge of structural model of the robot system.

  @MastersThesis{ 2008biswas,
  abstract  = {Robotic systems are vulnerable to faults. If faults in
  robot systems are not handled quickly enough it might lead
  to a catastrophe. Hence real-time fault detection and
  diagnosis on robots are important. Since robots are desired
  to work autonomously there is usually no human to maintain
  the robot instantly. This necessitates a fault detection
  and diagno- sis module on the robot. Again this module
  itself should not induce fault in the robot system and kept
  separate from the robot system as much as possible. Hence
  the idea of a black box is conceived, which would act as an
  external fault observer to the robot system. Considering
  the requirement of real-time operation, computationally
  light ap- proaches should be implemented. Since correlation
  measure gives hint about how system components might be
  interdependent on the same system variables, an approach
  based on correlation is investigated in this work. In order
  to detect fault the observable sensors and actuators values
  along with structural model of the system are available to
  the detection module. This works tries to keep the
  correlation based approach as general as possible. Finally
  after investigating deeply, an approach based on
  correlation change is proposed which is able to detect more
  faults. Upon detection of fault a simple diagnosis is done
  based on the knowledge of structural model of the robot
  system.},
  address  = {Grantham-Allee 20, 53757 St. Augustin, Germany},
  annote  = {[SS13] [BRSU] - [Deployment of a Correlation Based
  Algorithm for a Robotic Black Box] Pl{\"o}ger, Breuer,
  Thoduka supervising},
  author  = {Saugata Biswas},
  month = {September},
  school  = {Bonn-Rhein-Sieg University of Applied Sciences},
  title = {Deployment of a Correlation Based Algorithm for a Robotic
  Black Box},
  year = {2017}
  }

2016

• D. Vázquez, “Video Analysis and Anomaly Detection in Human Gait Patterns from a Fast Moving Camera – Development of an Outdoor Gait Analysis System,” Master Thesis, Grantham-Allee 20, 53757 St. Augustin, Germany, 2016.
  [BibTeX] [Abstract]
  

  Today, running has become an important activity in the life of many people as it mainly has a good impact on their health. Unfortunately, some individuals do not exercise in an appropriate way provoking several injuries to their bodies. Indoor gait analysis has been used to detect and treat abnormal patterns but it has been proven that the performance exhibited in indoor laboratories is not the same as the performance exhibited outdoors. If the gait patterns of the runners could be analyzed in environments in which they perform their outdoor running activities, better diagnostics would be obtained. Therefore, this Master thesis proposes a method where a moving robot could be used to record the runner’s performance by means of a depth camera. The proposed approach relies on a stereo vision system and color markers. Given a frontal stereo view of a subject, the location of the markers can be determined and used to create a 3D model of the human skeleton. Based on this model, several joint angles, and other features are created. Then, a Support Vector Machine is trained to differentiate between normal and abnormal gait patterns. The proposed method showed a classification rate of 76\% proving that frontal gait analysis is feasible.

  @MastersThesis{ 2016vazquezurenad,
  abstract  = {Today, running has become an important activity in the
  life of many people as it mainly has a good impact on their
  health. Unfortunately, some individuals do not exercise in
  an appropriate way provoking several injuries to their
  bodies. Indoor gait analysis has been used to detect and
  treat abnormal patterns but it has been proven that the
  performance exhibited in indoor laboratories is not the
  same as the performance exhibited outdoors. If the gait
  patterns of the runners could be analyzed in environments
  in which they perform their outdoor running activities,
  better diagnostics would be obtained. Therefore, this
  Master thesis proposes a method where a moving robot could
  be used to record the runner's performance by means of a
  depth camera. The proposed approach relies on a stereo
  vision system and color markers. Given a frontal stereo
  view of a subject, the location of the markers can be
  determined and used to create a 3D model of the human
  skeleton. Based on this model, several joint angles, and
  other features are created. Then, a Support Vector Machine
  is trained to differentiate between normal and abnormal
  gait patterns. The proposed method showed a classification
  rate of 76\% proving that frontal gait analysis is
  feasible.},
  address  = {Grantham-Allee 20, 53757 St. Augustin, Germany},
  annote  = {WS2014 - HBRS - Prof. Prassler, Prof. P\"oger, F\"uller},
  author  = {Daniel V\'azquez},
  month = {November},
  school  = {Bonn-Rhein-Sieg University of Applied Sciences},
  title = {Video Analysis and Anomaly Detection in Human Gait
  Patterns from a Fast Moving Camera - Development of an
  Outdoor Gait Analysis System},
  year = {2016}
  }

• S. Thoduka, “Motion Detection for Mobile Robots Using Vision,” Master Thesis, Grantham-Allee 20, 53757 St. Augustin, Germany, 2016.
  [BibTeX] [Abstract]
  

  Motion detection is an important skill for an autonomous mobile robot op- erating in dynamic environments. Motion in the environment could, for instance, indicate the presence of an obstacle, a human attracting the atten- tion of the robot, unwanted disturbance of the environment by the robot etc. Vision-based motion detection is particularly challenging when the robot’s camera is in motion. In addition to detecting independently moving objects while in motion, the robot must be able to distinguish between motions in the environment and motions of its own parts, such as its manipulator. In this thesis, we use a Fourier-Mellin transform-based image registra- tion method to compensate for camera motion before applying two different methods of motion detection: temporal differencing and feature tracking. Self-masking using the robot’s state and model is used to ignore motions of visible robot parts. The approach is evaluated on a set of navigation and manipulation se- quences recorded on a Care-O-bot 3 and a youBot and was also integrated and tested on a real Care-O-bot 3. The image registration method is able to compensate for most of the camera motion, but is still inaccurate at depth discontinuities and when there is large depth variance in the scene. The tem- poral difference method performs better than feature tracking, with a more consistent true positive rate and a lower false discovery rate.

  @MastersThesis{ 2016thoduka,
  abstract  = {Motion detection is an important skill for an autonomous
  mobile robot op- erating in dynamic environments. Motion in
  the environment could, for instance, indicate the presence
  of an obstacle, a human attracting the atten- tion of the
  robot, unwanted disturbance of the environment by the robot
  etc. Vision-based motion detection is particularly
  challenging when the robot’s camera is in motion. In
  addition to detecting independently moving objects while in
  motion, the robot must be able to distinguish between
  motions in the environment and motions of its own parts,
  such as its manipulator. In this thesis, we use a
  Fourier-Mellin transform-based image registra- tion method
  to compensate for camera motion before applying two
  different methods of motion detection: temporal
  differencing and feature tracking. Self-masking using the
  robot’s state and model is used to ignore motions of
  visible robot parts. The approach is evaluated on a set of
  navigation and manipulation se- quences recorded on a
  Care-O-bot 3 and a youBot and was also integrated and
  tested on a real Care-O-bot 3. The image registration
  method is able to compensate for most of the camera motion,
  but is still inaccurate at depth discontinuities and when
  there is large depth variance in the scene. The tem- poral
  difference method performs better than feature tracking,
  with a more consistent true positive rate and a lower false
  discovery rate.},
  address  = {Grantham-Allee 20, 53757 St. Augustin, Germany},
  annote  = {WS13, Pl\"{o}ger, Kraetzschmar, Hegger supervising},
  author  = {Santosh Thoduka},
  month = {September},
  school  = {Bonn-Rhein-Sieg University of Applied Sciences},
  title = {{M}otion {D}etection for {M}obile {R}obots {U}sing
  {V}ision},
  year = {2016}
  }

• C. K. Tan, “Neural Maps for Robotics – A Biologically-Inspired Approach for Obstacle Avoidance,” Master Thesis, Grantham-Allee 20, 53757 St. Augustin, Germany, 2016.
  [BibTeX] [Abstract]
  

  A general aim of autonomous robots is to create agents that can operate independently for long periods of time. One key aspect for them will be the control of movement, since a stationary robot would have less utility than a moving one. One way of fulfilling this aim is to build robots such that they mimic the way humans and animals execute movements, since they are extremely versatile and have the ability to learn and generalize their movements to many situations. This would require studies in neuroscience to discover mechanisms behind such adaptability. We propose the use of Dynamic Neural Fields (DNFs) as a tool to build maps of neural activities, which we term “neural maps” in this thesis, to model what is happening within the brain and nervous system during its operation. For example, they can be used as an internal representation of a robot in the context of localization. Another use of such maps is that they are useful in providing a visual representation correlating neural activity and physical activity, aiding understand of what goes on in the neural networks during execution of a task. In our thesis, we extended a generalized motor program from [Stringer et al. 2003] for the task of obstacle avoidance. This program utilizes memorized motor sequences within a neural network for its movement. By adding a decision component into their model, we show that an agent with such a motor program is able to select memorized trajectories that allows it to avoid obstacle while travelling towards a goal. Simulations show that our model is capable of obstacle avoidance. Our extended model can also be viewed in context of a middle-level controller that utilizes simple rules to decide on an action to take, without the intervention from a high-level planner.

  @MastersThesis{ 2016tan,
  abstract  = {A general aim of autonomous robots is to create agents
  that can operate independently for long periods of time.
  One key aspect for them will be the control of movement,
  since a stationary robot would have less utility than a
  moving one. One way of fulfilling this aim is to build
  robots such that they mimic the way humans and animals
  execute movements, since they are extremely versatile and
  have the ability to learn and generalize their movements to
  many situations. This would require studies in neuroscience
  to discover mechanisms behind such adaptability. We propose
  the use of Dynamic Neural Fields (DNFs) as a tool to build
  maps of neural activities, which we term "neural maps" in
  this thesis, to model what is happening within the brain
  and nervous system during its operation. For example, they
  can be used as an internal representation of a robot in the
  context of localization. Another use of such maps is that
  they are useful in providing a visual representation
  correlating neural activity and physical activity, aiding
  understand of what goes on in the neural networks during
  execution of a task. In our thesis, we extended a
  generalized motor program from [Stringer et al. 2003] for
  the task of obstacle avoidance. This program utilizes
  memorized motor sequences within a neural network for its
  movement. By adding a decision component into their model,
  we show that an agent with such a motor program is able to
  select memorized trajectories that allows it to avoid
  obstacle while travelling towards a goal. Simulations show
  that our model is capable of obstacle avoidance. Our
  extended model can also be viewed in context of a
  middle-level controller that utilizes simple rules to
  decide on an action to take, without the intervention from
  a high-level planner.},
  address  = {Grantham-Allee 20, 53757 St. Augustin, Germany},
  annote  = {WS 2013/14, Pl{\"o}ger, Kraetzschmar, Trappenberg},
  author  = {Chun Kwang Tan},
  month = {January},
  school  = {Bonn-Rhein-Sieg University of Applied Sciences},
  title = {Neural Maps for Robotics - A Biologically-Inspired
  Approach for Obstacle Avoidance},
  year = {2016}
  }

• I. Goodfellow, Y. Bengio, and A. Courville, Deep Learning, MIT Press, 2016.
  [BibTeX]

  @Book{ goodfellow2016,
  author  = {Ian Goodfellow and Yoshua Bengio and Aaron Courville},
  title = {{Deep Learning}},
  publisher  = {MIT Press},
  note = {\url{http://www.deeplearningbook.org}},
  year = {2016}
  }

• L. Wang, Y. Xiong, Z. Wang, Y. Qiao, D. Lin, X. Tang, and L. van Gool, “Temporal Segment Networks: Towards Good Practices for Deep Action Recognition,” in Computer Vision – ECCV 2016, 2016, p. 20–36.
  [BibTeX]

  @InProceedings{ wang2016,
  author  = {Wang, Limin and Xiong, Yuanjun and Wang, Zhe and Qiao, Yu
  and Lin, Dahua and Tang, Xiaoou and van Gool, Luc},
  title = {{Temporal Segment Networks: Towards Good Practices for
  Deep Action Recognition}},
  year = {2016},
  booktitle  = {Computer Vision -- ECCV 2016},
  publisher  = {Springer International Publishing},
  pages = {20--36}
  }

• E. Yildiz, P. {G. Ploeger}, S. Alda, and K. Pervoelz, “Fault-Tolerant Software Architecture for the Unmanned Surface Vehicle Roboship Declaration of Authorship,” Master Thesis, Grantham-Allee 20, 53757 St. Augustin, Germany, 2016.
  [BibTeX] [Abstract]
  

  One of the major goals of research in Robotics is to develop truly autonomous vehicles that can be utilized to assist humans in dangerous environments, especially where human life is under threat. In this regard, Unmanned Surface Vehicles (USVs) are developed and utilized to cope up with military scenarios which put human life at potential risk. One of the very common scenarios is harbor security, where USVs are expected to conduct surveillance and feedback to strengthen the harbor’s security. The (USV) Roboship has been designed and developed for harbor patrolling and coastal surveillance. It goes without saying, in such an application domain dependability plays a major role. USVs must not endanger the mission, face unexpected situations and allow no human intervention in case of a problem. At this point, it should be acknowledged that dependability is strongly coupled with fault tolerance when it comes to autonomous robots. However, due to complicated design schemes of USVs and unpredictable environmental factors of harbor environments, USVs are prone to fail. As it is almost certain that the USVs will fail, in order to not to endanger the mission, USVs should fail in a way that the recovery from the failures is possible, and mission integrity is protected. This paper highlights a fault tolerant software architecture for the USV Roboship. We conduct a comprehensive reliability analysis and propose a fault tolerant scheme, wherein the sensitive points of the existing architecture are spotted, and the fault tolerant techniques are incorporated. Since the approach is model-free and its decisions are made at a high frequency, the system is able to deal with highly dynamic scenarios. We used the UWSim simulation environment to simulate the scenarios in which the USV was supposed to navigate safely using its sensors. Field tests have proven the performance and reliability of the system to be satisfactory, yielding 53.57% decrease in faults. We plan to deploy the USV in Eckernfoerde harbor region of Germany to test the scheme in the future.

  @MastersThesis{ yildiz2016,
  abstract  = {One of the major goals of research in Robotics is to
  develop truly autonomous vehicles that can be utilized to
  assist humans in dangerous environments, especially where
  human life is under threat. In this regard, Unmanned
  Surface Vehicles (USVs) are developed and utilized to cope
  up with military scenarios which put human life at
  potential risk. One of the very common scenarios is harbor
  security, where USVs are expected to conduct surveillance
  and feedback to strengthen the harbor's security. The (USV)
  Roboship has been designed and developed for harbor
  patrolling and coastal surveillance. It goes without
  saying, in such an application domain dependability plays a
  major role. USVs must not endanger the mission, face
  unexpected situations and allow no human intervention in
  case of a problem. At this point, it should be acknowledged
  that dependability is strongly coupled with fault tolerance
  when it comes to autonomous robots. However, due to
  complicated design schemes of USVs and unpredictable
  environmental factors of harbor environments, USVs are
  prone to fail. As it is almost certain that the USVs will
  fail, in order to not to endanger the mission, USVs should
  fail in a way that the recovery from the failures is
  possible, and mission integrity is protected. This paper
  highlights a fault tolerant software architecture for the
  USV Roboship. We conduct a comprehensive reliability
  analysis and propose a fault tolerant scheme, wherein the
  sensitive points of the existing architecture are spotted,
  and the fault tolerant techniques are incorporated. Since
  the approach is model-free and its decisions are made at a
  high frequency, the system is able to deal with highly
  dynamic scenarios. We used the UWSim simulation environment
  to simulate the scenarios in which the USV was supposed to
  navigate safely using its sensors. Field tests have proven
  the performance and reliability of the system to be
  satisfactory, yielding 53.57% decrease in faults. We plan
  to deploy the USV in Eckernfoerde harbor region of Germany
  to test the scheme in the future.},
  address  = {Grantham-Allee 20, 53757 St. Augustin, Germany},
  annote  = {SS13, Fraunhofer Institute IAIS - Roboship, Pl{\"o}ger,
  Alda, Perv{\"o}lz supervising},
  author  = {Yildiz, Erenus and {G. Ploeger}, Paul and Alda, Sascha and
  Pervoelz, Kai},
  month = {October 2016},
  school  = {Bonn-Rhein-Sieg University of Applied Sciences},
  title = {{Fault-Tolerant Software Architecture for the Unmanned
  Surface Vehicle Roboship Declaration of Authorship}},
  year = {2016}
  }

• S. C. Wong, A. Gatt, V. Stamatescu, and M. D. McDonnell, “Understanding Data Augmentation for Classification: When to Warp?,” in 2016 International Conference on Digital Image Computing: Techniques and Applications, DICTA 2016, 2016, p. 1 – 6.
  [BibTeX]

  @InProceedings{ wong2016,
  author  = {Wong, Sebastien C. and Gatt, Adam and Stamatescu, Victor
  and McDonnell, Mark D.},
  booktitle  = {2016 International Conference on Digital Image Computing:
  Techniques and Applications, DICTA 2016},
  title = {{Understanding Data Augmentation for Classification: When
  to Warp?}},
  year = {2016},
  publisher  = {Institute of Electrical and Electronics Engineers Inc.},
  pages = {1 -- 6}
  }

• M. U. Tahir, “Analysis of live 3D mapping approaches with different sensors on various hardware platforms,” Master Thesis, Grantham-Allee 20, 53757 St. Augustin, Germany, 2016.
  [BibTeX] [Abstract]
  

  SLAM is an active research problem and there are various implementations available under ROS (Robot Operating System) which are real-time but because of low range sensors and hardware limitations, the data quality is not as good as in industrial 3D scanners (e.g. FARO Focus3D scanner) which are not real-time though. So, there is a need to have a system which is real-time and could also deliver high data quality under given circumstances. Certain real-time SLAM approaches may deliver higher data quality under certain environments and system specifications. For this, first an overview of the available algorithms and sensors have been provided followed by their evaluation in terms of sensor trajectory estimation (RMSE) and system performance (CPU + memory). “FARO Laser Tracker Vantage” has been employed to serve as the ground truth for the carried out experiments which has an accuracy of up to 0.015mm. After acquiring and analyzing the results from the experiments it has been concluded that “rtabmap” paired with “Asus Xtion Pro Live” on an Intel NUC board is a better option and is more accurate. The proposed system is real-time, efficient and low cost at the same time as compared to other existing low-cost counter-parts.

  @MastersThesis{ 2016tahir,
  abstract  = {SLAM is an active research problem and there are various
  implementations available under ROS (Robot Operating
  System) which are real-time but because of low range
  sensors and hardware limitations, the data quality is not
  as good as in industrial 3D scanners (e.g. FARO Focus3D
  scanner) which are not real-time though. So, there is a
  need to have a system which is real-time and could also
  deliver high data quality under given circumstances.
  Certain real-time SLAM approaches may deliver higher data
  quality under certain environments and system
  specifications. For this, first an overview of the
  available algorithms and sensors have been provided
  followed by their evaluation in terms of sensor trajectory
  estimation (RMSE) and system performance (CPU + memory).
  "FARO Laser Tracker Vantage" has been employed to serve as
  the ground truth for the carried out experiments which has
  an accuracy of up to 0.015mm. After acquiring and analyzing
  the results from the experiments it has been concluded that
  "rtabmap" paired with "Asus Xtion Pro Live" on an Intel NUC
  board is a better option and is more accurate. The proposed
  system is real-time, efficient and low cost at the same
  time as compared to other existing low-cost
  counter-parts.},
  address  = {Grantham-Allee 20, 53757 St. Augustin, Germany},
  annote  = {WS12/13 FARO Europe - Master Thesis Pl{\"o}ger,
  Kraetzschmar, Zweigle},
  author  = {Muhammad Umair Tahir},
  month = {August},
  school  = {Bonn-Rhein-Sieg University of Applied Sciences},
  title = {Analysis of live 3D mapping approaches with different
  sensors on various hardware platforms},
  year = {2016}
  }

• I. Siddique, “Comparative Analysis on Sensor Configurations of Autonomous Vehicles for Robust Obstacle Avoidance,” Master Thesis, Grantham-Allee 20, 53757 St. Augustin, Germany, 2016.
  [BibTeX] [Abstract]
  

  Autonomous vehicle working in close proximity to human raises a safety concern. Such vehicles have to be equipped with flawless obstacle avoidance techniques. Robust obstacle avoidance can only be achieved if the environment could be perceived flawlessly. Unfortunately the sensors to perceive the environment are not flawless. This forced modern day autonomous vehicles to use different combination of sensors to compensate each others limitations. This comes with additional agitation of fusing different sensors data. Besides in spite of combining different sensors there might still be situations which hamper the nominal performance of the sensor configuration. This makes successful and safe deployment of autonomous vehicle in public a very challenging task. Hence a comprehensive survey on sensor configurations is of immense importance. Here in this thesis work different sensors and sensor configurations of recent wheeled autonomous vehicles in indoor and outdoor scenarios have been investigated. The investigated sensor conflagrations were analyzed and compared to find a general trend in these different scenarios. Performance of the individual sensors used in those vehicles were also analyzed and compared.

  @MastersThesis{ 2016siddiqueisnain,
  abstract  = {Autonomous vehicle working in close proximity to human
  raises a safety concern. Such vehicles have to be equipped
  with flawless obstacle avoidance techniques. Robust
  obstacle avoidance can only be achieved if the environment
  could be perceived flawlessly. Unfortunately the sensors to
  perceive the environment are not flawless. This forced
  modern day autonomous vehicles to use different combination
  of sensors to compensate each others limitations. This
  comes with additional agitation of fusing different sensors
  data. Besides in spite of combining different sensors there
  might still be situations which hamper the nominal
  performance of the sensor configuration. This makes
  successful and safe deployment of autonomous vehicle in
  public a very challenging task. Hence a comprehensive
  survey on sensor configurations is of immense importance.
  Here in this thesis work different sensors and sensor
  configurations of recent wheeled autonomous vehicles in
  indoor and outdoor scenarios have been investigated. The
  investigated sensor conflagrations were analyzed and
  compared to find a general trend in these different
  scenarios. Performance of the individual sensors used in
  those vehicles were also analyzed and compared. },
  address  = {Grantham-Allee 20, 53757 St. Augustin, Germany},
  annote  = {WS13/14 FH-HBRS - Comparative Analysis on Sensor
  Configurations of Autonomous Vehicles for Robust Obstacle
  Avoidance, Prassler, Pl{\"o}ger},
  author  = {Isnain Siddique},
  month = {October},
  school  = {Bonn-Rhein-Sieg University of Applied Sciences},
  title = {Comparative Analysis on Sensor Configurations of
  Autonomous Vehicles for Robust Obstacle Avoidance},
  year = {2016}
  }

• A. Mitrevski, “Improving the Manipulation Skills of Service Robots by Refining Action Representations,” Master Thesis, Grantham-Allee 20, 53757 Sankt Augustin, Germany, 2016.
  [BibTeX] [Abstract]
  

  Releasing objects is an error-prone robot manipulation activity often due to insufficient knowledge about the preconditions of releasing actions. As the problem has a semantic nature that stems from the variations in physical behaviour between different object categories, neither symbolic nor geometric models alone are enough for specifying how and where a releasing action should be executed. In principle, those two have to be combined into a more general representation that maximises the execution success and minimises the probability of failures; however, symbolic and geometric action representations are frequently studied in isolation and one of them is taken for granted. This thesis investigates the exact nature of releasing actions, the problem of learning reusable models of such actions, and the manner in which those models can be utilised for execution. Starting with an examination of multiple planning domain learning algorithms and geometric reasoning procedures, we develop a template-based representation of actions that simplifies the acquisition and application of symbolic and geometric models; in particular, we show that template models, which we call $\delta$ and $\delta^\varnothing$ problems, can be learned in special configurations and that pairwise object relations should be an inherent part of the models’ nature. The models are organised in an object-centred memory model, called $\Delta$ memory, in which storage and retrieval depend on semantic object information. With the help of a simulated environment that abstracts away various manipulation activities, we show that properly constructed $\delta$ and $\delta^\varnothing$ problems can be a reliable representation of releasing actions, thereby justifying the use of template-based action representations, but also demonstrate that their utility depends on a number of conventions and can benefit from an additional ontology.

  @MastersThesis{ 2016mitrevski,
  title = {Improving the Manipulation Skills of Service Robots by
  Refining Action Representations},
  author  = {Aleksandar Mitrevski},
  abstract  = {Releasing objects is an error-prone robot manipulation
  activity often due to insufficient knowledge about the
  preconditions of releasing actions. As the problem has a
  semantic nature that stems from the variations in physical
  behaviour between different object categories, neither
  symbolic nor geometric models alone are enough for
  specifying how and where a releasing action should be
  executed. In principle, those two have to be combined into
  a more general representation that maximises the execution
  success and minimises the probability of failures; however,
  symbolic and geometric action representations are
  frequently studied in isolation and one of them is taken
  for granted.
  This thesis investigates the exact nature of releasing
  actions, the problem of learning reusable models of such
  actions, and the manner in which those models can be
  utilised for execution. Starting with an examination of
  multiple planning domain learning algorithms and geometric
  reasoning procedures, we develop a template-based
  representation of actions that simplifies the acquisition
  and application of symbolic and geometric models; in
  particular, we show that template models, which we call
  $\delta$ and $\delta^\varnothing$ problems, can be learned
  in special configurations and that pairwise object
  relations should be an inherent part of the models' nature.
  The models are organised in an object-centred memory model,
  called $\Delta$ memory, in which storage and retrieval
  depend on semantic object information.
  With the help of a simulated environment that abstracts
  away various manipulation activities, we show that properly
  constructed $\delta$ and $\delta^\varnothing$ problems can
  be a reliable representation of releasing actions, thereby
  justifying the use of template-based action
  representations, but also demonstrate that their utility
  depends on a number of conventions and can benefit from an
  additional ontology.},
  school  = {Bonn-Rhein-Sieg University of Applied Sciences},
  address  = {Grantham-Allee 20, 53757 Sankt Augustin, Germany},
  month = {January},
  year = {2016},
  annote  = {WS13/14 HBRS - Pl{\"o}ger, Witt, Kuestenmacher
  supervising}
  }

• A. M. Sundaram, “Planning Realistic Force Interactions for Bimanual Grasping and Manipulation,” Master Thesis, Grantham-Allee 20, 53757 St. Augustin, Germany, 2016.
  [BibTeX] [Abstract]
  

  Dexterous robot hands offer a wide variety of grasping and interaction possibilities with objects. In order to select the best grasp, it is critical to count with a reliable grasp quality measure. Traditional grasp analysis methods use quality measures that allow a relative comparison of grasps for the same object, without an associated physical meaning for the resulting quality. The focus of this thesis is to establish an improved grasp analysis method that will result in a quality measure that can be directly interpreted in the force domain. One of the most commonly used grasp qualities is the largest minimum resisted wrench, which indicates the maximum perturbation wrench that a grasp can resist in any direction. Two efficient ways to calculate this quality are identified: (i) incremental grasp wrench space algorithm, and (ii) ray shooting algorithm. However, existing algorithms for such methods make several assumptions to avoid computational complexities in analyzing the 6D wrench space of a grasp. Important properties like hand actuation, realizable contact forces, friction at the contacts, and geometry of the object to be grasped are either neglected or greatly simplified. In this thesis, these assumptions are improved to bring those algorithms closer to reality. In the case of bimanual grasps, the number of contacts significantly increases, which in turn increases the computational complexity of the process. Suitable algorithms to handle a higher number of contacts are also proposed in this thesis. For grasping an object successfully, considering the hand and the object for analysis are necessary but not sufficient requirements. The capabilities of the robotic arm to which the hand is attached are equally important. Different manipulability measures are considered for the arm, corresponding to single and dual hand grasps, and they are later unified with the physically relevant grasp quality to obtain an overall measure of the goodness of a particular grasp. Based on the updated grasp quality, a complete grasp planning architecture is established. It also includes methods for bimanual grasp synthesis and grasp filtering based on properties like collision with the environment and arm reachability. The thesis includes application examples that illustrate the applicability of the approach. Finally, the developed algorithms can be generalized to a different type of force interaction task, namely a humanoid robot balancing with multiple contacts with the environment. A customized ray shooting algorithm is used to find the stability region of a humanoid legged robot standing on an uneven terrain or making multiple contacts with its hands and legs. In contrast to the regular zero-moment point based method, the stability region is found by analyzing the wrench space of the robot, which makes the method independent of the number of contacts or the contact configuration. Different examples show the direct and intuitive interpretation of the results obtained with the proposed method.

  @MastersThesis{ 2016meenakshisundaram,
  abstract  = {Dexterous robot hands offer a wide variety of grasping and
  interaction possibilities with objects. In order to select
  the best grasp, it is critical to count with a reliable
  grasp quality measure. Traditional grasp analysis methods
  use quality measures that allow a relative comparison of
  grasps for the same object, without an associated physical
  meaning for the resulting quality. The focus of this thesis
  is to establish an improved grasp analysis method that will
  result in a quality measure that can be directly
  interpreted in the force domain.
  One of the most commonly used grasp qualities is the
  largest minimum resisted wrench, which indicates the
  maximum perturbation wrench that a grasp can resist in any
  direction. Two efficient ways to calculate this quality are
  identified: (i) incremental grasp wrench space algorithm,
  and (ii) ray shooting algorithm. However, existing
  algorithms for such methods make several assumptions to
  avoid computational complexities in analyzing the 6D wrench
  space of a grasp. Important properties like hand actuation,
  realizable contact forces, friction at the contacts, and
  geometry of the object to be grasped are either neglected
  or greatly simplified. In this thesis, these assumptions
  are improved to bring those algorithms closer to reality.
  In the case of bimanual grasps, the number of contacts
  significantly increases, which in turn increases the
  computational complexity of the process. Suitable
  algorithms to handle a higher number of contacts are also
  proposed in this thesis.
  For grasping an object successfully, considering the hand
  and the object for analysis are necessary but not
  sufficient requirements. The capabilities of the robotic
  arm to which the hand is attached are equally important.
  Different manipulability measures are considered for the
  arm, corresponding to single and dual hand grasps, and they
  are later unified with the physically relevant grasp
  quality to obtain an overall measure of the goodness of a
  particular grasp. Based on the updated grasp quality, a
  complete grasp planning architecture is established. It
  also includes methods for bimanual grasp synthesis and
  grasp filtering based on properties like collision with the
  environment and arm reachability. The thesis includes
  application examples that illustrate the applicability of
  the approach.
  Finally, the developed algorithms can be generalized to a
  different type of force interaction task, namely a humanoid
  robot balancing with multiple contacts with the
  environment. A customized ray shooting algorithm is used to
  find the stability region of a humanoid legged robot
  standing on an uneven terrain or making multiple contacts
  with its hands and legs. In contrast to the regular
  zero-moment point based method, the stability region is
  found by analyzing the wrench space of the robot, which
  makes the method independent of the number of contacts or
  the contact configuration. Different examples show the
  direct and intuitive interpretation of the results obtained
  with the proposed method.},
  address  = {Grantham-Allee 20, 53757 St. Augustin, Germany},
  annote  = {SS13 DLR - Planning Realistic Force Interactions for
  Bimanual Grasping and Manipulation Kraetzschmar,
  Albu-Schaeffer, Roa, Schneider supervising},
  author  = {Ashok Meenakshi Sundaram},
  month = {June},
  school  = {Bonn-Rhein-Sieg University of Applied Sciences},
  title = {Planning Realistic Force Interactions for Bimanual
  Grasping and Manipulation},
  year = {2016}
  }

• O. L. Carrion, “Task Planning, Execution and Monitoring for Mobile Manipulators in Industrial Domains,” Master Thesis, Grantham-Allee 20, 53757 St. Augustin, Germany, 2016.
  [BibTeX] [Abstract]
  

  Decision making is a necessary skill for autonomous mobile manipulators working in industrial environments. Task planning is a well-established field with a large research community that continues to produce new heuristics, tools and algorithms that enable agents to produce plans that achieve their goals. In this work we demonstrate the applicability of satisficing task planning with action costs by integrating a state-of-the-art planner, the Mercury planner, into the KUKA youBot mobile manipulator and using it to solve basic transportation and insertion tasks. In contrast to optimal planners which minimize total action costs in a plan, satisficing planners minimize plan generation time, yielding sub-optimal solutions but in less time compared to optimal planners. The planner uses a delete-list relaxation heuristic to quickly prune the search space and generate satisfactory solutions. The developed planning framework is modular, re-usable and well documented. It brings together two major standards in robotics and task planning: ROS and PDDL, similar to ROSPlan. Unlike ROSPlan, this work is able to plan with cost information. Moreover, it is more modular, enabling the community to use the various components at their own discretion. Finally, while ROSPlan allows only one re-planning strategy, our framework enables users to quickly implement their own strategies. The resulting system demonstrates that the agent’s behavior is optimized, it is able to flexibly handle unexpected situations, and it can robustly handle failures by re-planning when needed. It is also easier to maintain and extend. The work presented here also highlights the benefits of conducting a domain analysis to gain the maximum benefit from the use of a given planner and domain.

  @MastersThesis{ 2016limaoscar,
  abstract  = {Decision making is a necessary skill for autonomous mobile
  manipulators working in industrial environments. Task
  planning is a well-established field with a large research
  community that continues to produce new heuristics, tools
  and algorithms that enable agents to produce plans that
  achieve their goals. In this work we demonstrate the
  applicability of satisficing task planning with action
  costs by integrating a state-of-the-art planner, the
  Mercury planner, into the KUKA youBot mobile manipulator
  and using it to solve basic transportation and insertion
  tasks. In contrast to optimal planners which minimize total
  action costs in a plan, satisficing planners minimize plan
  generation time, yielding sub-optimal solutions but in less
  time compared to optimal planners. The planner uses a
  delete-list relaxation heuristic to quickly prune the
  search space and generate satisfactory solutions. The
  developed planning framework is modular, re-usable and well
  documented. It brings together two major standards in
  robotics and task planning: ROS and PDDL, similar to
  ROSPlan. Unlike ROSPlan, this work is able to plan with
  cost information. Moreover, it is more modular, enabling
  the community to use the various components at their own
  discretion. Finally, while ROSPlan allows only one
  re-planning strategy, our framework enables users to
  quickly implement their own strategies. The resulting
  system demonstrates that the agent's behavior is optimized,
  it is able to flexibly handle unexpected situations, and it
  can robustly handle failures by re-planning when needed. It
  is also easier to maintain and extend. The work presented
  here also highlights the benefits of conducting a domain
  analysis to gain the maximum benefit from the use of a
  given planner and domain.},
  address  = {Grantham-Allee 20, 53757 St. Augustin, Germany},
  annote  = {SS04/13 FH-BRS - RoboCup@Work Pl{\"o}ger, Kraetzschmar,
  Awaad supervising},
  author  = {Oscar Lima Carrion},
  month = {April},
  school  = {Bonn-Rhein-Sieg University of Applied Sciences},
  title = {Task Planning, Execution and Monitoring for Mobile
  Manipulators in Industrial Domains},
  year = {2016}
  }

• D. Nair, “Predicting Object Locations using Spatio-Temporal Information by a Domestic Service Robot: A Bayesian Learning Approach,” Master Thesis, Grantham-Allee 20, 53757 St. Augustin, Germany, 2016.
  [BibTeX] [Abstract]
  

  One of the ways domestic service robots can better assist humans is by providing personalized, predictive and context-aware services. Robots can observe human activities and work patterns and provide time-based contextual assistance. This thesis aims to enable domestic robots to empirically learn about human behaviour and preferences. In the current literature, user preferences are learned for a generic home environment, on the contrary we learn preferences over a specific home. Robots generate a lot of information using the raw data from their sensors, which is often discarded after use. If this information is recorded, it can be used to generate new knowledge. The thesis proposes models to generate knowledge about user preferences using stored information. The developed approaches in this thesis cover the following two knowledge generation topics: (1) learning user location preferences (2) learning user preferences in object placement. All knowledge generation techniques developed in this thesis are based on Bayesian modelling and have been implemented using probabilistic programming languages. The learned user preferences were used by the robot for predicting: (a) location of non-stationary objects (b) location of users in home and (c) room occupancy. The models were evaluated on three datasets collected over several months containing person and object occurrences in home and office environments. The efficiency of the models was accessed by measuring the accuracy score of each models. Our model for predicting location of non-stationary objects (a) was able to predict with 70% accuracy for 26 objects, while The model for user location preference (b) was able to predict with 63% accuracy and model for room occupancy (c) could predict for 3 rooms with more than 80% accuracy.

  @MastersThesis{ 2016nairdeebul,
  abstract  = {One of the ways domestic service robots can better assist
  humans is by providing personalized, predictive and
  context-aware services. Robots can observe human activities
  and work patterns and provide time-based contextual
  assistance. This thesis aims to enable domestic robots to
  empirically learn about human behaviour and preferences. In
  the current literature, user preferences are learned for a
  generic home environment, on the contrary we learn
  preferences over a specific home. Robots generate a lot of
  information using the raw data from their sensors, which is
  often discarded after use. If this information is recorded,
  it can be used to generate new knowledge. The thesis
  proposes models to generate knowledge about user
  preferences using stored information. The developed
  approaches in this thesis cover the following two knowledge
  generation topics: (1) learning user location preferences
  (2) learning user preferences in object placement. All
  knowledge generation techniques developed in this thesis
  are based on Bayesian modelling and have been implemented
  using probabilistic programming languages. The learned user
  preferences were used by the robot for predicting: (a)
  location of non-stationary objects (b) location of users in
  home and (c) room occupancy. The models were evaluated on
  three datasets collected over several months containing
  person and object occurrences in home and office
  environments. The efficiency of the models was accessed by
  measuring the accuracy score of each models. Our model for
  predicting location of non-stationary objects (a) was able
  to predict with 70% accuracy for 26 objects, while The
  model for user location preference (b) was able to predict
  with 63% accuracy and model for room occupancy (c) could
  predict for 3 rooms with more than 80% accuracy.},
  address  = {Grantham-Allee 20, 53757 St. Augustin, Germany},
  annote  = {SS14 Pl{\"o}ger, Lakemeyer, Niemueller supervising},
  author  = {Deebul Nair},
  month = {September},
  school  = {Bonn-Rhein-Sieg University of Applied Sciences},
  title = {Predicting Object Locations using Spatio-Temporal
  Information by a Domestic Service Robot: A Bayesian
  Learning Approach},
  year = {2016}
  }

• A. Ortega Sáinz, “Multi-Robot Path Planning in Confined Dynamic Environments,” Master Thesis, Grantham-Allee 20, 53757 St. Augustin, Germany, 2016.
  [BibTeX] [Abstract]
  

  The automation of transportation tasks in logistics using autonomous robots has been gaining popularity in the last years, due in most part to the advantages they present against other automation technologies or manually operated solutions. The productivity and efficiency of an application employing a Multi-Robot System (MRS) is a direct result of the path planning and coordination strategy used. The Multi-Robot Path Planning (MRPP) problem has been widely researched throughout the years, and given the popularity of the topic and the amount of algorithms available, selecting the best approach for a given application is not as straightforward as it may seem. First of all, there is still a gap between the problems solved by the academic community in the state of the art and the capabilities of modern technologies used in real applications. Numerous efforts are being made to close this gap, but given the complexities of evaluating multi-robot systems with real robots, there is still a very long way to go. Furthermore, the lack of a standard methodology to evaluate MRPP problems makes the selection of an approach for a given application difficult, particularly since the reported results are not directly comparable. Using a hospital transportation task as an example use case, this thesis will benchmark decentralized path planning approaches, keeping in mind real-life conditions, oriented towards implementation rather than theory. A comparative analysis will asses the selected approaches in regards to their scalability in the number of robots and their robustness to dynamic environments populated by moving obstacles. An analysis of the evaluation methodology and performance metrics reported in the state of the art will be use as a basis for the proposed set of guidelines and performance metrics to benchmark MRPP approaches hereafter. Finally, the groundwork for a benchmarking framework will be presented.

  @MastersThesis{ 2016ortega,
  abstract  = {The automation of transportation tasks in logistics using
  autonomous robots has been gaining popularity in the last
  years, due in most part to the advantages they present
  against other automation technologies or manually operated
  solutions. The productivity and efficiency of an
  application employing a Multi-Robot System (MRS) is a
  direct result of the path planning and coordination
  strategy used. The Multi-Robot Path Planning (MRPP) problem
  has been widely researched throughout the years, and given
  the popularity of the topic and the amount of algorithms
  available, selecting the best approach for a given
  application is not as straightforward as it may seem. First
  of all, there is still a gap between the problems solved by
  the academic community in the state of the art and the
  capabilities of modern technologies used in real
  applications. Numerous efforts are being made to close this
  gap, but given the complexities of evaluating multi-robot
  systems with real robots, there is still a very long way to
  go. Furthermore, the lack of a standard methodology to
  evaluate MRPP problems makes the selection of an approach
  for a given application difficult, particularly since the
  reported results are not directly comparable. Using a
  hospital transportation task as an example use case, this
  thesis will benchmark decentralized path planning
  approaches, keeping in mind real-life conditions, oriented
  towards implementation rather than theory. A comparative
  analysis will asses the selected approaches in regards to
  their scalability in the number of robots and their
  robustness to dynamic environments populated by moving
  obstacles. An analysis of the evaluation methodology and
  performance metrics reported in the state of the art will
  be use as a basis for the proposed set of guidelines and
  performance metrics to benchmark MRPP approaches hereafter.
  Finally, the groundwork for a benchmarking framework will
  be presented.},
  address  = {Grantham-Allee 20, 53757 St. Augustin, Germany},
  annote  = {WS13/14 HBRS Prassler, Pl{\"o}ger, F{\"u}ller supervising},
  author  = {Argentina {Ortega S{\'a}inz}},
  keywords  = {multi-robot systems; multi-robot path planning},
  month = {July},
  school  = {Bonn-Rhein-Sieg University of Applied Sciences},
  title = {Multi-Robot Path Planning in Confined Dynamic
  Environments},
  year = {2016}
  }

• S. Sathanam, “Emotion detection from German text by sentiment analysis,” Master Thesis, Grantham-Allee 20, 53757 St. Augustin, Germany, 2016.
  [BibTeX] [Abstract]
  

  Social assistive robots(SAR) providing emotional care and support is investigated in Emorobot project to help the patients affected by dementia. One of the functions of the SAR is to detect emotion from the patients to communicate effectively. Detecting emotions directly via facial expressions or vocal features have certain demerits with the elderly people. Detection of emotions from the spoken text is considered in this work. The six basic emotions Happy, Sad, Anger, Fear, Surprise and disgust are considered in addition to emotionless(none) condition. Emotion detection which comes under sentiment analysis has major issues with the complexity of feature vectors representation and preserving semantic meanings. The emergence of word embeddings is a recent breakthrough in natural language processing, where huge amount of data are used to learn semantically preserved word and sentence vectors in an unsupervised manner. Application of word vectors and document vectors in emotion detection is studied in this work. Detecting emotions from text is handled in two phases. The first phase involves generation of feature vectors from vector models. Vector models require huge amount of dataset to generate high quality feature vectors. Dataset containing nearly 10 million sentences are created by streaming online tweets. Another dataset around 18 million wikipedia sentences are also used for comparison of the nature and quality of the vectors generated. Second phase is estimating emotions from the sentences with the use of feature vectors. A sentence could be represented both by document vectors as well as average of word vectors in the sentence. Training dataset is manually generated by applying automatic labeling on twitter dataset to label the tweets for emotions based on the hashtags attached. Evaluation of the model is performed on the annotated sentences from Emorobot project as ground truth. Due to lesser data, a test dataset from twitter is also evaluated. A detailed evaluation of the vector models is provided for the use case considered. In total, two datasets Twitter and Wiki are applied on two word vector models CBOW, SG and two document vector models DBOW and DM to generate feature vectors which are evaluated against two test datasets Twitter and annotated dataset. A result of 16 cases shows that generally averaged word vectors perform better than document vectors. In word vector models, CBOW model outperforms SG model but by only a small value. Document vectors are unpredictable in nature as the results vary drastically for each runs. Eventhough, the vectors generated from wikipedia datasets contains better features as evaluated by analogy test sentences. The vectors generated from twitter datasets achieves better results than the feature vectors learnt from wikipedia datasets.

  @MastersThesis{ 2016sathanam,
  abstract  = {Social assistive robots(SAR) providing emotional care and
  support is investigated in Emorobot project to help the
  patients affected by dementia. One of the functions of the
  SAR is to detect emotion from the patients to communicate
  effectively. Detecting emotions directly via facial
  expressions or vocal features have certain demerits with
  the elderly people. Detection of emotions from the spoken
  text is considered in this work. The six basic emotions
  Happy, Sad, Anger, Fear, Surprise and disgust are
  considered in addition to emotionless(none) condition.
  Emotion detection which comes under sentiment analysis has
  major issues with the complexity of feature vectors
  representation and preserving semantic meanings. The
  emergence of word embeddings is a recent breakthrough in
  natural language processing, where huge amount of data are
  used to learn semantically preserved word and sentence
  vectors in an unsupervised manner. Application of word
  vectors and document vectors in emotion detection is
  studied in this work. Detecting emotions from text is
  handled in two phases. The first phase involves generation
  of feature vectors from vector models. Vector models
  require huge amount of dataset to generate high quality
  feature vectors. Dataset containing nearly 10 million
  sentences are created by streaming online tweets. Another
  dataset around 18 million wikipedia sentences are also used
  for comparison of the nature and quality of the vectors
  generated. Second phase is estimating emotions from the
  sentences with the use of feature vectors. A sentence could
  be represented both by document vectors as well as average
  of word vectors in the sentence. Training dataset is
  manually generated by applying automatic labeling on
  twitter dataset to label the tweets for emotions based on
  the hashtags attached. Evaluation of the model is performed
  on the annotated sentences from Emorobot project as ground
  truth. Due to lesser data, a test dataset from twitter is
  also evaluated. A detailed evaluation of the vector models
  is provided for the use case considered. In total, two
  datasets Twitter and Wiki are applied on two word vector
  models CBOW, SG and two document vector models DBOW and DM
  to generate feature vectors which are evaluated against two
  test datasets Twitter and annotated dataset. A result of 16
  cases shows that generally averaged word vectors perform
  better than document vectors. In word vector models, CBOW
  model outperforms SG model but by only a small value.
  Document vectors are unpredictable in nature as the results
  vary drastically for each runs. Eventhough, the vectors
  generated from wikipedia datasets contains better features
  as evaluated by analogy test sentences. The vectors
  generated from twitter datasets achieves better results
  than the feature vectors learnt from wikipedia datasets.},
  address  = {Grantham-Allee 20, 53757 St. Augustin, Germany},
  annote  = {WS2013 - Emotion detection from German text by sentiment
  analysis, Pl{\"o}ger, Kraetzschmar, F{\"u}ller supervising},
  author  = {Sivasurya Sathanam},
  month = {April},
  school  = {Bonn-Rhein-Sieg University of Applied Sciences},
  title = {Emotion detection from German text by sentiment analysis},
  year = {2016}
  }

• D. E. Ramos Avila, “A Study on Swarm Intelligence: Towards Nature-Inspired Robot Navigation,” Master Thesis, Grantham-Allee 20, 53757 St. Augustin, Germany, 2016.
  [BibTeX] [Abstract]
  

  Swarm Intelligence (SI) is a problem-solving approach whose inspiration comes from cooperative behaviours and well-defined social structures observed in different animal species. Within those structures, simple interactions among “unsophisticated” individuals produce complex and interesting properties which aid flocks of birds, schools of fish, packs of predators and colonies of insects in solving daily tasks in a very efficient way. Several meta-heuristics such as the particle swarm optimization (PSO) and the artificial bee colony algorithm (ABC) are based on this metaphor and entail important benefits. One of their most remarkable advantages is their ability to find solutions to NP-hard problems in a very short time. In the field of robotics, motion planning is a well-known problem considered to be NP-hard. This work studies SI as a possible alternative to state-of-the-art global path planners, with the main objective of producing robot-traversable paths in a competitive amount of time without further optimization. In particular, the usage of the novel grey wolf optimizer (GWO) is proposed and tested against other SI-based meta-heuristics, and then again against popular sampling-based planners. The experimental evaluation indicates that GWO usually has a higher success rate and faster convergence than other SI-based algorithms like ABC and the firefly algorithm (FFA). Moreover, the results show that even though rapidly-exploring random trees (RRT) are much faster, the proposed strategy produces shorter and smoother paths and even often outperforms probabilistic roadmaps (PRM) in terms of computational time. In spite of the limitations that the greedy nature of the planner still poses when dealing with complex environments, the results highlight a promising line of research that might alleviate current issues in motion planning. The approach is further extended to address multiple objective functions using the multi-objective grey wolf optimizer (MOGWO) based on Pareto dominance, as a typical requirement of robot motion planning is not to produce the shortest possible path, but rather a short-enough, smooth-enough and safe-enough path.

  @MastersThesis{ 2016ramosavila,
  title = {A Study on Swarm Intelligence: Towards Nature-Inspired
  Robot Navigation},
  author  = {Ramos Avila, Diego Enrique},
  school  = {Bonn-Rhein-Sieg University of Applied Sciences},
  year = {2016},
  address  = {Grantham-Allee 20, 53757 St. Augustin, Germany},
  month = {April},
  abstract  = {Swarm Intelligence (SI) is a problem-solving approach
  whose inspiration comes from cooperative behaviours and
  well-defined social structures observed in different animal
  species. Within those structures, simple interactions among
  “unsophisticated” individuals produce complex and
  interesting properties which aid flocks of birds, schools
  of fish, packs of predators and colonies of insects in
  solving daily tasks in a very efficient way. Several
  meta-heuristics such as the particle swarm optimization
  (PSO) and the artificial bee colony algorithm (ABC) are
  based on this metaphor and entail important benefits. One
  of their most remarkable advantages is their ability to
  find solutions to NP-hard problems in a very short time. In
  the field of robotics, motion planning is a well-known
  problem considered to be NP-hard. This work studies SI as a
  possible alternative to state-of-the-art global path
  planners, with the main objective of producing
  robot-traversable paths in a competitive amount of time
  without further optimization. In particular, the usage of
  the novel grey wolf optimizer (GWO) is proposed and tested
  against other SI-based meta-heuristics, and then again
  against popular sampling-based planners. The experimental
  evaluation indicates that GWO usually has a higher success
  rate and faster convergence than other SI-based algorithms
  like ABC and the firefly algorithm (FFA). Moreover, the
  results show that even though rapidly-exploring random
  trees (RRT) are much faster, the proposed strategy produces
  shorter and smoother paths and even often outperforms
  probabilistic roadmaps (PRM) in terms of computational
  time. In spite of the limitations that the greedy nature of
  the planner still poses when dealing with complex
  environments, the results highlight a promising line of
  research that might alleviate current issues in motion
  planning. The approach is further extended to address
  multiple objective functions using the multi-objective grey
  wolf optimizer (MOGWO) based on Pareto dominance, as a
  typical requirement of robot motion planning is not to
  produce the shortest possible path, but rather a
  short-enough, smooth-enough and safe-enough path.},
  annote  = {WS13/14 H-BRS - A Study on Swarm Intelligence: Towards
  Nature-Inspired Robot Navigation Pl{\"o}ger, Asteroth
  supervising}
  }

• N. Deshpande, “Using Semantic Information for Robot Navigation,” Master Thesis, Grantham-Allee 20, 53757 St. Augustin, Germany, 2016.
  [BibTeX] [Abstract]
  

  Current research in robotics focuses on making autonomous robots for domestic purposes which have to work along with humans in real environments. One of the prerequisites for such robots is the ability to navigate intelligently in their environment. Traditional navigation systems focus on finding a shortest collision free path by avoiding all obstacles. This approach has worked well in laboratories and structured environments. However, for robots navigating in real environments that are filled with humans, often, finding a shortest collision free path is not the desired behaviour. The robot needs to navigate in a more intelligent manner by adapting its behaviour based on the context at hand. Moreover, real environments are unstructured and there are often situations where a robot needs to reposition objects to navigate and reach its goal. This work addresses the problem of using semantic information for making navigation adaptive to different situations. Towards this goal, different forms of semantic information useful for navigation have been identified. An architecture is proposed to represent this information and use it for different aspects of navigation. The proposed architecture also uses contextual information about the tasks at hand for navigation. The architecture has been developed and implemented using the Robot Operating System(ROS) framework. Tests have been performed in simulation using the Pioneer 3DX robot platform. Preliminary results prove the validity of the proposed architecture. The robot was able to navigate in a more desired and intelligent manner by using semantic and contextual information.

  @MastersThesis{ 2016deshpande,
  title = {Using Semantic Information for Robot Navigation},
  author  = {Niranjan Deshpande},
  school  = {Bonn-Rhein-Sieg University of Applied Sciences},
  year = {2016},
  address  = {Grantham-Allee 20, 53757 St. Augustin, Germany},
  month = {March},
  note = {Advisors: Prof. Dr. Paul G. Pl{\"o}ger, Prof. Dr. Anne
  Spalanzani, Mr. Sven Schneider},
  abstract  = {Current research in robotics focuses on making autonomous
  robots for domestic purposes which have to work along with
  humans in real environments. One of the prerequisites for
  such robots is the ability to navigate intelligently in
  their environment. Traditional navigation systems focus on
  finding a shortest collision free path by avoiding all
  obstacles. This approach has worked well in laboratories
  and structured environments. However, for robots navigating
  in real environments that are filled with humans, often,
  finding a shortest collision free path is not the desired
  behaviour. The robot needs to navigate in a more
  intelligent manner by adapting its behaviour based on the
  context at hand. Moreover, real environments are
  unstructured and there are often situations where a robot
  needs to reposition objects to navigate and reach its goal.
  This work addresses the problem of using semantic
  information for making navigation adaptive to different
  situations. Towards this goal, different forms of semantic
  information useful for navigation have been identified. An
  architecture is proposed to represent this information and
  use it for different aspects of navigation. The proposed
  architecture also uses contextual information about the
  tasks at hand for navigation. The architecture has been
  developed and implemented using the Robot Operating
  System(ROS) framework. Tests have been performed in
  simulation using the Pioneer 3DX robot platform.
  Preliminary results prove the validity of the proposed
  architecture. The robot was able to navigate in a more
  desired and intelligent manner by using semantic and
  contextual information.}
  }

2015

• G. Sannino, I. {De Falco}, and G. {De Pietro}, “A Supervised Approach to Automatically Extract a Set of Rules to Support Fall Detection in an mHealth System,” Applied Soft Computing Journal, vol. 34, p. 205 – 216, 2015.
  [BibTeX]

  @Article{ sannino2015,
  author  = {Sannino, Giovanna and {De Falco}, Ivanoe and {De Pietro},
  Giuseppe},
  journal  = {Applied Soft Computing Journal},
  title = {{A Supervised Approach to Automatically Extract a Set of
  Rules to Support Fall Detection in an mHealth System}},
  year = {2015},
  volume  = {34},
  pages = {205 -- 216},
  issn = {1568-4946}
  }

• S. Ioffe and C. Szegedy, “Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift,” CoRR, 2015.
  [BibTeX]

  @Article{ sergey2015,
  author  = {Ioffe, Sergey and Szegedy, Christian},
  title = {{Batch Normalization: Accelerating Deep Network Training
  by Reducing Internal Covariate Shift}},
  journal  = {CoRR},
  year = {2015}
  }

• Y. Yun, C. Innocenti, G. Nero, H. Linden, and I. Gu, “Fall Detection in RGB-D Videos for Elderly Care,” in 2015 17th International Conference on E-health Networking, Application Services (HealthCom), 2015, p. 422 – 427.
  [BibTeX]

  @InProceedings{ yun2015,
  author  = {Yun, Yixiao and Innocenti, Christopher and Nero, Gustav
  and Linden, Henrik and Gu, Irene},
  booktitle  = {2015 17th International Conference on E-health Networking,
  Application Services (HealthCom)},
  title = {{Fall Detection in RGB-D Videos for Elderly Care}},
  year = {2015},
  pages = {422 -- 427}
  }

• Z. Zhang, C. Conly, and V. Athitsos, “A Survey on Vision-based Fall Detection,” in Proceedings of the 8th ACM International Conference on PErvasive Technologies Related to Assistive Environments – PETRA ’15, 2015, p. 46:1–46:7.
  [BibTeX]

  @InProceedings{ zhang2015,
  author  = {Zhang, Zhong and Conly, Christopher and Athitsos,
  Vassilis},
  title = {{A Survey on Vision-based Fall Detection}},
  booktitle  = {Proceedings of the 8th ACM International Conference on
  PErvasive Technologies Related to Assistive Environments -
  PETRA '15},
  publisher  = {ACM},
  pages = {46:1--46:7},
  year = {2015}
  }

• F. Caba Heilbron, V. Escorcia, B. Ghanem, and J. Carlos Niebles, “ActivityNet: A Large-Scale Video Benchmark for Human Activity Understanding,” in The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015.
  [BibTeX]

  @InProceedings{ heilbron2015,
  author  = {Caba Heilbron, Fabian and Escorcia, Victor and Ghanem,
  Bernard and Carlos Niebles, Juan},
  booktitle  = {The IEEE Conference on Computer Vision and Pattern
  Recognition (CVPR)},
  title = {{ActivityNet: A Large-Scale Video Benchmark for Human
  Activity Understanding}},
  year = {2015}
  }

• M. Zolghadr, “Semantic Similarity Between Objects in Home Environments,” Master Thesis, Grantham-Allee 20, 53757 St. Augustin, Germany, 2015.
  [BibTeX] [Abstract]
  

  When a service robot is acting in a home environment, there are situations where some of the objects necessary for completing the task may not be available, because either they are missing or the preconditions for using them are not met. In such situations, humans show incredible flexibility by making substitutions. However, for a robot identifying the similarity between objects and finding an appropriate substitute for an unavailable object is not a straightforward task. Most of the approaches which tried to solve the problem have focused on lifting, where another instance of the same object class is used as a substitute. These approaches apply a limited range of semantics. Most approaches mea- sure similarity between objects using perception. Perception-based approaches tend to be time-consuming. However, the application of semantic-based similarity measures has been successfully and efficiently used in the genetics domain. The approach presented here combines similarity measurement with knowledge base queries in order to find appropriate substitutes. In our ontological approach, similarity measurement is carried out based on a broad range of objects semantics, including the functional affordances, object features, part-whole relations, and spatial proximity of the objects in the home environment. In this research, existing tools, previously used in the biological domains, for measuring the similarity of individuals within an ontology are evaluated and an analysis of the most useful measures is conducted. The experiments which were conducted were used to guide the creation of a methodology for the modeling of the objects within the knowledge base. The results show that the Jaccard Similarity Coefficient is particularly well-suited to our goal. The results also highlight the limitations of this approach and suggest that a com- bination of ontology-based and perception-based approaches may be optimal in order to find a suitable substitute for the unavailable objects.

  @MastersThesis{ 2015zolghadr,
  abstract  = { When a service robot is acting in a home environment,
  there are situations where some of the objects necessary
  for completing the task may not be available, because
  either they are missing or the preconditions for using them
  are not met. In such situations, humans show incredible
  flexibility by making substitutions. However, for a robot
  identifying the similarity between objects and finding an
  appropriate substitute for an unavailable object is not a
  straightforward task. Most of the approaches which tried to
  solve the problem have focused on lifting, where another
  instance of the same object class is used as a substitute.
  These approaches apply a limited range of semantics. Most
  approaches mea- sure similarity between objects using
  perception. Perception-based approaches tend to be
  time-consuming. However, the application of semantic-based
  similarity measures has been successfully and efficiently
  used in the genetics domain. The approach presented here
  combines similarity measurement with knowledge base queries
  in order to find appropriate substitutes. In our
  ontological approach, similarity measurement is carried out
  based on a broad range of objects semantics, including the
  functional affordances, object features, part-whole
  relations, and spatial proximity of the objects in the home
  environment. In this research, existing tools, previously
  used in the biological domains, for measuring the
  similarity of individuals within an ontology are evaluated
  and an analysis of the most useful measures is conducted.
  The experiments which were conducted were used to guide the
  creation of a methodology for the modeling of the objects
  within the knowledge base. The results show that the
  Jaccard Similarity Coefficient is particularly well-suited
  to our goal. The results also highlight the limitations of
  this approach and suggest that a com- bination of
  ontology-based and perception-based approaches may be
  optimal in order to find a suitable substitute for the
  unavailable objects. },
  address  = {Grantham-Allee 20, 53757 St. Augustin, Germany},
  annote  = {WS08/09 Kraetzschmar, Pl{\"o}ger, Awaad supervising},
  author  = {Mahan Zolghadr},
  month = {April},
  school  = {Bonn-Rhein-Sieg University of Applied Sciences},
  title = {Semantic Similarity Between Objects in Home Environments},
  year = {2015}
  }

• F. Kilic, “Application and improvement of the TRRL (Transport and Road Research Laboratory) high-speed laser profilometer algorithm with sensor fusion,” Master Thesis, Grantham-Allee 20, 53757 St. Augustin, Germany, 2015.
  [BibTeX] [Abstract]
  

  Maintenance of the highway networks is considered to be vital in order to ensure transportation safety and quality. Within different quality criterions used for a highway, road roughness stands out as the most important factor which deteriorates the driving comfort and endangers the passengers. High-speed profilometers are developed in order to be able to monitor road roughness on the highways without affecting the normal traffic flow. These devices record road elevation profiles while driving at highways speeds. Transportation and Road Research Laboratory (TRRL) has developed a unique high- speed profilometer system. Their design contains four laser displacement transducers which are placed along a long trailer to be towed by a vehicle. As the trailer moves forward, the leading sensor measures the road elevation and the other sensors provide a reference for the new measurements. This thesis focuses on the improvement and the implementation of TRRL-type high speed profilometers. In order to enable faster and safer operations on the highways, the geometric design of the original system is changed to make it more compact. Using a more compact design increases the observed pitch angles on the system and the increased angles are expected to introduce a new source of error for the measurements. The thesis investigates the effects of vehicle pitching on the measured profiles, and it suggests a method for pitch angle estimation in order to correct the laser measurements. The error analysis seen in the original work is extended and the algorithm is adjusted for the new geometric design. The factor that affects the measurement accuracy the most is determined to be the surface texture due to their randomness. The performance of the originally suggested method for eliminating the texture-caused errors is observed to be insufficient. Therefore, the thesis proposes a new method with a better performance which eliminates the texture-caused errors by modelling them with quadratic functions. The overall performance of the presented profilometer is evaluated by conducting experiments on a road with a known true profile. The accuracy and the repeatability of the observed results indicate that the developed profilometer can be used for measuring true profiles with some further improvement.

  @MastersThesis{ 2015kilic,
  abstract  = {Maintenance of the highway networks is considered to be
  vital in order to ensure transportation safety and quality.
  Within different quality criterions used for a highway,
  road roughness stands out as the most important factor
  which deteriorates the driving comfort and endangers the
  passengers. High-speed profilometers are developed in order
  to be able to monitor road roughness on the highways
  without affecting the normal traffic flow. These devices
  record road elevation profiles while driving at highways
  speeds.
  Transportation and Road Research Laboratory (TRRL) has
  developed a unique high- speed profilometer system. Their
  design contains four laser displacement transducers which
  are placed along a long trailer to be towed by a vehicle.
  As the trailer moves forward, the leading sensor measures
  the road elevation and the other sensors provide a
  reference for the new measurements.
  This thesis focuses on the improvement and the
  implementation of TRRL-type high speed profilometers. In
  order to enable faster and safer operations on the
  highways, the geometric design of the original system is
  changed to make it more compact. Using a more compact
  design increases the observed pitch angles on the system
  and the increased angles are expected to introduce a new
  source of error for the measurements. The thesis
  investigates the effects of vehicle pitching on the
  measured profiles, and it suggests a method for pitch angle
  estimation in order to correct the laser measurements. The
  error analysis seen in the original work is extended and
  the algorithm is adjusted for the new geometric design. The
  factor that affects the measurement accuracy the most is
  determined to be the surface texture due to their
  randomness. The performance of the originally suggested
  method for eliminating the texture-caused errors is
  observed to be insufficient. Therefore, the thesis proposes
  a new method with a better performance which eliminates the
  texture-caused errors by modelling them with quadratic
  functions.
  The overall performance of the presented profilometer is
  evaluated by conducting experiments on a road with a known
  true profile. The accuracy and the repeatability of the
  observed results indicate that the developed profilometer
  can be used for measuring true profiles with some further
  improvement.},
  address  = {Grantham-Allee 20, 53757 St. Augustin, Germany},
  annote  = {WS12 measX GmbH & Co. KG - BAST (High-Speed Profilometer)
  Ploeger, Breuer, Hilsmann supervising},
  author  = {Furkan Kilic},
  month = {January},
  school  = {Bonn-Rhein-Sieg University of Applied Sciences},
  title = {Application and improvement of the TRRL (Transport and
  Road Research Laboratory) high-speed laser profilometer
  algorithm with sensor fusion},
  year = {2015}
  }

• J. Sanchez, “Robust and Safe Manipulation by Sensor Fusion of Robotic Manipulators and End-Effectors,” Master Thesis, Grantham-Allee 20, 53757 St. Augustin, Germany, 2015.
  [BibTeX] [Abstract]
  

  A continuously increasing demand of staff able to take care of the elderly has generated interest in using robots as caregivers. However, the current state of the art limits the complexity of the tasks a robot can achieve. Another important constraint, is the ability of the robot to safely react to unexpected behaviors of a person. Thus, the first step is ensuring the robot can detect these events. In this work, a multi-sensor system based on design patterns is proposed to detect an object (e.g. a person’s arm) slipping from the robot’s grasp. Tactile sensors are used in combination with a force-torque sensor to provide complementary information. Over 500 experiments on a Care-O-bot 3 provided a comparative evaluation of a slip detector implementation. The results show an improved performance when combining both modalities (tactile and force). Furthermore, the proposed implementation proved is able to operate with different grasp shapes while maintaining a high accuracy. Lastly, the evaluation also exhibited the difficulties encountered at detecting motions of a human arm being grasped by a robot.

  @MastersThesis{ 2015sanchez,
  abstract  = {A continuously increasing demand of staff able to take
  care of the elderly has generated interest in using robots
  as caregivers. However, the current state of the art limits
  the complexity of the tasks a robot can achieve. Another
  important constraint, is the ability of the robot to safely
  react to unexpected behaviors of a person. Thus, the first
  step is ensuring the robot can detect these events. In this
  work, a multi-sensor system based on design patterns is
  proposed to detect an object (e.g. a person's arm) slipping
  from the robot's grasp. Tactile sensors are used in
  combination with a force-torque sensor to provide
  complementary information. Over 500 experiments on a
  Care-O-bot 3 provided a comparative evaluation of a slip
  detector implementation. The results show an improved
  performance when combining both modalities (tactile and
  force). Furthermore, the proposed implementation proved is
  able to operate with different grasp shapes while
  maintaining a high accuracy. Lastly, the evaluation also
  exhibited the difficulties encountered at detecting motions
  of a human arm being grasped by a robot.},
  address  = {Grantham-Allee 20, 53757 St. Augustin, Germany},
  annote  = {WS12/13 BRSU - RoboCup Pl{\"o}ger, Kraetzschmar,
  Schneider, supervising},
  author  = {Jose Sanchez},
  month = {March},
  school  = {Bonn-Rhein-Sieg University of Applied Sciences},
  title = {Robust and Safe Manipulation by Sensor Fusion of Robotic
  Manipulators and End-Effectors},
  year = {2015}
  }

• M. Vayugundla, “Experimental Evaluation and Improvement of a Viewframe-Based Navigation Method,” Master Thesis, Grantham-Allee 20, 53757 St. Augustin, Germany, 2015.
  [BibTeX] [Abstract]
  

  Insects like ants and bees navigate robustly in their environments in spite of their small brains using vision as their primary sensor. Inspired by this, researchers at DLR are working on a range free navigation system using visual features. This ability is specially useful for autonomous navigation in large environments and also for computationally limited small robots. Each location in the environment is represented as a viewframe. A viewframe is a set of landmark observations where each landmark observation contains landmark I.D, descriptor and corresponding angle with respect to the robot’s location. Binary Robust Invariant Scalable Keypoints (BRISK) features extracted from the omnidirectional images were used as landmarks in this work. The environment is represented as a Trail-Map which preserves the relationship between adjacent viewframes and is efficient at both storing and pruning the map size when required. This work experimentally evaluates the current system and improves it. In this work, as an extension to the Trail-Map representation, topological knowledge was extracted with the help of dimensionality reduction techniques and by defining dissimilarity measures between any two viewframes. Using this topological knowledge, a pose graph is developed adding edges between viewframes based on how close they are in addition to the adjacency connections. With the help of this map, shorter paths were identified for homing. The topological mapping pipeline was implemented on the robot and experiments were performed in both indoor and outdoor environments. The performance of different dissimilarity measures and dimensionality reduction techniques in building a topological map of viewframes was evaluated. The experiments showed that using this pose graph representation, the robot could take shorter paths which are a subset of the long exploration paths by using the intersections of the paths.

  @MastersThesis{ 2015vayugundla,
  abstract  = {Insects like ants and bees navigate robustly in their
  environments in spite of their small brains using vision as
  their primary sensor. Inspired by this, researchers at DLR
  are working on a range free navigation system using visual
  features. This ability is specially useful for autonomous
  navigation in large environments and also for
  computationally limited small robots.
  Each location in the environment is represented as a
  viewframe. A viewframe is a set of landmark observations
  where each landmark observation contains landmark I.D,
  descriptor and corresponding angle with respect to the
  robot's location. Binary Robust Invariant Scalable
  Keypoints (BRISK) features extracted from the
  omnidirectional images were used as landmarks in this work.
  The environment is represented as a Trail-Map which
  preserves the relationship between adjacent viewframes and
  is efficient at both storing and pruning the map size when
  required. This work experimentally evaluates the current
  system and improves it.
  In this work, as an extension to the Trail-Map
  representation, topological knowledge was extracted with
  the help of dimensionality reduction techniques and by
  defining dissimilarity measures between any two viewframes.
  Using this topological knowledge, a pose graph is developed
  adding edges between viewframes based on how close they are
  in addition to the adjacency connections. With the help of
  this map, shorter paths were identified for homing. The
  topological mapping pipeline was implemented on the robot
  and experiments were performed in both indoor and outdoor
  environments. The performance of different dissimilarity
  measures and dimensionality reduction techniques in
  building a topological map of viewframes was evaluated.
  The experiments showed that using this pose graph
  representation, the robot could take shorter paths which
  are a subset of the long exploration paths by using the
  intersections of the paths.},
  address  = {Grantham-Allee 20, 53757 St. Augustin, Germany},
  annote  = {SS11 Heiden, Kraetzschmar, Stelzer supervising},
  author  = {Mallikarjuna Vayugundla},
  month = {December},
  school  = {Bonn-Rhein-Sieg University of Applied Sciences},
  title = {Experimental Evaluation and Improvement of a
  Viewframe-Based Navigation Method},
  year = {2015}
  }

• D. Hernandez, “Robust Localization and Path-tracking for a Mobile Outdoor Robot,” Master Thesis, Grantham-Allee 20, 53757 St. Augustin, Germany, 2015.
  [BibTeX] [Abstract]
  

  The aim of this master’s thesis is the development of a robust localization and path-tracking system for mobile outdoor robots. The system is implemented in the context of a mobile robot as an assistive device. The developed localization system is suitable for robots with applicability in pedestrian and urban areas, running along walking or bicycles paths. The algorithm is an implementation of a particle filter framework. Data from low-cost GPS, odometry sensors, digital maps and the novelty of a visual road-detection algorithm are fused to estimate robot’s location. The results show a consistent estimated robot’s location with a close-loop error of about 1 meter of accuracy. The robustness of the approach is demonstrated by showing experimental results containing different map configurations highlighting the weaknesses of low-cost GPS and a good algorithm performance even with the unavailability of some of the data.

  @MastersThesis{ 2015hernandez,
  abstract  = {The aim of this master's thesis is the development of a
  robust localization and path-tracking system for mobile
  outdoor robots. The system is implemented in the context of
  a mobile robot as an assistive device. The developed
  localization system is suitable for robots with
  applicability in pedestrian and urban areas, running along
  walking or bicycles paths. The algorithm is an
  implementation of a particle filter framework. Data from
  low-cost GPS, odometry sensors, digital maps and the
  novelty of a visual road-detection algorithm are fused to
  estimate robot's location. The results show a consistent
  estimated robot's location with a close-loop error of about
  1 meter of accuracy. The robustness of the approach is
  demonstrated by showing experimental results containing
  different map configurations highlighting the weaknesses of
  low-cost GPS and a good algorithm performance even with the
  unavailability of some of the data.},
  address  = {Grantham-Allee 20, 53757 St. Augustin, Germany},
  annote  = {WS12/13 Locomotec - RUFUS Prassler, Asteroth, Blumenthal},
  author  = {David Hernandez},
  month = {June},
  school  = {Bonn-Rhein-Sieg University of Applied Sciences},
  title = {Robust Localization and Path-tracking for a Mobile Outdoor
  Robot},
  year = {2015}
  }

2014

• N. Srivastava, G. Hinton, A. Krizhevsky, I. Sutskever, and R. Salakhutdinov, “Dropout: A Simple Way to Prevent Neural Networks from Overfitting,” Journal of Machine Learning Research, vol. 15, pp. 1929-1958, 2014.
  [BibTeX]

  @Article{ nitish2014,
  author  = {Nitish Srivastava and Geoffrey Hinton and Alex Krizhevsky
  and Ilya Sutskever and Ruslan Salakhutdinov},
  title = {{Dropout: A Simple Way to Prevent Neural Networks from
  Overfitting}},
  journal  = {Journal of Machine Learning Research},
  year = {2014},
  volume  = {15},
  pages = {1929-1958}
  }

• A. Karpathy, G. Toderici, S. Shetty, T. Leung, R. Sukthankar, and L. Fei-Fei, “Large-scale Video Classification with Convolutional Neural Networks,” in The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2014, p. 1725 – 1732.
  [BibTeX]

  @InProceedings{ karpathy2014,
  author  = {Karpathy, Andrej and Toderici, George and Shetty, Sanketh
  and Leung, Thomas and Sukthankar, Rahul and Fei-Fei, Li},
  title = {{Large-scale Video Classification with Convolutional
  Neural Networks}},
  booktitle  = {The IEEE Conference on Computer Vision and Pattern
  Recognition (CVPR)},
  year = {2014},
  pages = {1725 -- 1732}
  }

• K. Simonyan and A. Zisserman, “Two-Stream Convolutional Networks for Action Recognition in Videos,” in Advances in Neural Information Processing Systems, Curran Associates, Inc., 2014, p. 568 – 576.
  [BibTeX]

  @InCollection{ simonyan2014,
  author  = {Simonyan, Karen and Zisserman, Andrew},
  title = {{Two-Stream Convolutional Networks for Action Recognition
  in Videos}},
  year = {2014},
  booktitle  = {Advances in Neural Information Processing Systems},
  pages = {568 -- 576},
  publisher  = {Curran Associates, Inc.}
  }

• D. Sun, S. Roth, and M. J. Black, “A Quantitative Analysis of Current Practices in Optical Flow Estimation and the Principles Behind Them,” International Journal of Computer Vision, vol. 106, p. 115 – 137, 2014.
  [BibTeX]

  @Article{ sun2013,
  author  = {Sun, Deqing and Roth, Stefan and Black, Michael J},
  journal  = {International Journal of Computer Vision},
  title = {{A Quantitative Analysis of Current Practices in Optical
  Flow Estimation and the Principles Behind Them}},
  year = {2014},
  pages = {115 -- 137},
  volume  = {106}
  }

• Z. Zhang, C. Conly, and V. Athitsos, “Evaluating Depth-Based Computer Vision Methods for Fall Detection under Occlusions,” in Advances in Visual Computing, 2014.
  [BibTeX]

  @InProceedings{ zhang2014,
  author  = {Zhang, Zhong and Conly, Christopher and Athitsos,
  Vassilis},
  title = {{Evaluating Depth-Based Computer Vision Methods for Fall
  Detection under Occlusions}},
  booktitle  = {Advances in Visual Computing},
  publisher  = {Springer International Publishing},
  year = {2014}
  }

• L. Wangm, Y. Qiao, and X. Tang, “Latent Hierarchical Model of Temporal Structure for Complex Activity Classification,” IEEE Transactions on Image Processing, vol. 23, p. 810 – 822, 2014.
  [BibTeX]

  @Article{ wang2014,
  author  = {Wangm, L. and Qiao, Y. and Tang, X.},
  journal  = {IEEE Transactions on Image Processing},
  title = {{Latent Hierarchical Model of Temporal Structure for
  Complex Activity Classification}},
  year = {2014},
  volume  = {23},
  pages = {810 -- 822}
  }

• C. Szegedy, W. Liu, Y. Jia, Pierre Sermanet, S. E. Reed, D. Anguelov, D. Erhan, V. Vanhoucke, and A. Rabinovich, “Going Deeper with Convolutions,” CoRR, p. 1 – 9, 2014.
  [BibTeX]

  @Article{ christian2014,
  author  = {Christian Szegedy and Wei Liu and Yangqing Jia and Pierre
  Sermanet and Scott E. Reed and Dragomir Anguelov and
  Dumitru Erhan and Vincent Vanhoucke and Andrew Rabinovich},
  title = {{Going Deeper with Convolutions}},
  journal  = {CoRR},
  year = {2014},
  pages = {1 -- 9}
  }

• R. K. Venkat, “Smart Person Counter in Mid-Ranging Environments,” Master Thesis, Grantham-Allee 20, 53757 St. Augustin, Germany, 2014.
  [BibTeX] [Abstract]
  

  Growing usage of train networks is pushing this vital public transport infrastructure to its physical capacity limitations in major cities globally. Capacity planning and people re-routing is increasingly becoming ubiquitous now a days to manage such increasingly crowded environments. With the current trends of growing technology, robotics and automation is an attractive option for solving this problem. A logical starting point to address the task of capacity planning is to understand the extent of the issue: for exam- ple, building automated algorithm for counting the number of people in such scenarios. Keeping this in mind, this thesis concentrates on achieving a robust people counting sys- tem which can handle real world scenarios posing very intricate and challenging situations. A few of the challenging situations are tracking multiple people walking across the sensor’s field of view(FOV) and counting them correctly within an acceptable time for the system to work in real time, tracking people moving very close to each other, tracking people with a variety of walking behaviors and tracking people walking past obstacles such as pillars etc. Out of the above mentioned challenges, this thesis mainly concentrates on the below two problems: 1. Counting multiple people (8-10 in number appearing simultaneously in a single ob- servation), in real time. 10 is approximately the maximum number of people which the primeSense sensor’s FOV can accommodate in a single observation within its working range. 2. Tracking closely moving people, in which case their corresponding blobs merge and the system no longer detects them belonging to two people. The first problem of counting multiple people appearing simultaneously in real time is handled by building the person counting system on top of already existing person detection systems and an existing particle filter. A background image subtraction step is added to improve the people detection rate. The latter problem of closely moving people is handled using an event graph-based approach with a step of identity mapping for which our approach uses the profile shape of people’s head and shoulders which has been shown to be useful in identifying individuals. The evaluation results show that the person counting system is robust enough to handle 8-10 people walking across the FOV in a single observation with an average execu- tion time of 240 ms and a maximum (worst case) execution time of 470 ms. Additionally, the system is robust to scenarios where people walk very closely to each other, maintaining an accurate person count. An empirical evaluation of the approach in a major Sydney public train station (N=522), and results demonstrating the methods in the complexities of this challenging environment are also presented. Furthermore, these results demon- strate that the novel methods contribute significantly to the person counting system, are real world viable and hence lay a foundation for the idea of people congestion awareness towards the goal of achieving efficient capacity planning.

  @MastersThesis{ 2014venkat,
  abstract  = {Growing usage of train networks is pushing this vital
  public transport infrastructure to its physical capacity
  limitations in major cities globally. Capacity planning and
  people re-routing is increasingly becoming ubiquitous now a
  days to manage such increasingly crowded environments. With
  the current trends of growing technology, robotics and
  automation is an attractive option for solving this
  problem. A logical starting point to address the task of
  capacity planning is to understand the extent of the issue:
  for exam- ple, building automated algorithm for counting
  the number of people in such scenarios. Keeping this in
  mind, this thesis concentrates on achieving a robust people
  counting sys- tem which can handle real world scenarios
  posing very intricate and challenging situations. A few of
  the challenging situations are tracking multiple people
  walking across the sensor's field of view(FOV) and counting
  them correctly within an acceptable time for the system to
  work in real time, tracking people moving very close to
  each other, tracking people with a variety of walking
  behaviors and tracking people walking past obstacles such
  as pillars etc. Out of the above mentioned challenges, this
  thesis mainly concentrates on the below two problems: 1.
  Counting multiple people (8-10 in number appearing
  simultaneously in a single ob- servation), in real time. 10
  is approximately the maximum number of people which the
  primeSense sensor's FOV can accommodate in a single
  observation within its working range. 2. Tracking closely
  moving people, in which case their corresponding blobs
  merge and the system no longer detects them belonging to
  two people. The first problem of counting multiple people
  appearing simultaneously in real time is handled by
  building the person counting system on top of already
  existing person detection systems and an existing particle
  filter. A background image subtraction step is added to
  improve the people detection rate. The latter problem of
  closely moving people is handled using an event graph-based
  approach with a step of identity mapping for which our
  approach uses the profile shape of people's head and
  shoulders which has been shown to be useful in identifying
  individuals. The evaluation results show that the person
  counting system is robust enough to handle 8-10 people
  walking across the FOV in a single observation with an
  average execu- tion time of 240 ms and a maximum (worst
  case) execution time of 470 ms. Additionally, the system is
  robust to scenarios where people walk very closely to each
  other, maintaining an accurate person count. An empirical
  evaluation of the approach in a major Sydney public train
  station (N=522), and results demonstrating the methods in
  the complexities of this challenging environment are also
  presented. Furthermore, these results demon- strate that
  the novel methods contribute significantly to the person
  counting system, are real world viable and hence lay a
  foundation for the idea of people congestion awareness
  towards the goal of achieving efficient capacity planning.
  },
  address  = {Grantham-Allee 20, 53757 St. Augustin, Germany},
  annote  = {SS11 FH-BRS - UTS,Sydney Pl{\"o}ger, Kraetzschmar,
  Kirchner supervising},
  author  = {Ravi Kumar Venkat},
  month = {May},
  school  = {Bonn-Rhein-Sieg University of Applied Sciences},
  title = {Smart Person Counter in Mid-Ranging Environments},
  year = {2014}
  }

• T. C. Hassan, “Dynamic Facial Expression Estimation by means of Model Fitting,” Master Thesis, Grantham-Allee 20, 53757 St. Augustin, Germany, 2014.
  [BibTeX] [Abstract]
  

  Analysis of facial expressions is an integral part of human behavioural research. The Facial Action Coding System (FACS) manual guides researchers in identifying and coding facial expressions in terms of basic facial movements called Action Units (AUs). However, coding faces manually, based on FACS, is a tedious process. Automating the FACS-based analysis of faces in images and image sequences would save a great amount of time, and thereby accelerate behavioural research. Automatic facial AU analysis would also be of value in developing technologies for affect-based human-computer (robot) interaction. This thesis deals with the problem of fully automatic estimation of AUs in image sequences. A model-based approach is pursued. The shape of the human face is represented in the form of an interconnected mesh of vertices. Linear models describe the shape in terms of deformation vectors controlled by a set of parameters. These deformation vectors correspond to changes in facial shape resulting from individual AUs. The parameters that control these deformations denote the intensity at which the AUs are expressed. Existing methods for model fitting can be used to determine the AU model parameters. However, these methods follow a frame-by-frame strategy and do not incorporate the dynamics of underlying motion. This causes two main problems. Firstly, the trajectories of estimated parameters are noisy. Secondly, ambiguities in AU parameter estimates cannot be resolved correctly. As a result, the AU estimation performance is poor. State estimation methods allow dynamic models of parameter evolution to be combined with noisy observations of textures or model vertices given by the model-fitting methods. In this thesis, the use of state estimation methods to improve AU estimation performance is investigated.

  @MastersThesis{ 2014hassan,
  abstract  = {Analysis of facial expressions is an integral part of
  human behavioural research. The Facial Action Coding System
  (FACS) manual guides researchers in identifying and coding
  facial expressions in terms of basic facial movements
  called Action Units (AUs). However, coding faces manually,
  based on FACS, is a tedious process. Automating the
  FACS-based analysis of faces in images and image sequences
  would save a great amount of time, and thereby accelerate
  behavioural research. Automatic facial AU analysis would
  also be of value in developing technologies for
  affect-based human-computer (robot) interaction. This
  thesis deals with the problem of fully automatic estimation
  of AUs in image sequences. A model-based approach is
  pursued. The shape of the human face is represented in the
  form of an interconnected mesh of vertices. Linear models
  describe the shape in terms of deformation vectors
  controlled by a set of parameters. These deformation
  vectors correspond to changes in facial shape resulting
  from individual AUs. The parameters that control these
  deformations denote the intensity at which the AUs are
  expressed. Existing methods for model fitting can be used
  to determine the AU model parameters. However, these
  methods follow a frame-by-frame strategy and do not
  incorporate the dynamics of underlying motion. This causes
  two main problems. Firstly, the trajectories of estimated
  parameters are noisy. Secondly, ambiguities in AU parameter
  estimates cannot be resolved correctly. As a result, the AU
  estimation performance is poor. State estimation methods
  allow dynamic models of parameter evolution to be combined
  with noisy observations of textures or model vertices given
  by the model-fitting methods. In this thesis, the use of
  state estimation methods to improve AU estimation
  performance is investigated.},
  address  = {Grantham-Allee 20, 53757 St. Augustin, Germany},
  annote  = {SS12 Fraunhofer IIS Prassler, Pl\"{o}ger, F\"{u}ller,
  Seuss supervising},
  author  = {Teena Chakkalayil Hassan},
  month = {June},
  school  = {Bonn-Rhein-Sieg University of Applied Sciences},
  title = {Dynamic Facial Expression Estimation by means of Model
  Fitting},
  year = {2014}
  }

• N. Giftsun, “‘Stack of Tasks’ Controller for Mobile Manipulators with a Path Planner,” Master Thesis, Grantham-Allee 20, 53757 St. Augustin, Germany, 2014.
  [BibTeX] [Abstract]
  

  Mobile Manipulators are usually redundant systems providing the flexibility of handling task space constraints. Prioritized Task Function based control schemes are quite popu- lar among the community of Humanoid Robots equipped with five times more than the degrees of freedom in a Mobile Manipulator. These control schemes don’t require a closed loop inverse kinematic solution and can avoid conflicts between multiple tasks. ‘Stack of Tasks'(SOT) is one such framework supported by state-of-the-art solver to handle equality and inequality constraints much efficiently. Though the task function based control strat- egy are quite attractive in terms of handling multiple tasks, they are only locally optimal. The controller needs the help of a global planner in avoiding the local minimum which is inherent in all Jacobian based controllers. This thesis focuses on developing a generic SOT controller for Mobile Manipulators capable of handling both motion-generation and path following tasks in real time scenarios. A study is done on the solver and the planning modules to bridge the planning component and the SOT controller. The performance of the controller is evaluated by experiments both in simulation and real PR2 robot.

  @MastersThesis{ 2014giftsun,
  abstract  = {Mobile Manipulators are usually redundant systems
  providing the flexibility of handling task space
  constraints. Prioritized Task Function based control
  schemes are quite popu- lar among the community of Humanoid
  Robots equipped with five times more than the degrees of
  freedom in a Mobile Manipulator. These control schemes
  don't require a closed loop inverse kinematic solution and
  can avoid conflicts between multiple tasks. 'Stack of
  Tasks'(SOT) is one such framework supported by
  state-of-the-art solver to handle equality and inequality
  constraints much efficiently. Though the task function
  based control strat- egy are quite attractive in terms of
  handling multiple tasks, they are only locally optimal. The
  controller needs the help of a global planner in avoiding
  the local minimum which is inherent in all Jacobian based
  controllers. This thesis focuses on developing a generic
  SOT controller for Mobile Manipulators capable of handling
  both motion-generation and path following tasks in real
  time scenarios. A study is done on the solver and the
  planning modules to bridge the planning component and the
  SOT controller. The performance of the controller is
  evaluated by experiments both in simulation and real PR2
  robot.},
  address  = {Grantham-Allee 20, 53757 St. Augustin, Germany},
  annote  = {SS2011 LAAS-CNRS - Factory in a Day Ploeger, Lamiraux,
  Kahl supervising},
  author  = {Nirmal Giftsun},
  month = {October},
  school  = {Bonn-Rhein-Sieg University of Applied Sciences},
  title = {‘Stack of Tasks' Controller for Mobile Manipulators with
  a Path Planner},
  year = {2014}
  }

• S. Junoh, “Development of a Cognitive UAV for Medical Assistance Application:Integration of a Soar Agent in ROS,” Master Thesis, Grantham-Allee 20, 53757 St. Augustin, Germany, 2014.
  [BibTeX] [Abstract]
  

  A scenario where a set of small-scale UAVs is used to deliver medical goods to assist injured or sick people in remote locations is proposed. A typical example would be a person having been bitten by a snake in a remote area and desperately needing snake antivenom. The UAV emergency system would then prepare a UAV with the snake antivenom and the UAV would then y to the injured person and deliver the medicine. As this medical assistance system should operate at all times and would often be based in remote locations, it is only feasible when it involves minimum human interaction. This is one of perfect scenarios by which UAV system benefits from a high degree of autonomy. The UAV has to handle the delivery of the medicine even in adverse conditions without relying on human intervention. Fortunately, over the past decades, there has been plenty of progress in the research of autonomy allowing us nowadays to design vehicles with an ever increasing degree of autonomy. One research effort in the eld of autonomy is cognitive architectures, with Soar being among the prominent. Soar tries to model the human cognitive process and ability in a software architecture. Their eorts (since 1983) resulted in a exible cognitive architecture which is available as Open Source software. In this thesis, the use of Soar in the context of the medical assistance UAV is investigated. While all autonomy functionalities are handled by the Soar framework, general capabilities and the middleware of the UAV are modeled using the widely used robot framework ROS. ROS has been providing many modules which allow for ying and navigating small-scale UAVs and therefore enabling a fast prototype development of the UAV software. In this thesis, both the system and the cognitive functionalities of the autonomous medical assistance UAV are designed. Consequently, a major part of this thesis is the integration of Soar into ROS (ROSied Soar). Starting from an overall architecture design including the UAV systems and the cognitive agent, a software implementation was derived. Finally, the UAV model was tested in a simulation environment (ROS/Gazebo) where the UAV performed a delivery mission. The simulation included the straightforward regular delivery as well as missions where stressors were applied which forced the Soar agent to react and make alternative decisions. Preliminary simulation data reveal that this approach has the potential for creating such a medical assistance UAV system. This also means that the synergy between Soar and ROS has been achieved, hence, has shown the usefulness of this integration for the use of UAVs deployed for complex mission.

  @MastersThesis{ 2014junoh,
  abstract  = {A scenario where a set of small-scale UAVs is used to
  deliver medical goods to assist injured or sick people in
  remote locations is proposed. A typical example would be a
  person having been bitten by a snake in a remote area and
  desperately needing snake antivenom. The UAV emergency
  system would then prepare a UAV with the snake antivenom
  and the UAV would then y to the injured person and deliver
  the medicine. As this medical assistance system should
  operate at all times and would often be based in remote
  locations, it is only feasible when it involves minimum
  human interaction. This is one of perfect scenarios by
  which UAV system benefits from a high degree of autonomy.
  The UAV has to handle the delivery of the medicine even in
  adverse conditions without relying on human intervention.
  Fortunately, over the past decades, there has been plenty
  of progress in the research of autonomy allowing us
  nowadays to design vehicles with an ever increasing degree
  of autonomy. One research effort in the eld of autonomy is
  cognitive architectures, with Soar being among the
  prominent. Soar tries to model the human cognitive process
  and ability in a software architecture. Their eorts (since
  1983) resulted in a exible cognitive architecture which is
  available as Open Source software. In this thesis, the use
  of Soar in the context of the medical assistance UAV is
  investigated. While all autonomy functionalities are
  handled by the Soar framework, general capabilities and the
  middleware of the UAV are modeled using the widely used
  robot framework ROS. ROS has been providing many modules
  which allow for ying and navigating small-scale UAVs and
  therefore enabling a fast prototype development of the UAV
  software. In this thesis, both the system and the cognitive
  functionalities of the autonomous medical assistance UAV
  are designed. Consequently, a major part of this thesis is
  the integration of Soar into ROS (ROSied Soar). Starting
  from an overall architecture design including the UAV
  systems and the cognitive agent, a software implementation
  was derived. Finally, the UAV model was tested in a
  simulation environment (ROS/Gazebo) where the UAV performed
  a delivery mission. The simulation included the
  straightforward regular delivery as well as missions where
  stressors were applied which forced the Soar agent to react
  and make alternative decisions. Preliminary simulation data
  reveal that this approach has the potential for creating
  such a medical assistance UAV system. This also means that
  the synergy between Soar and ROS has been achieved, hence,
  has shown the usefulness of this integration for the use of
  UAVs deployed for complex mission.},
  address  = {Grantham-Allee 20, 53757 St. Augustin, Germany},
  annote  = {SS09 Silver Atena Electronic Systems Engineering GmbH -
  Autonomy Kreatzschmar, Heni, Stenger supervising},
  author  = {Shahmi Junoh},
  month = {November},
  school  = {Bonn-Rhein-Sieg University of Applied Sciences},
  title = {Development of a Cognitive UAV for Medical Assistance
  Application:Integration of a Soar Agent in ROS},
  year = {2014}
  }

• Pramanujam, “Robust Navigation of a Mobile Manipulator in a Dynamic Environment,” Master Thesis, Grantham-Allee 20, 53757 St. Augustin, Germany, 2014.
  [BibTeX] [Abstract]
  

  Obstacle avoidance is one of the essential components to achieve autonomy in robot navigation. Algorithms for obstacle avoidance have been developed for over past two decades and the recent advances made in these algorithms do not only deal static but also dynamic obstacles. However, in the case of highly cluttered environments, these algorithms stop navigating the robot and wait for the environment to clear, which is undesirable in crowded environments. Another situation during which these algorithms break is when the external sensors fail or are unable to perceive the environment. The condition of the robot in such scenarios is similar to a human navigating in the dark. The thesis is motivated by the above arguments and addresses the problem similar to how human copes up with trying to find a path in a heavily crowded environment or in the dark. The sense of touch and pressure is used to recover from such scenarios. In order to achieve this, the concept of \emph{compliance} has been adapted from the field of robot manipulation and applied to robot navigation, whereby the collision is considered as an external force being exerted on the system. The thesis proposes a coherent framework to combine compliance in navigation to the existing navigation algorithms, thereby, allowing robots to deal with real as well as virtual forces of the environment. In order to realize the framework, kinetic analysis on a four wheeled omni-directional robot was performed. As a result, the robot can be torque controlled given a certain path to be followed. This provides an estimate of the force exerted by the robot for its motion. A disturbance observer based on robot’s momentum was implemented in order to detect undesirable collisions. The experiments show encouraging results for detecting obstacles at low velocity.

  @MastersThesis{ 2014ramanujam,
  abstract  = {Obstacle avoidance is one of the essential components to
  achieve autonomy in robot navigation. Algorithms for
  obstacle avoidance have been developed for over past two
  decades and the recent advances made in these algorithms do
  not only deal static but also dynamic obstacles. However,
  in the case of highly cluttered environments, these
  algorithms stop navigating the robot and wait for the
  environment to clear, which is undesirable in crowded
  environments. Another situation during which these
  algorithms break is when the external sensors fail or are
  unable to perceive the environment. The condition of the
  robot in such scenarios is similar to a human navigating in
  the dark.
  The thesis is motivated by the above arguments and
  addresses the problem similar to how human copes up with
  trying to find a path in a heavily crowded environment or
  in the dark. The sense of touch and pressure is used to
  recover from such scenarios. In order to achieve this, the
  concept of \emph{compliance} has been adapted from the
  field of robot manipulation and applied to robot
  navigation, whereby the collision is considered as an
  external force being exerted on the system.
  The thesis proposes a coherent framework to combine
  compliance in navigation to the existing navigation
  algorithms, thereby, allowing robots to deal with real as
  well as virtual forces of the environment. In order to
  realize the framework, kinetic analysis on a four wheeled
  omni-directional robot was performed. As a result, the
  robot can be torque controlled given a certain path to be
  followed. This provides an estimate of the force exerted by
  the robot for its motion. A disturbance observer based on
  robot's momentum was implemented in order to detect
  undesirable collisions. The experiments show encouraging
  results for detecting obstacles at low velocity. },
  address  = {Grantham-Allee 20, 53757 St. Augustin, Germany},
  annote  = {SS11 FH-BRS Pl{\"o}ger, Prassler, Blumenthal},
  author  = {Pramanujam},
  month = {August},
  school  = {Bonn-Rhein-Sieg University of Applied Sciences},
  title = {Robust Navigation of a Mobile Manipulator in a Dynamic
  Environment },
  year = {2014}
  }

• M. Valdenegro, “Fast Text Detection for Road Scenes,” Master Thesis, Grantham-Allee 20, 53757 St. Augustin, Germany, 2014.
  [BibTeX] [Abstract]
  

  Extraction of text information from visual sources is an important component of many modern applications, for example, extracting the text from traffic signs on a road scene in an autonomous vehicle. For natural images or road scenes this is a unsolved problem. In this thesis the use of histogram of stroke widths (HSW) for character and non-character region classification is presented. Stroke widths are extracted using two methods. One is based on the Stroke Width Transform and another based on run lengths. The HSW is combined with two simple region features– aspect and occupancy ratios– and then a linear SVM is used as classifier. One advantage of our method over the state of the art is that it is script-independent and can also be used to verify detected text regions with the purpose of reducing false positives. Our experiments on generated datasets of Latin, CJK, Hiragana and Katakana characters show that the HSW is able to correctly classify at least 90 % of the character regions, a similar figure is obtained for non-character regions. This performance is also obtained when training the HSW with one script and testing with a different one, and even when characters are rotated. On the English and Kannada portions of the Chars74K dataset we obtained over 95% correctly classified character regions. The use of raycasting for text line grouping is also proposed. By combining it with our HSW-based character classifier, a text detector based on Maximally Stable Extremal Regions (MSER) was implemented. The text detector was evaluated on our own dataset of road scenes from the German Autobahn, where 65% precision, 72% recall with a f-score of 69% was obtained. Using the HSW as a text verifier increases precision while slightly reducing recall. Our HSW feature allows the building of a script-independent and low parameter count classifier for character and non-character regions

  @MastersThesis{ 2014valdenegro,
  abstract  = {Extraction of text information from visual sources is an
  important component of many modern applications, for
  example, extracting the text from traffic signs on a road
  scene in an autonomous vehicle. For natural images or road
  scenes this is a unsolved problem. In this thesis the use
  of histogram of stroke widths (HSW) for character and
  non-character region classification is presented. Stroke
  widths are extracted using two methods. One is based on the
  Stroke Width Transform and another based on run lengths.
  The HSW is combined with two simple region features–
  aspect and occupancy ratios– and then a linear SVM is
  used as classifier. One advantage of our method over the
  state of the art is that it is script-independent and can
  also be used to verify detected text regions with the
  purpose of reducing false positives. Our experiments on
  generated datasets of Latin, CJK, Hiragana and Katakana
  characters show that the HSW is able to correctly classify
  at least 90 % of the character regions, a similar figure is
  obtained for non-character regions. This performance is
  also obtained when training the HSW with one script and
  testing with a different one, and even when characters are
  rotated. On the English and Kannada portions of the
  Chars74K dataset we obtained over 95% correctly classified
  character regions. The use of raycasting for text line
  grouping is also proposed. By combining it with our
  HSW-based character classifier, a text detector based on
  Maximally Stable Extremal Regions (MSER) was implemented.
  The text detector was evaluated on our own dataset of road
  scenes from the German Autobahn, where 65% precision, 72%
  recall with a f-score of 69% was obtained. Using the HSW as
  a text verifier increases precision while slightly reducing
  recall. Our HSW feature allows the building of a
  script-independent and low parameter count classifier for
  character and non-character regions},
  address  = {Grantham-Allee 20, 53757 St. Augustin, Germany},
  annote  = {WS2012 - Fraunhofer IAIS Pl{\"o}ger, Kraetzschmar,
  Eickeler supervising},
  author  = {Matias Valdenegro},
  month = {September},
  school  = {Bonn-Rhein-Sieg University of Applied Sciences},
  title = {Fast Text Detection for Road Scenes},
  year = {2014}
  }

• O. Alaqtash, “Creating a Focused HTN Planning Problem using DL Inferencing,” Master Thesis, Grantham-Allee 20, 53757 St. Augustin, Germany, 2014.
  [BibTeX] [Abstract]
  

  Automated planning allows systems to automatically generate plans in order to execute tasks or achieve goals. To accomplish this, planners use knowledge of the actions that can be performed, when they can be carried out as well as how these actions affect the world. This knowledge represents the planning domain. Moreover, planners need a description of the world. A planning problem consists of these two items as well as the task to be achieved. Hierarchical Task Network (HTN) planning is a popular automated planning approach. It encodes domain information that defines recipes on how to carry out a task. This results in a pruned search space and provides quality plans. However, the search space remains large and this can lead to intractability. Many real-world planning problems are still difficult to solve in a reasonable time by autonomous agents. The reason for this is that the planning problems are not preprocessed to include only the relevant parts of the domain and the description of the world. This preprocessing is carried out by a human for systems without full autonomy. Researchers continue to optimize planning approaches and planners, but very few have tackled the problem of identifying relevant parts of the planning problem. The work of Hartanto (2011) is among the first to propose a solution which involves modeling the planning domain in the Web Ontology Language (OWL). However, his solution has a number of deficits. Most importantly, his modeling of the planning domain is not flexible enough to be reused. In this work, we adopt Hartanto’s approach of representing the planning domain in OWL to enable inference and create a focused planning problem. To do this, we model the planning domain and the description of the world in a decidable fragment of OWL, namely OWL 2 DL, and create a component that queries this knowledge by generating a set of SPARQL-DL queries to identify the relevant parts of the description of the world and the domain. The solution is able to generate a focused planning problem in OWL 2 DL, which can later be transformed into the JSHOP2 syntax for the planner to use. The results of our experiments showed a reduction of between 58% to 81% from the original size of the planning problem. Such a decrease in the problem size should lead to a much faster plan generation process.

  @MastersThesis{ 2014alaqtash,
  abstract  = {Automated planning allows systems to automatically
  generate plans in order to execute tasks or achieve goals.
  To accomplish this, planners use knowledge of the actions
  that can be performed, when they can be carried out as well
  as how these actions affect the world. This knowledge
  represents the planning domain. Moreover, planners need a
  description of the world. A planning problem consists of
  these two items as well as the task to be achieved.
  Hierarchical Task Network (HTN) planning is a popular
  automated planning approach. It encodes domain information
  that defines recipes on how to carry out a task. This
  results in a pruned search space and provides quality
  plans. However, the search space remains large and this can
  lead to intractability. Many real-world planning problems
  are still difficult to solve in a reasonable time by
  autonomous agents. The reason for this is that the planning
  problems are not preprocessed to include only the relevant
  parts of the domain and the description of the world. This
  preprocessing is carried out by a human for systems without
  full autonomy. Researchers continue to optimize planning
  approaches and planners, but very few have tackled the
  problem of identifying relevant parts of the planning
  problem. The work of Hartanto (2011) is among the first to
  propose a solution which involves modeling the planning
  domain in the Web Ontology Language (OWL). However, his
  solution has a number of deficits. Most importantly, his
  modeling of the planning domain is not flexible enough to
  be reused. In this work, we adopt Hartanto's approach of
  representing the planning domain in OWL to enable inference
  and create a focused planning problem. To do this, we model
  the planning domain and the description of the world in a
  decidable fragment of OWL, namely OWL 2 DL, and create a
  component that queries this knowledge by generating a set
  of SPARQL-DL queries to identify the relevant parts of the
  description of the world and the domain. The solution is
  able to generate a focused planning problem in OWL 2 DL,
  which can later be transformed into the JSHOP2 syntax for
  the planner to use. The results of our experiments showed a
  reduction of between 58% to 81% from the original size of
  the planning problem. Such a decrease in the problem size
  should lead to a much faster plan generation process.},
  address  = {Grantham-Allee 20, 53757 St. Augustin, Germany},
  annote  = {WS12/13 H-BRS - Creating a Focused HTN Planning Problem
  using DL Inferencing Kr{\"a}tzschmar, Pl{\"o}ger, Awaad
  supervising},
  author  = {Obada Alaqtash},
  month = {December},
  school  = {Bonn-Rhein-Sieg University of Applied Sciences},
  title = {Creating a Focused HTN Planning Problem using DL
  Inferencing},
  year = {2014}
  }

2013

• M. Mubashir, L. Shao, and L. Seed, “A Survey on Fall Detection: Principles and Approaches,” Neurocomputing, vol. 100, p. 144 – 152, 2013.
  [BibTeX]

  @Article{ mubashir2013,
  author  = {Mubashir, Muhammad and Shao, Ling and Seed, Luke},
  title = {{A Survey on Fall Detection: Principles and Approaches}},
  journal  = {Neurocomputing},
  volume  = {100},
  pages = {144 -- 152},
  year = {2013},
  issn = {0925-2312}
  }

• I. Charfi, J. Miteran, J. Dubois, M. Atri, and R. Tourki, “Optimized Spatio-temporal Descriptors for Real-time Fall Detection: Comparison of Support Vector Machine and Adaboost-based Classification,” Journal of Electronic Imaging, vol. 22, p. 1 – 18, 2013.
  [BibTeX]

  @Article{ tourki2013,
  author  = {Imen Charfi and Johel Miteran and Julien Dubois and
  Mohamed Atri and Rached Tourki},
  title = {{Optimized Spatio-temporal Descriptors for Real-time Fall
  Detection: Comparison of Support Vector Machine and
  Adaboost-based Classification}},
  volume  = {22},
  journal  = {Journal of Electronic Imaging},
  publisher  = {SPIE},
  pages = {1 -- 18},
  year = {2013}
  }

• M. Vallejo, C. V. Isaza, and J. D. López, “Artificial Neural Networks as an Alternative to Traditional Fall Detection Methods,” in Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC), 2013, pp. 1648-1651.
  [BibTeX]

  @InProceedings{ vallejo2013,
  author  = {Vallejo, Marcela and Isaza, Claudia V. and L{\'{o}}pez,
  Jos{\'{e}} D.},
  booktitle  = {Annual International Conference of the IEEE Engineering in
  Medicine and Biology Society (EMBC)},
  title = {{Artificial Neural Networks as an Alternative to
  Traditional Fall Detection Methods}},
  pages = {1648-1651},
  year = {2013}
  }

• M. Lin, Q. Chen, and S. Yan, “Network In Network,” Neural and Evolutionary Computing, 2013.
  [BibTeX]

  @Article{ lin2013,
  author  = {Lin, Min and Chen, Qiang and Yan, Shuicheng},
  title = {{Network In Network}},
  year = {2013},
  journal  = {Neural and Evolutionary Computing}
  }

• M. W. Butt, “A computational visual attention system for diver’s hand signs and gestures recognition,” Master Thesis, Grantham Allee 20, 53757 St. Augustin, Germany, 2013.
  [BibTeX]

  @MastersThesis{ butt2013a-computational,
  address  = {Grantham Allee 20, 53757 St. Augustin, Germany},
  annote  = {[2013] [Fintrope], [Ploeger] supervising},
  author  = {Mohsin Wahab Butt},
  date-added  = {2016-08-28 08:11:03 +0000},
  date-modified  = {2016-08-28 08:13:31 +0000},
  month = {August},
  school  = {Bonn-Rhein-Sieg University of Applied Sciences},
  title = {A computational visual attention system for diver's hand
  signs and gestures recognition},
  year = {2013}
  }

• M. Pathare, “Household Device Recognition & Information Retrieval using a Smart-phone,” Master Thesis, Grantham-Allee 20, 53757 St. Augustin, Germany, 2013.
  [BibTeX] [Abstract]
  

  Present household devices are not like their predecessors, having a single on/off switch, instead they have numerous settings and functionalities. User’s need to have manual of such devices handy. In case they forget which, and how to adjust the device parameters or how to carry out certain ‘not so common tasks’ on the device. Sometimes the device interfaces can be difficult to understand, not only for elders but also for common people. Having the device’s operational information on our finger tips is a major requirement and also the information should be easily accessible. This issue is addressed in our work. A smart-phone can be a very handy tool in such situations. We have developed an application on a smart-phone that can recognize known household devices, retrieve its operational information, and present it to the user in an interactive way. Smaller, vision deprived robots can also benefit from our application. They can outsource the device recognition task to our application and will receive a feedback about the device and its details. Vision based techniques are used to find the device, that user needs information on. Feature base approach is used to recognize the device. The device’s operational information is stored in local database and is retrieved after its successful recognition. This information contains functionality of the controls present on the device and the tasks that can be performed with it. Users can view this information interactively. Robots can access the location information of the recognized device in the scene image. After evaluation, it was found that our application is robust in recognizing devices under illumination, scale and rotation variations and also under slight occlusions. No false positive detections were observed during tests, due to geometrical and text based validation checks. For five devices in database, application takes about 20 seconds to recognize a device, including focusing and image capturing time. The observed mean square error in estimating distance to device and horizontal angle to the device measured from optical axis, were 0.3097cm and 1.393$^\circ$ respectively. Our application proves useful to users, as well as robots in getting to know a device’s operational information. An end-to-end prototype was also developed, showing a robot mounted with a smart-phone running our application, searching for a device, recognizing it, localizing it and grasping it successfully.

  @MastersThesis{ 2013pathare,
  abstract  = {Present household devices are not like their predecessors,
  having a single on/off switch, instead they have numerous
  settings and functionalities. User's need to have manual of
  such devices handy. In case they forget which, and how to
  adjust the device parameters or how to carry out certain
  'not so common tasks' on the device. Sometimes the device
  interfaces can be difficult to understand, not only for
  elders but also for common people.
  Having the device's operational information on our finger
  tips is a major requirement and also the information should
  be easily accessible. This issue is addressed in our work.
  A smart-phone can be a very handy tool in such situations.
  We have developed an application on a smart-phone that can
  recognize known household devices, retrieve its operational
  information, and present it to the user in an interactive
  way. Smaller, vision deprived robots can also benefit from
  our application. They can outsource the device recognition
  task to our application and will receive a feedback about
  the device and its details.
  Vision based techniques are used to find the device, that
  user needs information on. Feature base approach is used to
  recognize the device. The device's operational information
  is stored in local database and is retrieved after its
  successful recognition. This information contains
  functionality of the controls present on the device and the
  tasks that can be performed with it. Users can view this
  information interactively. Robots can access the location
  information of the recognized device in the scene image.
  After evaluation, it was found that our application is
  robust in recognizing devices under illumination, scale and
  rotation variations and also under slight occlusions. No
  false positive detections were observed during tests, due
  to geometrical and text based validation checks. For five
  devices in database, application takes about 20 seconds to
  recognize a device, including focusing and image capturing
  time. The observed mean square error in estimating distance
  to device and horizontal angle to the device measured from
  optical axis, were 0.3097cm and 1.393$^\circ$ respectively.
  Our application proves useful to users, as well as robots
  in getting to know a device's operational information. An
  end-to-end prototype was also developed, showing a robot
  mounted with a smart-phone running our application,
  searching for a device, recognizing it, localizing it and
  grasping it successfully.},
  address  = {Grantham-Allee 20, 53757 St. Augustin, Germany},
  annote  = {WS10 H-BRS Pl{\"o}ger,Breuer supervising},
  author  = {Mandar Pathare},
  month = {April},
  school  = {Bonn-Rhein-Sieg University of Applied Sciences},
  title = {Household Device Recognition & Information Retrieval using
  a Smart-phone},
  year = {2013}
  }

• F. Rouatbi, “Two SVM-based methods for the classification of airborne LiDAR data,” Master Thesis, Grantham-Allee 20, 53757 St. Augustin, Germany, 2013.
  [BibTeX] [Abstract]
  

  Airborne Light Detection And Ranging (LiDAR) is a remote sensing method used to collect high resolution information of the earth’s surface. This technology uses laser light beams emitted from a LiDAR system mounted on an airborne platform to scan the landscape topology and the different objects above the bare-earth such as buildings, vegetation, cars and roads. The collected data is an elevation model in the form of a point cloud which is useful for a wide range of applications such as 3D modeling, change detection analysis and objects recognition. In order to extract relevant information from the scanned areas, the generated point cloud needs to be classified. In this thesis, we propose two different methods for the classification of airborne LiDAR data into ground, trees, and buildings. The first method is a point-based classification approach which works on each point independently and uses geometrical features derived from the local neighborhood to make a decision about the class attributes. Therefore, a classification system composed of a cascade of three independent binary classifiers (tree, ground and buildings) has been implemented. Each classifier is based on a support vector machine (SVM) which recognizes the points of a particular class and removes them from the dataset. First, the tree classifier is used to detects the tree points as they have the most discriminating features. Then, the remaining non-tree points are passed to the ground classifier which recognizes the ground points and eliminates them from the dataset. Finally, the building classifier is applied in order to separate the points belonging to the buildings from other small objects which have no predefined class such as cars. The second method is an object-based classification approach which segments the data and performs a classification of of the resulting segments. The surface growing algorithm was used to segment the proximal points having geometrical similarities into a set of disjoint objects. The ground objects are first identified based on the size and the median height of the points in each segment. Then, a SVM-based classifier is used to distinguish the objects corresponding to the buildings. Therefore, a set of object-based features derived from the geometrical attributes of the points within the same segment has been used. Both methods have been tested on a labeled dataset composed of more than one million LiDAR points and applied to a set of real data.

  @MastersThesis{ 2013rouatbi,
  abstract  = {Airborne Light Detection And Ranging (LiDAR) is a remote
  sensing method used to collect high resolution information
  of the earth's surface. This technology uses laser light
  beams emitted from a LiDAR system mounted on an airborne
  platform to scan the landscape topology and the different
  objects above the bare-earth such as buildings, vegetation,
  cars and roads. The collected data is an elevation model in
  the form of a point cloud which is useful for a wide range
  of applications such as 3D modeling, change detection
  analysis and objects recognition. In order to extract
  relevant information from the scanned areas, the generated
  point cloud needs to be classified. In this thesis, we
  propose two different methods for the classification of
  airborne LiDAR data into ground, trees, and buildings. The
  first method is a point-based classification approach which
  works on each point independently and uses geometrical
  features derived from the local neighborhood to make a
  decision about the class attributes. Therefore, a
  classification system composed of a cascade of three
  independent binary classifiers (tree, ground and buildings)
  has been implemented. Each classifier is based on a support
  vector machine (SVM) which recognizes the points of a
  particular class and removes them from the dataset. First,
  the tree classifier is used to detects the tree points as
  they have the most discriminating features. Then, the
  remaining non-tree points are passed to the ground
  classifier which recognizes the ground points and
  eliminates them from the dataset. Finally, the building
  classifier is applied in order to separate the points
  belonging to the buildings from other small objects which
  have no predefined class such as cars. The second method is
  an object-based classification approach which segments the
  data and performs a classification of of the resulting
  segments. The surface growing algorithm was used to segment
  the proximal points having geometrical similarities into a
  set of disjoint objects. The ground objects are first
  identified based on the size and the median height of the
  points in each segment. Then, a SVM-based classifier is
  used to distinguish the objects corresponding to the
  buildings. Therefore, a set of object-based features
  derived from the geometrical attributes of the points
  within the same segment has been used. Both methods have
  been tested on a labeled dataset composed of more than one
  million LiDAR points and applied to a set of real data.},
  address  = {Grantham-Allee 20, 53757 St. Augustin, Germany},
  annote  = {SS08 FH-BRS - Fraunhofer FKIE Two SVM-based methods for
  the classification of airborne LiDAR data Ploeger, Koch
  supervising},
  author  = {Fahmi Rouatbi},
  month = {November},
  school  = {Bonn-Rhein-Sieg University of Applied Sciences},
  title = {Two SVM-based methods for the classification of airborne
  LiDAR data},
  year = {2013}
  }

• S. Schneider, “Design of a declarative language for task-oriented grasping and tool-use with dextrous robotic hands,” Master Thesis, Grantham-Allee 20, 53757 St. Augustin, Germany, 2013.
  [BibTeX] [Abstract]
  

  Apparently simple manipulation tasks for a human such as transportation or tool use are challenging to replicate in an autonomous service robot. Nevertheless, dextrous ma- nipulation is an important aspect for a robot in many daily tasks. While it is possible to manufacture special-purpose hands for one specific task in industrial settings, a general- purpose service robot in households must have flexible hands which can adapt to many tasks. Intelligently using tools enables the robot to perform tasks more efficiently and even beyond the designed capabilities. In this work a declarative domain-specific language, called Grasp Domain Definition Language (GDDL), is presented that allows the specification of grasp planning problems independently of a specific grasp planner. This design goal resembles the idea of the Planning Domain Definition Language (PDDL). The specification of GDDL requires a detailed analysis of the research in grasping in order to identify best practices in different domains that contribute to a grasp. These domains describe for instance physical as well as semantic properties of objects and hands. Grasping always has a purpose which is captured in the task domain definition. It enables the robot to grasp an object in a task- dependent manner. Suitable representations in these domains have to be identified and formalized for which a domain-driven software engineering approach is applied. This kind of modeling allows the specification of constraints which guide the composition of domain entity specifications. The domain-driven approach fosters reuse of domain concepts while the constraints enable the validation of models already during design time. A proof of concept implementation of GDDL into the GraspIt! grasp planner is developed. Preliminary results of this thesis have been published and presented on the IEEE International Conference on Robotics and Automation (ICRA).

  @MastersThesis{ 2013schneider,
  abstract  = {Apparently simple manipulation tasks for a human such as
  transportation or tool use are challenging to replicate in
  an autonomous service robot. Nevertheless, dextrous ma-
  nipulation is an important aspect for a robot in many daily
  tasks. While it is possible to manufacture special-purpose
  hands for one specific task in industrial settings, a
  general- purpose service robot in households must have
  flexible hands which can adapt to many tasks. Intelligently
  using tools enables the robot to perform tasks more
  efficiently and even beyond the designed capabilities. In
  this work a declarative domain-specific language, called
  Grasp Domain Definition Language (GDDL), is presented that
  allows the specification of grasp planning problems
  independently of a specific grasp planner. This design goal
  resembles the idea of the Planning Domain Definition
  Language (PDDL). The specification of GDDL requires a
  detailed analysis of the research in grasping in order to
  identify best practices in different domains that
  contribute to a grasp. These domains describe for instance
  physical as well as semantic properties of objects and
  hands. Grasping always has a purpose which is captured in
  the task domain definition. It enables the robot to grasp
  an object in a task- dependent manner. Suitable
  representations in these domains have to be identified and
  formalized for which a domain-driven software engineering
  approach is applied. This kind of modeling allows the
  specification of constraints which guide the composition of
  domain entity specifications. The domain-driven approach
  fosters reuse of domain concepts while the constraints
  enable the validation of models already during design time.
  A proof of concept implementation of GDDL into the GraspIt!
  grasp planner is developed. Preliminary results of this
  thesis have been published and presented on the IEEE
  International Conference on Robotics and Automation
  (ICRA).},
  address  = {Grantham-Allee 20, 53757 St. Augustin, Germany},
  note = {WS09/10 H-BRS - RoboCup Kraetzschmar, Pl{\"o}ger,
  Hochgeschwender},
  author  = {Sven Schneider},
  month = {May},
  school  = {Bonn-Rhein-Sieg University of Applied Sciences},
  title = {Design of a declarative language for task-oriented
  grasping and tool-use with dextrous robotic hands},
  year = {2013}
  }

• Z. B. Kasim, “HMM-Based Diagnosis of Known Exogenous Interventions in Mobile Manipulators,” Master Thesis, Grantham-Allee 20, 53757 St. Augustin, Germany, 2013.
  [BibTeX] [Abstract]
  

  A mobile manipulator lives in a dynamic environment where changes occur over time. The changes are external, where they occur outside the mobile manipulator’s control, however, they affect the results for a task assigned for the mobile manipulator. These external events are called exogenous interventions. The main purpose of this thesis is to provide a model based, probabilistic method as an approach for known exogenous interventions in mobile manipulators. With the thought of the mobile platform and manipulator can work in parallel, we proposed two variations of continuous observation hidden Markov model, namely the classical hidden Markov model and the parallel hidden Markov model. We elaborated several use cases for exogenous interventions and with these we collected data from real and simulated mobile manipulators. Due the characteristics of exogenous interventions which are rare, random and versatile; we also obtained statistics for exogenous interventions to build up the model. We then tested the models on a simulated environment. The result shows that diagnosing exogenous interventions are precise and sensitive. With correct modeling, the hidden Markov Model is proven to be able to diagnose the exogenous interventions correctly.

  @MastersThesis{ 2013kasim,
  abstract  = {A mobile manipulator lives in a dynamic environment where
  changes occur over time. The changes are external, where
  they occur outside the mobile manipulator's control,
  however, they affect the results for a task assigned for
  the mobile manipulator. These external events are called
  exogenous interventions. The main purpose of this thesis is
  to provide a model based, probabilistic method as an
  approach for known exogenous interventions in mobile
  manipulators. With the thought of the mobile platform and
  manipulator can work in parallel, we proposed two
  variations of continuous observation hidden Markov model,
  namely the classical hidden Markov model and the parallel
  hidden Markov model. We elaborated several use cases for
  exogenous interventions and with these we collected data
  from real and simulated mobile manipulators. Due the
  characteristics of exogenous interventions which are rare,
  random and versatile; we also obtained statistics for
  exogenous interventions to build up the model. We then
  tested the models on a simulated environment. The result
  shows that diagnosing exogenous interventions are precise
  and sensitive. With correct modeling, the hidden Markov
  Model is proven to be able to diagnose the exogenous
  interventions correctly.},
  address  = {Grantham-Allee 20, 53757 St. Augustin, Germany},
  annote  = {SS09 Pl\"oger, von der Hude, K\"ustenmacher supervising},
  author  = {Zinnirah Binti Kasim},
  month = {July},
  school  = {Bonn-Rhein-Sieg University of Applied Sciences},
  title = {HMM-Based Diagnosis of Known Exogenous Interventions in
  Mobile Manipulators},
  year = {2013}
  }

2012

• S. Sharma, “Unified Approach to Motion Planning by Coordinating Mobility and Manipulability for the KUKA youBot,” Master Thesis, Grantham-Allee 20, 53757 St. Augustin, Germany, 2012.
  [BibTeX] [Abstract]
  

  Recent trend in robotics is the shift from using fixed base manipulators to mobile manipulators. Mobile manipulators have become of high interest to industry and service robotics sector because of the increased flexibility and effectiveness they offer due to added mobility. However, the combination of the mobility provided by a mobile platform and the manipulation capabilities provided by a robot arm leads to complex analytical problems for research. These problems can be studied very well on the KUKA youBot, a mobile manipulator designed for education and research applications. This thesis aims to achieve seamless integration and synchronization of mobility and manipulation capabilities for performing mobile manipulation tasks with the KUKA youBot. To do this, we propose a novel approach to perform unified motion planning for the KUKA youBot based on Inverse Kinematics Bi-directional Rapidly Exploring Random Trees (IKBiRRT) algorithm using Workspace Goal Regions. A closed form inverse kinematics solution for the unified kinematics of the KUKA youBot is used to implement the planner and resolve redundancies. A motion planning framework is developed that is capable of updating the elements in the world model in an online manner. Experiments are performed both in simulation and on the robot. To test the coordination between arm and base motions, the mobile manipulator plans a collision free path so that the it goes under a table or bar to simulate a limbo movement. A new approach is described to represent the workspace capabilities of the youBot into a map. This map containing the useful workspace of the mobile manipulator is termed as feasibility map. The full six dimensional continuous workspace of the youBot has been mapped to a reduced subspace in two dimensions without loss of information. The feasibility map describes the reachability and redundancy due to mobility of the youBot, in its workspace. The reachability is represented in a reachability map and is constrained by joint limits and singularities whereas a redundancy map describes the redundancy range at a particular point in the workspace. Applications of the feasibility map in motion planning, grasp planning and online obstacle avoidance scenarios are discussed.

  @MastersThesis{ 2012sharma,
  abstract  = {Recent trend in robotics is the shift from using fixed
  base manipulators to mobile manipulators. Mobile
  manipulators have become of high interest to industry and
  service robotics sector because of the increased
  flexibility and effectiveness they offer due to added
  mobility. However, the combination of the mobility provided
  by a mobile platform and the manipulation capabilities
  provided by a robot arm leads to complex analytical
  problems for research. These problems can be studied very
  well on the KUKA youBot, a mobile manipulator designed for
  education and research applications. This thesis aims to
  achieve seamless integration and synchronization of
  mobility and manipulation capabilities for performing
  mobile manipulation tasks with the KUKA youBot. To do this,
  we propose a novel approach to perform unified motion
  planning for the KUKA youBot based on Inverse Kinematics
  Bi-directional Rapidly Exploring Random Trees (IKBiRRT)
  algorithm using Workspace Goal Regions. A closed form
  inverse kinematics solution for the unified kinematics of
  the KUKA youBot is used to implement the planner and
  resolve redundancies. A motion planning framework is
  developed that is capable of updating the elements in the
  world model in an online manner. Experiments are performed
  both in simulation and on the robot. To test the
  coordination between arm and base motions, the mobile
  manipulator plans a collision free path so that the it goes
  under a table or bar to simulate a limbo movement. A new
  approach is described to represent the workspace
  capabilities of the youBot into a map. This map containing
  the useful workspace of the mobile manipulator is termed as
  feasibility map. The full six dimensional continuous
  workspace of the youBot has been mapped to a reduced
  subspace in two dimensions without loss of information. The
  feasibility map describes the reachability and redundancy
  due to mobility of the youBot, in its workspace. The
  reachability is represented in a reachability map and is
  constrained by joint limits and singularities whereas a
  redundancy map describes the redundancy range at a
  particular point in the workspace. Applications of the
  feasibility map in motion planning, grasp planning and
  online obstacle avoidance scenarios are discussed.},
  address  = {Grantham-Allee 20, 53757 St. Augustin, Germany},
  annote  = {WS 09 HBRS, KUKA Laboratories GmbH - BRICS Kraetzschmar,
  Scheurer},
  author  = {Shashank Sharma},
  month = {September},
  school  = {Bonn-Rhein-Sieg University of Applied Sciences},
  title = {Unified Approach to Motion Planning by Coordinating
  Mobility and Manipulability for the KUKA youBot},
  year = {2012}
  }

• M. Thosar, “A Naive Approach For Learning The Semantics Of Effects Of An Action,” Master Thesis, Grantham-Allee 20, 53757 St. Augustin, Germany, 2012.
  [BibTeX] [Abstract]
  

  Classical artificial intelligence emphasizes that cognitive abilities such as inferring conclusions from the sensor information acquired over time, devising suitable plans, reasoning about the environment etc. are grounded in a mental representation of the world and the resulting intelligent behavior is an outcome of the correct reasoning using these representations. One of the use of these abilities is to reason about effects of an action where an agent reasons how the environment changes after executing an action. In this work, we have proposed an approach to learn the effects of an action. The effects of an action can be learned either by simply stating what changes in the environment given the current state and action description or by stating how these changes are related to the action. The latter part forms our approach. We have developed a system called EVOLT (EVOLving Theories) which for a given set of state descriptions (before and after an action is executed) and a set of action parameters of the action, represented in a predefined format, learns a theory consisting of a formal description of effects of an action in stages such that a theory relates the common changes in the state to the action parameter/s according to the given mathematical/logical operator definitions, expressed in a logic language which is referred as a semantic of the action in this setting. The main contribution of this work is twofold: one, we have proposed an extended Inductive Logic Programming(ILP) technique in a way that the extended ILP does not rely on labeled examples unlike the conventional ILP. Moreover, it is provided with a background knowledge which is general in nature and is not limited to a single action-effect theory learning task, but can be reused for a variety of learning action-effect theory tasks. Second, the representation formalism proposed in this task. The formalism offers a uniform representation for the input data in a sense that the input data is distributed in one or more groups according to some property such that the property wraps systematically the data according to a representation rule. Due to this uniformity, programmer do not have to worry about what and how much data is enlisted in each group while programming. Also, using this representation formalism, variety of inductive learning problems can be aimed. The experimental evaluation is done to examine the applicability and behavior of EVOLT in different environmental settings. The other aim was to gain insights during the experimentation which can further be used to improve the system. As the system is in its early developmental stages, we have discussed the strengths and weaknesses of the system realized during the development and evaluation. The results of the evaluation were quiet reasonable which has formed the basis for the future work we have discussed at the end of this report.

  @MastersThesis{ 2012thosar,
  abstract  = {Classical artificial intelligence emphasizes that
  cognitive abilities such as inferring conclusions from the
  sensor information acquired over time, devising suitable
  plans, reasoning about the environment etc. are grounded in
  a mental representation of the world and the resulting
  intelligent behavior is an outcome of the correct reasoning
  using these representations. One of the use of these
  abilities is to reason about effects of an action where an
  agent reasons how the environment changes after executing
  an action. In this work, we have proposed an approach to
  learn the effects of an action. The effects of an action
  can be learned either by simply stating what changes in the
  environment given the current state and action description
  or by stating how these changes are related to the action.
  The latter part forms our approach. We have developed a
  system called EVOLT (EVOLving Theories) which for a given
  set of state descriptions (before and after an action is
  executed) and a set of action parameters of the action,
  represented in a predefined format, learns a theory
  consisting of a formal description of effects of an action
  in stages such that a theory relates the common changes in
  the state to the action parameter/s according to the given
  mathematical/logical operator definitions, expressed in a
  logic language which is referred as a semantic of the
  action in this setting. The main contribution of this work
  is twofold: one, we have proposed an extended Inductive
  Logic Programming(ILP) technique in a way that the extended
  ILP does not rely on labeled examples unlike the
  conventional ILP. Moreover, it is provided with a
  background knowledge which is general in nature and is not
  limited to a single action-effect theory learning task, but
  can be reused for a variety of learning action-effect
  theory tasks. Second, the representation formalism proposed
  in this task. The formalism offers a uniform representation
  for the input data in a sense that the input data is
  distributed in one or more groups according to some
  property such that the property wraps systematically the
  data according to a representation rule. Due to this
  uniformity, programmer do not have to worry about what and
  how much data is enlisted in each group while programming.
  Also, using this representation formalism, variety of
  inductive learning problems can be aimed. The experimental
  evaluation is done to examine the applicability and
  behavior of EVOLT in different environmental settings. The
  other aim was to gain insights during the experimentation
  which can further be used to improve the system. As the
  system is in its early developmental stages, we have
  discussed the strengths and weaknesses of the system
  realized during the development and evaluation. The results
  of the evaluation were quiet reasonable which has formed
  the basis for the future work we have discussed at the end
  of this report.},
  address  = {Grantham-Allee 20, 53757 St. Augustin, Germany},
  annote  = {[WS08/09] [Kahl], [Mueller], [Ploeger] supervising},
  author  = {Madhura Thosar},
  month = {November},
  school  = {Bonn-Rhein-Sieg University of Applied Sciences},
  title = {A Naive Approach For Learning The Semantics Of Effects Of
  An Action},
  year = {2012}
  }

• M. Kepski and B. Kwolek, “Fall Detection on Embedded Platform Using Kinect and Wireless Accelerometer.” 2012, p. 407–414.
  [BibTeX]

  @InProceedings{ kepski2012,
  author  = {Kepski, Michal and Kwolek, Bogdan},
  journal  = {ICCHP},
  title = {{Fall Detection on Embedded Platform Using Kinect and
  Wireless Accelerometer}},
  year = {2012},
  issn = {01692607},
  pages = {407--414}
  }

• K. Soomro, A. Roshan Zamir, and M. Shah, “UCF101: A Dataset of 101 Human Actions Classes From Videos in The Wild,” CoRR, 2012.
  [BibTeX]

  @Article{ khurram2012,
  author  = {Soomro, Khurram and Roshan Zamir, Amir and Shah, Mubarak},
  title = {{{UCF101:} {A} Dataset of 101 Human Actions Classes From
  Videos in The Wild}},
  journal  = {CoRR},
  year = {2012}
  }

• T. Mathew, “A Computer Game based Motivation System for Human Muscle Strength Testing,” Master Thesis, Grantham-Allee 20, 53757 St. Augustin, Germany, 2012.
  [BibTeX] [Abstract]
  

  The objective of this thesis is to implement a computer game based motivation system for maximal strength testing on the Biodex System 3 Isokinetic Dynamometer. The prototype game has been designed to improve the peak torque produced in an isometric knee extensor strength test. An extensive analysis is performed on a torque data set from a previous study. The torque responses for five second long maximal voluny contractions of the knee extensor are analyzed to understand torque response characteristics of di fferent subjects. The parameters identified in the data analysis are used in the implementation of the ‘Shark and School of Fish’ game. The behavior of the game for different torque responses is analyzed on a different torque data set from the previous study. The evaluationshows that the game rewards and motivates continuously over a repetition to reach the peak torque value. The evaluation also shows that the game rewards the user more if he overcomes a baseline torque value within the first second and then gradually increase the torque to reach peak torque.

  @MastersThesis{ 2012mathew,
  abstract  = {The objective of this thesis is to implement a computer
  game based motivation system for maximal strength testing
  on the Biodex System 3 Isokinetic Dynamometer. The
  prototype game has been designed to improve the peak torque
  produced in an isometric knee extensor strength test. An
  extensive analysis is performed on a torque data set from a
  previous study. The torque responses for five second long
  maximal voluny contractions of the knee extensor are
  analyzed to understand torque response characteristics of
  di fferent subjects. The parameters identified in the data
  analysis are used in the implementation of the 'Shark and
  School of Fish' game. The behavior of the game for
  different torque responses is analyzed on a different
  torque data set from the previous study. The
  evaluationshows that the game rewards and motivates
  continuously over a repetition to reach the peak torque
  value. The evaluation also shows that the game rewards the
  user more if he overcomes a baseline torque value within
  the first second and then gradually increase the torque to
  reach peak torque.},
  address  = {Grantham-Allee 20, 53757 St. Augustin, Germany},
  annote  = {[Semester you sted the MAS Program] [Institute of
  Aerospace Medicine,German Aerospace Center] - [A Computer
  Game based Motivation System for Human Muscle Strength
  Testing] [Herpers], [Rittweger] supervising},
  author  = {Tintu Mathew},
  month = {},
  school  = {Bonn-Rhein-Sieg University of Applied Sciences},
  title = {A Computer Game based Motivation System for Human Muscle
  Strength Testing},
  year = {2012}
  }

• M. Füller, “Multi-step motion planning of climbing robots in 3D environments under kinodynamic constraints,” Master Thesis, Grantham-Allee 20, 53757 St. Augustin, Germany, 2012.
  [BibTeX] [Abstract]
  

  New locomotion techniques for service robots aim to overcome the difficulties with obstacle avoidance and dynamic settings in household environments by changing the robot’s working plane to the wall and ceiling. Besides the question of the mechanical implementation of these kind of robots, the motion planning of the robot requires another view and special attention. A spider-type robot is assumed in this thesis that is able to move along given footholds in the environment. The motion planning task is to determine the required steps to move to the goal along available docking points at the walls and the ceiling. A previous study of the author developed a multi-step motion planner for such kind of robot. The planner is able to plan multi-step motions while being aware of kinematic constraints and object collisions. One aspect that was identified during that work but not implemented is the awareness of dynamic constraints of such a robot. The motion of a legged robot – especially of climbing ones – requires special attention to the dynamic limitations. These dynamic limitations are for example maximum possible joint torques during the climbing motions. This work is the next step toward a more realistic multi-step motion planner for 3D indoor environments along walls and ceilings. It extends the existing kinematic multi-step motion planner presented by Füller (2011) with additional awareness of dynamic constraints. The planner is further integrated into a real-world physic simulation environment. The final multi-step planner can be used to evaluate required components for a real implementation of such kind of robot. It can identify problems in the planning process due to limits on the robot design and the used components before and during the development of such a robot. Additionally, the thesis presents a possible approach for an extension of a pure kinematic sample-based motion planner toward dynamic awareness.

  @MastersThesis{ 2012fueller,
  author  = {Matthias F{\"u}ller},
  title = {Multi-step motion planning of climbing robots in 3D
  environments under kinodynamic constraints},
  school  = {Bonn-Rhein-Sieg University of Applied Sciences},
  year = {2012},
  address  = {Grantham-Allee 20, 53757 St. Augustin, Germany},
  month = {March},
  abstract  = {New locomotion techniques for service robots aim to
  overcome the difficulties with obstacle avoidance and
  dynamic settings in household environments by changing the
  robot's working plane to the wall and ceiling. Besides the
  question of the mechanical implementation of these kind of
  robots, the motion planning of the robot requires another
  view and special attention. A spider-type robot is assumed
  in this thesis that is able to move along given footholds
  in the environment. The motion planning task is to
  determine the required steps to move to the goal along
  available docking points at the walls and the ceiling.
  A previous study of the author developed a multi-step
  motion planner for such kind of robot. The planner is able
  to plan multi-step motions while being aware of kinematic
  constraints and object collisions. One aspect that was
  identified during that work but not implemented is the
  awareness of dynamic constraints of such a robot. The
  motion of a legged robot - especially of climbing ones -
  requires special attention to the dynamic limitations.
  These dynamic limitations are for example maximum possible
  joint torques during the climbing motions. This work is the
  next step toward a more realistic multi-step motion planner
  for 3D indoor environments along walls and ceilings. It
  extends the existing kinematic multi-step motion planner
  presented by F{\"u}ller (2011) with additional awareness of
  dynamic constraints. The planner is further integrated into
  a real-world physic simulation environment. The final
  multi-step planner can be used to evaluate required
  components for a real implementation of such kind of robot.
  It can identify problems in the planning process due to
  limits on the robot design and the used components before
  and during the development of such a robot. Additionally,
  the thesis presents a possible approach for an extension of
  a pure kinematic sample-based motion planner toward dynamic
  awareness.},
  annote  = {W09, Prassler, Forsman supervising}
  }

• M. Arasi, “Particle Filter Based Approach for the Diagnosis of Unknown External Faults in a Mobile Manipulator,” Master Thesis, Grantham-Allee 20, 53757 St. Augustin, Germany, 2012.
  [BibTeX] [Abstract]
  

  Monitoring complex systems (e.g. Mobile Manipulator) for the sake of estimation and fault diagnosis, is a topic of considerable dominance in scientific literature. Autonomous systems often stop performing their tasks because of occurrences of unexpected faults. Here, unexpected or unknown fault is defined as the fault that occurs in the system’s dynamic environment and cannot be observed by the system’s sensors because of their limitations. Fault diagnosis is a prerequisite for robust operation of a system in a hazardous or changing environment. Fault diagnosis comprises of different procedures represented by fault detection, isolation and fault identification. In recent years, many researchers have investigated diagnosis problems, and many of them proposed particle filter (PF) as a solution. Particle filter is a Bayesian approach used in many applications such as fault diagnosis to approximate the belief distribution of the state of the system as soon as the observation is available. It has successfully implemented on a complete system model that is running under internal faults. In the recent works, for the systems dealing with unknown faults, researchers assume that the approach (i.e. Particle Filter) is working on un-modeled or imperfectly modeled system. This assumption leads only to the detection of the unknown faults but cannot identify the main cause of the failure. However, this information is not enough, especially for the recovery procedure. Since it is important and useful for the fault diagnosis approach to incorporate the disturbances and unexpected faults that occur in a system, this is the main issue handled in the thesis. In brief, developing an efficient fault diagnosis approach for the unknown external faults is suggested to be the solution for the aforementioned problems. In addition, robustness, correctness and efficiency of the proposed approach has been investigated, in diagnosing the state of the system. Our fault diagnosis approach is based on Particle Filter (PF). As an evidence of working of the approach, we use a set of test case scenarios for a simulated mobile manipulator, in openRAVE [22]. On the same set of test case scenarios, we applied the Classical Particle Filter (CPF) and the Gaussian Particle Filter (GPF). The approach is applied on a hybrid dynamic system in the application of reliable navigation for a mobile robot and efficient object transfer from the initial to the target position for a robot manipulator. In our approach, we have assumed that during the occurrences of unknown fault the system’s hardware and software keeps functioning perfectly. Extensive simulations have been carried out to test the performance of the approach under different number of particles. The experimental results show the effectiveness of the approach on the given set of use case scenarios. The results show that both the approaches, CPF and GPF, are equally good at diagnosing unknown external faults but the GPF functions a lot better for the diagnosis of the unknown external faults even under less number of particles. The proposed diagnosis approach is able not only to diagnose the fault, but also to estimate some valuable informations needed for the recovery process, such as collision position for a navigated mobile robot.

  @MastersThesis{ 2012arasi,
  abstract  = {Monitoring complex systems (e.g. Mobile Manipulator) for
  the sake of estimation and fault diagnosis, is a topic of
  considerable dominance in scientific literature. Autonomous
  systems often stop performing their tasks because of
  occurrences of unexpected faults. Here, unexpected or
  unknown fault is defined as the fault that occurs in the
  system's dynamic environment and cannot be observed by the
  system's sensors because of their limitations. Fault
  diagnosis is a prerequisite for robust operation of a
  system in a hazardous or changing environment. Fault
  diagnosis comprises of different procedures represented by
  fault detection, isolation and fault identification. In
  recent years, many researchers have investigated diagnosis
  problems, and many of them proposed particle filter (PF) as
  a solution. Particle filter is a Bayesian approach used in
  many applications such as fault diagnosis to approximate
  the belief distribution of the state of the system as soon
  as the observation is available. It has successfully
  implemented on a complete system model that is running
  under internal faults. In the recent works, for the systems
  dealing with unknown faults, researchers assume that the
  approach (i.e. Particle Filter) is working on un-modeled or
  imperfectly modeled system. This assumption leads only to
  the detection of the unknown faults but cannot identify the
  main cause of the failure. However, this information is not
  enough, especially for the recovery procedure. Since it is
  important and useful for the fault diagnosis approach to
  incorporate the disturbances and unexpected faults that
  occur in a system, this is the main issue handled in the
  thesis. In brief, developing an efficient fault diagnosis
  approach for the unknown external faults is suggested to be
  the solution for the aforementioned problems. In addition,
  robustness, correctness and efficiency of the proposed
  approach has been investigated, in diagnosing the state of
  the system. Our fault diagnosis approach is based on
  Particle Filter (PF). As an evidence of working of the
  approach, we use a set of test case scenarios for a
  simulated mobile manipulator, in openRAVE [22]. On the same
  set of test case scenarios, we applied the Classical
  Particle Filter (CPF) and the Gaussian Particle Filter
  (GPF). The approach is applied on a hybrid dynamic system
  in the application of reliable navigation for a mobile
  robot and efficient object transfer from the initial to the
  target position for a robot manipulator. In our approach,
  we have assumed that during the occurrences of unknown
  fault the system's hardware and software keeps functioning
  perfectly. Extensive simulations have been carried out to
  test the performance of the approach under different number
  of particles. The experimental results show the
  effectiveness of the approach on the given set of use case
  scenarios. The results show that both the approaches, CPF
  and GPF, are equally good at diagnosing unknown external
  faults but the GPF functions a lot better for the diagnosis
  of the unknown external faults even under less number of
  particles. The proposed diagnosis approach is able not only
  to diagnose the fault, but also to estimate some valuable
  informations needed for the recovery process, such as
  collision position for a navigated mobile robot.},
  address  = {Grantham-Allee 20, 53757 St. Augustin, Germany},
  annote  = {WS08/09 FH-BRS - Particle Filter Based Approach for the
  Diagnosis of Unknown External Faults in a Mobile
  Manipulator Pl{\"o}ger, M{\"u}ller, K{\"u}stenmacher
  supervising},
  author  = {Musherah Arasi},
  month = {October},
  school  = {Bonn-Rhein-Sieg University of Applied Sciences},
  title = {Particle Filter Based Approach for the Diagnosis of
  Unknown External Faults in a Mobile Manipulator},
  year = {2012}
  }

• P. Banerjee, “Design and Implementation of a Software Development Toolkit to provide Perception Functionalities for Robots,” Master Thesis, Grantham-Allee 20, 53757 St. Augustin, Germany, 2012.
  [BibTeX] [Abstract]
  

  Robotics is a vast, interdisciplinary field of study that demands expertise of multiple disciplines like com- puter science, mathematics, electronics, mechanics, etc. Leaving the hardware part aside, even robotic software development involves integration of multiple processing components focused on different chal- lenges like, control, manipulation, navigation, perception, planning, etc. In recent development scenarios, a group of researchers usually focus on different sub-domains of expertise individually and then the sub- domain applications are integrated into complete robotic application software. To ease the software development process and reduce requirement of comprehensive knowledge of multiple sub-domains, Software Development Toolkits (SDKs) can be used. The SDKs will provide a set of high-level Application Programming Interfaces (APIs) for different robotic-domains. The APIs can be used to trigger predefined processing steps. To use the APIs only working knowledge of the corre- sponding functionalities will be required. This will enable development of a complete robotic application with comprehensive knowledge of certain sub-domains and abstract working information for the rest. For example, a planning domain expert can implement the software logics for task planning and simply reuse the APIs from perception and manipulation domain to trigger the perception and manipulation processes as governed by the implemented planning module. However, in the current state of the art of open-source robotic softwares, Software Development Toolkits for robotic manipulation, control and perception are very limited. This work aims to develop a SDK in the field of 3D robotic perception that should be able to provide dedicated interfaces to trigger high level perception functionalities like ”object detection”, ”6D pose estimation” and so on. The interface will be designed considering two classes of users for the SDK, domain-users 1 and domain-experts 2 . Based on available domain knowledge the developers can use high-level APIs to parametrize and trigger a processing step or low-level APIs to re-configure the software module and setup the algorithms and parameters to be used for processing.

  @MastersThesis{ 2012banerjee,
  abstract  = {Robotics is a vast, interdisciplinary field of study that
  demands expertise of multiple disciplines like com- puter
  science, mathematics, electronics, mechanics, etc. Leaving
  the hardware part aside, even robotic software development
  involves integration of multiple processing components
  focused on different chal- lenges like, control,
  manipulation, navigation, perception, planning, etc. In
  recent development scenarios, a group of researchers
  usually focus on different sub-domains of expertise
  individually and then the sub- domain applications are
  integrated into complete robotic application software. To
  ease the software development process and reduce
  requirement of comprehensive knowledge of multiple
  sub-domains, Software Development Toolkits (SDKs) can be
  used. The SDKs will provide a set of high-level Application
  Programming Interfaces (APIs) for different
  robotic-domains. The APIs can be used to trigger predefined
  processing steps. To use the APIs only working knowledge of
  the corre- sponding functionalities will be required. This
  will enable development of a complete robotic application
  with comprehensive knowledge of certain sub-domains and
  abstract working information for the rest. For example, a
  planning domain expert can implement the software logics
  for task planning and simply reuse the APIs from perception
  and manipulation domain to trigger the perception and
  manipulation processes as governed by the implemented
  planning module. However, in the current state of the art
  of open-source robotic softwares, Software Development
  Toolkits for robotic manipulation, control and perception
  are very limited. This work aims to develop a SDK in the
  field of 3D robotic perception that should be able to
  provide dedicated interfaces to trigger high level
  perception functionalities like ”object detection”,
  ”6D pose estimation” and so on. The interface will be
  designed considering two classes of users for the SDK,
  domain-users 1 and domain-experts 2 . Based on available
  domain knowledge the developers can use high-level APIs to
  parametrize and trigger a processing step or low-level APIs
  to re-configure the software module and setup the
  algorithms and parameters to be used for processing. },
  address  = {Grantham-Allee 20, 53757 St. Augustin, Germany},
  annote  = {WS-09,GPS EU-BRICS, Prassler, Blumenthal, Zakharov},
  author  = {Pinaki Banerjee},
  month = {February},
  year = {2012},
  school  = {Bonn-Rhein-Sieg University of Applied Sciences},
  title = {Design and Implementation of a Software Development
  Toolkit to provide Perception Functionalities for Robots}
  }

• E. Dayangac, “Vision-based 6DoF Input Device Development with Ground Truth Evaluation,” Master Thesis, Grantham-Allee 20, 53757 St. Augustin, Germany, 2012.
  [BibTeX] [Abstract]
  

  Different visual tracking approaches exists, among them feature-based and template- based are prominent, for determining position and orientation of an object or camera. Visual tracking has been used for several applications ranging from visual odometry, augmented reality, 3D object modeling, medical imaging, visual simultaneous localization and mapping to surveillance. The work presented in this thesis designs a visual tracking system to recover the full 6 degree-of-freedom pose of a camera with developing an evaluation methodology for performance analysis of the system. The work presented in this thesis includes design of a visual tracking system to recover the full 6 degree-of-freedom pose of a camera and also development of an evaluation methodology to analyze the performance of the tracking system. Marker-based tracking and visualization system is implemented for CAVE-type virtual reality environments. The system consists of a wired CCD camera as the only sensor and multiple unique markers. The markers are placed into a video picture and projected within a scene using a projector Except the calibration of projector and camera, the system does not need additional calibration between the scene and the tracking system. The system is affordable comparing to other existing tracking systems. The drawbacks of the system are the high power consumption and cabling requirement due to the CCD camera and the visibility of the markers within the scene. However, rapid development in imaging hardware makes the system affordable, and useful in the near future. The LCD projectors are used, hence the visualized marker within the scene is in the visible range of human eye under the current implementation. The system still needs a solution of a light module at invisible wavelength. Camera pose estimation is very common in the field of Computer Vision, which makes the implementation of a tracking system a trivial task mentioned above except; performance analysis is a challenging task. The extensive evaluation of a tracking algorithm is still an open issue in research, even though there are some attempts to address the problem. In this thesis, we categorized evaluation methods on visual tracking systems. We can conclude that the existing approaches solves the problem partially, and insufficient to discover weaknesses and strengths of an algorithm. Knowledge from the experiments does not clearly show the failures of the algorithms to improve the results, because their main purpose was to compare some particular approaches. Also, an evaluation methodology has been proposed and a framework is implemented. The framework includes the camera and visualization system additionally with a 5DoF computer-controlled robotic arm and an optical tracking system. The arm is used to move the camera systematically, while optical tracking system is used to get a rough idea about the pose of the robot-base. A new system and distributed software architectures are presented in the report. This framework is capable of collecting static image sequences repeatedly around a 3D point-list. Featuring different texture images can be collected in a short time. During run-time, it does on-line image processing with stitching reference data. The data analysis part is done manually afterwards, which is one of the incapabilities of the system. Due to the deficient robotic arm, the motion pattern is limited and collected images are static. Consequently, the implemented marker based tracking system and the optical tracking system are evaluated partially.

  @MastersThesis{ 2012dayangac,
  abstract  = {Different visual tracking approaches exists, among them
  feature-based and template- based are prominent, for
  determining position and orientation of an object or
  camera. Visual tracking has been used for several
  applications ranging from visual odometry, augmented
  reality, 3D object modeling, medical imaging, visual
  simultaneous localization and mapping to surveillance.
  The work presented in this thesis designs a visual tracking
  system to recover the full 6 degree-of-freedom pose of a
  camera with developing an evaluation methodology for
  performance analysis of the system. The work presented in
  this thesis includes design of a visual tracking system to
  recover the full 6 degree-of-freedom pose of a camera and
  also development of an evaluation methodology to analyze
  the performance of the tracking system.
  Marker-based tracking and visualization system is
  implemented for CAVE-type virtual reality environments. The
  system consists of a wired CCD camera as the only sensor
  and multiple unique markers. The markers are placed into a
  video picture and projected within a scene using a projector
  Except the calibration of projector and camera, the system
  does not need additional calibration between the scene and
  the tracking system. The system is affordable comparing to
  other existing tracking systems. The drawbacks of the
  system are the high power consumption and cabling
  requirement due to the CCD camera and the visibility of the
  markers within the scene. However, rapid development in
  imaging hardware makes the system affordable, and useful in
  the near future. The LCD projectors are used, hence the
  visualized marker within the scene is in the visible range
  of human eye under the current implementation. The system
  still needs a solution of a light module at invisible
  wavelength.
  Camera pose estimation is very common in the field of
  Computer Vision, which makes the implementation of a
  tracking system a trivial task mentioned above except;
  performance analysis is a challenging task.
  The extensive evaluation of a tracking algorithm is still
  an open issue in research, even though there are some
  attempts to address the problem. In this thesis, we
  categorized evaluation methods on visual tracking systems.
  We can conclude that the existing approaches solves the
  problem partially, and insufficient to discover weaknesses
  and strengths of an algorithm. Knowledge from the
  experiments does not clearly show the failures of the
  algorithms to improve the results, because their main
  purpose was to compare some particular approaches.
  Also, an evaluation methodology has been proposed and a
  framework is implemented. The framework includes the camera
  and visualization system additionally with a 5DoF
  computer-controlled robotic arm and an optical tracking
  system. The arm is used to move the camera systematically,
  while optical tracking system is used to get a rough idea
  about the pose of the robot-base. A new system and
  distributed software architectures are presented in the
  report.
  This framework is capable of collecting static image
  sequences repeatedly around a 3D point-list. Featuring
  different texture images can be collected in a short time.
  During run-time, it does on-line image processing with
  stitching reference data. The data analysis part is done
  manually afterwards, which is one of the incapabilities of
  the system. Due to the deficient robotic arm, the motion
  pattern is limited and collected images are static.
  Consequently, the implemented marker based tracking system
  and the optical tracking system are evaluated partially.},
  address  = {Grantham-Allee 20, 53757 St. Augustin, Germany},
  annote  = {SS08/09,Herpers, Hinkenjann, Saitov supervising},
  author  = {Enes Dayangac},
  month = {April},
  school  = {Bonn-Rhein-Sieg University of Applied Sciences},
  title = {Vision-based 6DoF Input Device Development with Ground
  Truth Evaluation},
  year = {2012}
  }

• R. Dwiputra, “Dynamic Modeling of KUKA youBot Manipulator,” Master Thesis, Grantham-Allee 20, 53757 St. Augustin, Germany, 2012.
  [BibTeX] [Abstract]
  

  Models and simulation tools are crucial in robotic research. Experimentation with models is cost and time efficient due to its flexibility to be automated, conditioned, and accelerated. In this master thesis, a model of the KUKA youBot manipulator was developed with Modelica. Modelica is a multi-domain modeling language description which is capable of bridging mechanical, electrical, hydraulic, and thermodynamic domain in a single model. With Modelica, the model developed incorporated dynamic parameters, motor specifications, and control system. The advantages of a robot model only holds true when its accuracy has been validated with the real robot. Before further research involving the model is executed, it should be ensured that the model behaviours reflect the actual system behaviours. Therefore, experiments have been performed in parallel with the model development to assure the model’s accuracy. The experiments performed in this master thesis are: 1. Validation of the robot controller model; 2. Friction model approximation; and 3. Validation of the robot model for different motions and modes. The results show that the model reflects the actual system behaviour to a certain extent. The experiment results and the model can be further used for experiment involving control system, load identification, and trajectory generation algorithm.

  @MastersThesis{ 2012dwiputra,
  author  = {Rhama Dwiputra},
  title = {Dynamic Modeling of KUKA youBot Manipulator},
  year = {2012},
  month = {October},
  abstract  = {Models and simulation tools are crucial in robotic
  research. Experimentation with models is cost and time
  efficient due to its flexibility to be automated,
  conditioned, and accelerated. In this master thesis, a
  model of the KUKA youBot manipulator was developed with
  Modelica. Modelica is a multi-domain modeling language
  description which is capable of bridging mechanical,
  electrical, hydraulic, and thermodynamic domain in a single
  model. With Modelica, the model developed incorporated
  dynamic parameters, motor specifications, and control
  system.
  The advantages of a robot model only holds true when its
  accuracy has been validated with the real robot. Before
  further research involving the model is executed, it should
  be ensured that the model behaviours reflect the actual
  system behaviours. Therefore, experiments have been
  performed in parallel with the model development to assure
  the model's accuracy. The experiments performed in this
  master thesis are: 1. Validation of the robot controller
  model; 2. Friction model approximation; and 3. Validation
  of the robot model for different motions and modes. The
  results show that the model reflects the actual system
  behaviour to a certain extent. The experiment results and
  the model can be further used for experiment involving
  control system, load identification, and trajectory
  generation algorithm.},
  address  = {Grantham-Allee 20, 53757 St. Augustin, Germany},
  annote  = {[Semester you started the MAS Program] [Project
  Affiliation] - [Project Name] [Last name of 1st
  Supervisor], [Last name of 2nd supervisor], [last name of
  third supervisor (if applicable)] supervising},
  school  = {Bonn-Rhein-Sieg University of Applied Sciences}
  }

• N. Akhtar, “Improving reliability of mobile manipulators against unknown external faults,” Master Thesis, Grantham-Allee 20, 53757 St. Augustin, Germany, 2012.
  [BibTeX] [Abstract]
  

  A robot (e.g. mobile manipulator) that interacts with its environment to perform its tasks, often faces situations in which it is unable to achieve its goals despite perfect functioning of its sensors and actuators. These situations occur when the behavior of the objects manipulated by the robot deviates from its expected course because of unforeseeable circumstances. These deviations are experienced by the robot as unknown external faults which result in failures of its actions. In this work we present an approach that increases the reliability of mobile manipulators against the unknown external faults. This approach focuses on the actions of manipulators which involve releasing of an object. The approach presented in this work is formulated as a three step scheme that takes an example simulation and a definition of a planning operator as its inputs. The example simulation is a simulation that shows the expected/desired behavior of the object upon the execution of the action corresponded by the planning operator. In its first step, the scheme finds a description of the behavior of the objects in the example simulation in terms of logical atoms. We refer to these logical atoms collectively as description vocabulary. The description of the simulation is used by the second step to find limits of the parameters of the manipulated object. Using randomly chosen values of the parameters within these limits, this step creates different examples of the releasing state of the object. These examples are labelled as desired or undesired according to the behavior of the object in the simulation. The description vocabulary is also used in labeling the examples. In the third step, an algorithm (i.e. N-Bins) uses the labelled examples to suggest the state for the object in which releasing it avoids the occurrence of any unknown external faults. The proposed N-Bins algorithm can also be used for binary classification problem. Therefore, in our experiments with the proposed approach we also test its prediction ability along with the analysis of the results of our approach. The results show that under the circumstances peculiar to our approach, N-Bins algorithm show reasonable prediction accuracy where other state of the art classification algorithms fail to do so. Thus, N-Bins also extends the ability of a robot to predict the behavior of the object to avoid unknown external faults. In this work we use simulation environment OPENRave that uses physics engine ODE to simulate the dynamics of rigid bodies.

  @MastersThesis{ 2012akhtar,
  address  = {Grantham-Allee 20, 53757 St. Augustin, Germany},
  author  = {Naveed Akhtar},
  date-modified  = {2012-02-18 17:07:45 +0100},
  keywords  = {external faults, mobile manipulators, binary
  classification},
  month = {February},
  school  = {Bonn-Rhein-Sieg University of Applied Sciences},
  title = {Improving reliability of mobile manipulators against
  unknown external faults},
  year = {2012},
  abstract  = {A robot (e.g. mobile manipulator) that interacts with its
  environment to perform its tasks, often faces situations in
  which it is unable to achieve its goals despite perfect
  functioning of its sensors and actuators. These situations
  occur when the behavior of the objects manipulated by the
  robot deviates from its expected course because of
  unforeseeable circumstances. These deviations are
  experienced by the robot as unknown external faults which
  result in failures of its actions. In this work we present
  an approach that increases the reliability of mobile
  manipulators against the unknown external faults. This
  approach focuses on the actions of manipulators which
  involve releasing of an object. The approach presented in
  this work is formulated as a three step scheme that takes
  an example simulation and a definition of a planning
  operator as its inputs. The example simulation is a
  simulation that shows the expected/desired behavior of the
  object upon the execution of the action corresponded by the
  planning operator. In its first step, the scheme finds a
  description of the behavior of the objects in the example
  simulation in terms of logical atoms. We refer to these
  logical atoms collectively as description vocabulary. The
  description of the simulation is used by the second step to
  find limits of the parameters of the manipulated object.
  Using randomly chosen values of the parameters within these
  limits, this step creates different examples of the
  releasing state of the object. These examples are labelled
  as desired or undesired according to the behavior of the
  object in the simulation. The description vocabulary is
  also used in labeling the examples. In the third step, an
  algorithm (i.e. N-Bins) uses the labelled examples to
  suggest the state for the object in which releasing it
  avoids the occurrence of any unknown external faults.
  The proposed N-Bins algorithm can also be used for binary
  classification problem. Therefore, in our experiments with
  the proposed approach we also test its prediction ability
  along with the analysis of the results of our approach. The
  results show that under the circumstances peculiar to our
  approach, N-Bins algorithm show reasonable prediction
  accuracy where other state of the art classification
  algorithms fail to do so. Thus, N-Bins also extends the
  ability of a robot to predict the behavior of the object to
  avoid unknown external faults. In this work we use
  simulation environment OPENRave that uses physics engine
  ODE to simulate the dynamics of rigid bodies.
  },
  annote  = {WS 09/10, H-BRS, Pl{\"o}ger, Asteroth, Kuestenmacher
  supervising }
  }

2011

• M. Shahzad, “Detection and tracking of pointing hand gestures for AR applications,” Master Thesis, Grantham-Allee 20, 53757 St. Augustin, Germany, 2011.
  [BibTeX] [Abstract]
  

  In the field of augmented reality (AR) user interfaces and service robotics, directly using user�s hand provide natural way to interact with the computers. Apart from other gestures, pointing gestures are one of the most commonly used gestures in everyday life. They can enable natural and convenient way for humans to interact with computers. In this thesis, a real time approach for detecting such natural pointing hand gestures for user interaction is presented. It uses data provided by a stereoscopic camera system capable of giving dense disparity maps and an estimation of 3D point cloud information. The developed approach allows the user to conveniently interact using natural pointing hand gestures with different components/modules of the biological laboratory Biolab. The Biolab is an International Standard Payload Rack (ISPR) of European Space Agency (ESA) and is operated by German Aerospace Center (Deutsches Zentrum f�r Luft- und Raumfahrt e.V. – DLR) during space missions. It is an integral part of the European science laboratory, Columbus, which is part of the International Space Station (ISS). The approach uses skin color and 3D data to extract probable hand regions in an image. These regions are further refined to detect the pointing hand contour using minimum distance approach by projecting the 3D contour data onto virtual plane representing the Biolab. Feature points (hand center and fingertip) are found in the detected contour to estimate the pointing direction and a 3D virtual beam is fitted between them. The fingertip projection is used to switch between two semantically modeled Biolab representations in two different levels. The virtual plane and estimated 3D beam is used in the identification of pointing targets in both the levels. The developed approach is able to work in real time without any markers or other sensors attached to the pointing hand. It is person independent and has the potential to cope with different skin colors and complex background and does not require any manual initialization procedure. Moreover, it does not put any constrain on the user to wear special clothing and detects pointing hand and identifies the pointed targets even in cases if the user is not wearing full sleeves shirt. The algorithm is thoroughly evaluated with 8 different persons under two different lighting conditions in sitting and standing postures. The pointing targets are correctly identified as can be verified by inspection for natural pointing hand gestures performed by different test subjects. Identification results presented later in the thesis illustrates the effectiveness of the approach.

  @MastersThesis{ 2011shahzad,
  abstract  = {In the field of augmented reality (AR) user interfaces and
  service robotics, directly using user�s hand provide
  natural way to interact with the computers. Apart from
  other gestures, pointing gestures are one of the most
  commonly used gestures in everyday life. They can enable
  natural and convenient way for humans to interact with
  computers. In this thesis, a real time approach for
  detecting such natural pointing hand gestures for user
  interaction is presented. It uses data provided by a
  stereoscopic camera system capable of giving dense
  disparity maps and an estimation of 3D point cloud
  information. The developed approach allows the user to
  conveniently interact using natural pointing hand gestures
  with different components/modules of the biological
  laboratory Biolab. The Biolab is an International Standard
  Payload Rack (ISPR) of European Space Agency (ESA) and is
  operated by German Aerospace Center (Deutsches Zentrum f�r
  Luft- und Raumfahrt e.V. - DLR) during space missions. It
  is an integral part of the European science laboratory,
  Columbus, which is part of the International Space Station
  (ISS). The approach uses skin color and 3D data to extract
  probable hand regions in an image. These regions are
  further refined to detect the pointing hand contour using
  minimum distance approach by projecting the 3D contour data
  onto virtual plane representing the Biolab. Feature points
  (hand center and fingertip) are found in the detected
  contour to estimate the pointing direction and a 3D virtual
  beam is fitted between them. The fingertip projection is
  used to switch between two semantically modeled Biolab
  representations in two different levels. The virtual plane
  and estimated 3D beam is used in the identification of
  pointing targets in both the levels. The developed approach
  is able to work in real time without any markers or other
  sensors attached to the pointing hand. It is person
  independent and has the potential to cope with different
  skin colors and complex background and does not require any
  manual initialization procedure. Moreover, it does not put
  any constrain on the user to wear special clothing and
  detects pointing hand and identifies the pointed targets
  even in cases if the user is not wearing full sleeves
  shirt. The algorithm is thoroughly evaluated with 8
  different persons under two different lighting conditions
  in sitting and standing postures. The pointing targets are
  correctly identified as can be verified by inspection for
  natural pointing hand gestures performed by different test
  subjects. Identification results presented later in the
  thesis illustrates the effectiveness of the approach. },
  address  = {Grantham-Allee 20, 53757 St. Augustin, Germany},
  annote  = {ws08 DLR - Detection and tracking of pointing hand
  gestures for AR applications Herpers, Pl�ger, Mittag},
  author  = {Muhammad Shahzad},
  month = {March},
  school  = {Bonn-Rhein-Sieg University of Applied Sciences},
  title = {Detection and tracking of pointing hand gestures for AR
  applications},
  year = {2011}
  }

• F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, J. Vanderplas, A. Passos, D. Cournapeau, M. Brucher, M. Perrot, and É. Duchesnay, “Scikit-learn: Machine Learning in Python,” Journal of Machine Learning Research, vol. 12, p. 2825 – 2830, 2011.
  [BibTeX]

  @Article{ fabian2011,
  author  = {Pedregosa, Fabian and Varoquaux, Ga{\"el} and Gramfort,
  Alexandre and Michel, Vincent and Thirion, Bertrand and
  Grisel, Olivier and Blondel, Mathieu and Prettenhofer,
  Peter and Weiss, Ron and Dubourg, Vincent and Vanderplas,
  Jake and Passos, Alexandre and Cournapeau, David and
  Brucher, Matthieu and Perrot, Matthieu and Duchesnay,
  {\'E}douard},
  journal  = {Journal of Machine Learning Research},
  title = {{Scikit-learn: Machine Learning in Python}},
  year = {2011},
  pages = {2825 -- 2830},
  volume  = {12}
  }

• H. Kuehne, H. Jhuang, E. Garrote, T. Poggio, and T. Serre, “HMDB: a large video database for human motion recognition,” in Proceedings of the International Conference on Computer Vision (ICCV), 2011.
  [BibTeX]

  @InProceedings{ kuehne2011,
  author  = {Kuehne, H. and Jhuang, H. and Garrote, E. and Poggio, T.
  and Serre, T.},
  title = {{{HMDB}: a large video database for human motion
  recognition}},
  booktitle  = {Proceedings of the International Conference on Computer
  Vision (ICCV)},
  year = {2011}
  }

• C. A. Mueller, “3D Object Shape Categorization in Domestic Environments,” Master Thesis, Grantham-Allee 20, 53757 St. Augustin, Germany, 2011.
  [BibTeX] [Abstract]
  

  In service robotics, tasks without the involvement of objects are barely applicable, like in searching, fetching or delivering tasks. Service robots are supposed to capture efficiently object related information in real world scenes while for instance considering clutter and noise, and also being flexible and scalable to memorize a large set of objects. Besides object perception tasks like object recognition where the object’s identity is analyzed, object categorization is an important visual object perception cue that associates unknown object instances based on their e.g. appearance or shape to a corresponding category. We present a pipeline from the detection of object candidates in a domestic scene over the description to the final shape categorization of detected candidates. In order to detect object related information in cluttered domestic environments an object detection method is proposed that copes with multiple plane and object occurrences like in cluttered scenes with shelves. Further a surface reconstruction method based on Growing Neural Gas (GNG) in combination with a shape distribution-based descriptor is proposed to reflect shape characteristics of object candidates. Beneficial properties provided by the GNG such as smoothing and denoising effects support a stable description of the object candidates which also leads towards a more stable learning of categories. Based on the presented descriptor a dictionary approach combined with a supervised shape learner is presented to learn prediction models of shape categories. Experimental results, of different shapes related to domestically appearing object shape categories such as cup, can, box, bottle, bowl, plate and ball, are shown. A classifica- tion accuracy of about 90% and a sequential execution time of lesser than two seconds for the categorization of an unknown object is achieved which proves the reasonableness of the proposed system design. Additional results are shown towards object tracking and false positive handling to enhance the robustness of the categorization. Also an initial approach towards incremental shape category learning is proposed that learns a new category based on the set of previously learned shape categories.

  @MastersThesis{ 2011mueller,
  author  = {Christian Atanas Mueller},
  title = {3D Object Shape Categorization in Domestic Environments},
  school  = {Bonn-Rhine-Sieg University of Applied Sciences},
  year = {2011},
  address  = {Grantham-Allee 20, 53757 St. Augustin, Germany},
  month = {December},
  abstract  = {In service robotics, tasks without the involvement of
  objects are barely applicable, like in searching, fetching
  or delivering tasks. Service robots are supposed to capture
  efficiently object related information in real world scenes
  while for instance considering clutter and noise, and also
  being flexible and scalable to memorize a large set of
  objects. Besides object perception tasks like object
  recognition where the object's identity is analyzed, object
  categorization is an important visual object perception cue
  that associates unknown object instances based on their
  e.g. appearance or shape to a corresponding category. We
  present a pipeline from the detection of object candidates
  in a domestic scene over the description to the final shape
  categorization of detected candidates. In order to detect
  object related information in cluttered domestic
  environments an object detection method is proposed that
  copes with multiple plane and object occurrences like in
  cluttered scenes with shelves. Further a surface
  reconstruction method based on Growing Neural Gas (GNG) in
  combination with a shape distribution-based descriptor is
  proposed to reflect shape characteristics of object
  candidates. Beneficial properties provided by the GNG such
  as smoothing and denoising effects support a stable
  description of the object candidates which also leads
  towards a more stable learning of categories. Based on the
  presented descriptor a dictionary approach combined with a
  supervised shape learner is presented to learn prediction
  models of shape categories. Experimental results, of
  different shapes related to domestically appearing object
  shape categories such as cup, can, box, bottle, bowl, plate
  and ball, are shown. A classifica- tion accuracy of about
  90% and a sequential execution time of lesser than two
  seconds for the categorization of an unknown object is
  achieved which proves the reasonableness of the proposed
  system design. Additional results are shown towards object
  tracking and false positive handling to enhance the
  robustness of the categorization. Also an initial approach
  towards incremental shape category learning is proposed
  that learns a new category based on the set of previously
  learned shape categories.},
  annote  = {[Winter term 2008][BRSU] - [RoboCup@Home][Ploeger],
  [Kraetzschmar], [Hochgeschwender] supervising}
  }

• M. Hoffmann, “A simulation environment for distributed stream analysis,” Master Thesis, Grantham-Allee 20, 53757 St. Augustin, Germany, 2011.
  [BibTeX] [Abstract]
  

  This work is part of the European research project LIFT – Using Local Inference in Massively Distributed Systems. The goal of LIFT is to apply local inference on nodes of a massively distributed system in order to analyze the systems global phenomena in near real-time. Nowadays distributed systems i.e. large data centers or huge distributed networks grow larger and larger. Out of these circumstances, the problem of monitoring and analyzing these systems becomes more and more complex and almost unmanageable. The goal of this work is to develop a simulation framework for the LIFT context. This framework will support and ease the implementation, testing and evaluation of different data stream filters and global phenomena- (or state-) determination algorithms. This simulation environment intends to make the platform independent prototyping process for distributed stream analysis manageable. The adaptation of an existing multi agent simulation environment will be shown, together with techniques on how to overcome limitation to enable simulation of massively distributed systems. Additionally a framework will be developed which will support the rapid prototyping of local filters. The adapted simulation environment will be evaluated by discussing its computational and parameter dependant scalability. Additionally a filter model, the privacy preserving spatial filter model will be introduced and applied to a real world scenario for validation.

  @MastersThesis{ 2011hoffmannmarius,
  author  = {Marius Hoffmann},
  title = {A simulation environment for distributed stream analysis},
  school  = {Bonn-Rhein-Sieg University of Applied Science},
  year = {2011},
  address  = {Grantham-Allee 20, 53757 St. Augustin, Germany},
  month = {September},
  annote  = {[SS09][LIFT] - [Using Local Inference in Massively
  Distributed Systems][PD Dr. Michael Mock (IAIS)][Prof. Dr.
  Paul G. Pl\"oger (H-BRS)]},
  abstract  = {This work is part of the European research project LIFT --
  Using Local Inference in Massively Distributed Systems. The
  goal of LIFT is to apply local inference on nodes of a
  massively distributed system in order to analyze the
  systems global phenomena in near real-time. Nowadays
  distributed systems i.e. large data centers or huge
  distributed networks grow larger and larger. Out of these
  circumstances, the problem of monitoring and analyzing
  these systems becomes more and more complex and almost
  unmanageable. The goal of this work is to develop a
  simulation framework for the LIFT context. This framework
  will support and ease the implementation, testing and
  evaluation of different data stream filters and global
  phenomena- (or state-) determination algorithms. This
  simulation environment intends to make the platform
  independent prototyping process for distributed stream
  analysis manageable. The adaptation of an existing multi
  agent simulation environment will be shown, together with
  techniques on how to overcome limitation to enable
  simulation of massively distributed systems. Additionally a
  framework will be developed which will support the rapid
  prototyping of local filters. The adapted simulation
  environment will be evaluated by discussing its
  computational and parameter dependant scalability.
  Additionally a filter model, the privacy preserving spatial
  filter model will be introduced and applied to a real world
  scenario for validation.}
  }

• Q. Fu, “The CUDA Acceleration of SIFT algorithm on GPU,” Master Thesis, Grantham Allee 20, 53757 St. Augustin, Germany, 2011.
  [BibTeX] [Abstract]
  

  Since the SIFT algorithm was invented in 1999, it has been well applied in several fields such as object recognition, robotic mapping and navigation, image stitching, 3D model- ing, gesture recognition, video tracking, and match moving. Because of the steadiness and distinctiveness of SIFT, it has become the most popular research topic or tool in computer vision. However, in order to apply SIFT in real time, speed is the biggest difficulty. Since in the SIFT algorithm, the original image has to be modified several times, such as convolution and calculate the Difference of Gaussian. Thus, if the image size is big, the workload would grow exponentially. It would take seconds to perform SIFT on an image with resolution 1024*768. In this thesis, a GPU is used to accelerate SIFT. By taking the advantage of parallel computing and proper management of memory on graphic card, the algorithm is accelerate about 4 times and the result shows that the size of image would not influence the time cost as big as in the CPU implementation.

  @MastersThesis{ 2011fu,
  abstract  = {Since the SIFT algorithm was invented in 1999, it has been
  well applied in several fields such as object recognition,
  robotic mapping and navigation, image stitching, 3D model-
  ing, gesture recognition, video tracking, and match moving.
  Because of the steadiness and distinctiveness of SIFT, it
  has become the most popular research topic or tool in
  computer vision. However, in order to apply SIFT in real
  time, speed is the biggest difficulty. Since in the SIFT
  algorithm, the original image has to be modified several
  times, such as convolution and calculate the Difference of
  Gaussian. Thus, if the image size is big, the workload
  would grow exponentially. It would take seconds to perform
  SIFT on an image with resolution 1024*768. In this thesis,
  a GPU is used to accelerate SIFT. By taking the advantage
  of parallel computing and proper management of memory on
  graphic card, the algorithm is accelerate about 4 times and
  the result shows that the size of image would not influence
  the time cost as big as in the CPU implementation. },
  address  = {Grantham Allee 20, 53757 St. Augustin, Germany},
  annote  = {WS 2006/07, Kraetschmar, Pl\"{o}ger supervising},
  author  = {Quiang Fu},
  month = {February},
  school  = {Bonn-Rhein-Sieg University of Applied Sciences},
  title = {The CUDA Acceleration of SIFT algorithm on GPU },
  year = {2011}
  }

• G. Giorgana, “Facial Expression Recognition from Video Sequences using Spatial and Spatiotemporal Face Descriptors,” Master Thesis, Grantham Allee 20, 53757 St. Augustin, Germany, 2011.
  [BibTeX] [Abstract]
  

  In the last decades, more natural and efficient channels of communication between humans and robots have been investigated. Most of the efforts have been toward analysis of spoken language, but the existing speech-based solutions are still highly error-prone. As a result, researchers have started to consider human facial expressions as an important form to improve the spoken communication. In this thesis we describe three effective methods for the extraction of features that can be used for the recognition of human facial expressions: 2D Gabor filters, Local Binary Patterns (LBP) and Local Binary Patterns from Three Orthogonal Planes (LBP-TOP). Moreover, we investigate the effect of using different sets of parameters on the three approaches. Depending on the technique, the recognition of the expressions is done in still images or complete video sequences. The Cohn-Kanade AU-Coded Facial Expression Database is used throughout this work. All the experiments in this report are carried out using the AdaBoost.MH algorithm. We describe how this ensemble learner can cope with our multi-class problem, while also reducing the number of features. Furthermore, the system is evaluated for two different kinds of weak learners, namely multi-threshold decision stumps and single-threshold decision stumps. Another question we address in this project is the effect of histogram equalization for the face images. Therefore, the impact of this technique in the three methods is analyzed. From the three explored approaches, LBP-TOP takes into account the facial motion that occurs due to facial expressions. On the other hand, 2D Gabor Filters and LBP only describe the instantaneous face appearance, ignoring temporal information. Taking the previous information into account, we compare the approaches and show the advantages of considering motion cues into the analysis. Finally, we present experimental evidence that the three techniques are suitable to recognize facial expressions from live video.

  @MastersThesis{ 2011giorgana,
  author  = {Geovanny Giorgana},
  title = {Facial Expression Recognition from Video Sequences using
  Spatial and Spatiotemporal Face Descriptors},
  school  = {Bonn-Rhein-Sieg University of Applied Sciences},
  year = {2011},
  address  = {Grantham Allee 20, 53757 St. Augustin, Germany},
  month = {February},
  abstract  = {In the last decades, more natural and efficient channels
  of communication between humans and robots have been
  investigated. Most of the efforts have been toward analysis
  of spoken language, but the existing speech-based solutions
  are still highly error-prone. As a result, researchers have
  started to consider human facial expressions as an
  important form to improve the spoken communication.
  In this thesis we describe three effective methods for the
  extraction of features that can be used for the recognition
  of human facial expressions: 2D Gabor filters, Local Binary
  Patterns (LBP) and Local Binary Patterns from Three
  Orthogonal Planes (LBP-TOP). Moreover, we investigate the
  effect of using different sets of parameters on the three
  approaches. Depending on the technique, the recognition of
  the expressions is done in still images or complete video
  sequences. The Cohn-Kanade AU-Coded Facial Expression
  Database is used throughout this work.
  All the experiments in this report are carried out using
  the AdaBoost.MH algorithm. We describe how this ensemble
  learner can cope with our multi-class problem, while also
  reducing the number of features. Furthermore, the system is
  evaluated for two different kinds of weak learners, namely
  multi-threshold decision stumps and single-threshold
  decision stumps.
  Another question we address in this project is the effect
  of histogram equalization for the face images. Therefore,
  the impact of this technique in the three methods is
  analyzed.
  From the three explored approaches, LBP-TOP takes into
  account the facial motion that occurs due to facial
  expressions. On the other hand, 2D Gabor Filters and LBP
  only describe the instantaneous face appearance, ignoring
  temporal information. Taking the previous information into
  account, we compare the approaches and show the advantages
  of considering motion cues into the analysis. Finally, we
  present experimental evidence that the three techniques are
  suitable to recognize facial expressions from live video.},
  annote  = {WS07/08 H-BRS - RoboCup@home Pl\"{o}ger, Kraetzschmar
  supervising}
  }

• F. Hegger, “3D People Detection in Domestic Environments,” Master Thesis, Grantham-Allee 20, 53757 St. Augustin, Germany, 2011.
  [BibTeX] [Abstract]
  

  The ability of detecting people has become a crucial subtask, especially in robotic systems which aim an application in public or domestic environments. Robots already provide their services e.g. in real home improvement markets and guide people to a desired product. In such a scenario many robot internal tasks would benefit from the knowledge of knowing the number and positions of people in the vicinity. The navigation for example could treat them as dynamical moving objects and also predict their next motion directions in order to compute a much safer path. Or the robot could specifically approach customers and offer its services. This requires to detect a person or even a group of people in a reasonable range in front of the robot. Challenges of such a real-world task are e.g. changing lightning conditions, a dynamic environment and different people shapes. In this thesis a 3D people detection approach based on point cloud data provided by the Microsoft Kinect is implemented and integrated on mobile service robot. A Top-Down/Bottom-Up segmentation is applied to increase the systems flexibility and provided the capability to the detect people even if they are partially occluded. A feature set is proposed to detect people in various pose configurations and motions using a machine learning technique. The system can detect people up to a distance of 5 meters. The experimental evaluation compared different machine learning techniques and showed that standing people can be detected with a rate of 87.29% and sitting people with 74.94% using a Random Forest classifier. Certain objects caused several false detections. To elimante those a verification is proposed which further evaluates the persons shape in the 2D space. The detection component has been implemented as s sequential (frame rate of 10 Hz) and a parallel application (frame rate of 16 Hz). Finally, the component has been embedded into complete people search task which explorates the environment, find all people and approach each detected person.

  @MastersThesis{ 2011hegger,
  abstract  = {The ability of detecting people has become a crucial
  subtask, especially in robotic systems which aim an
  application in public or domestic environments. Robots
  already provide their services e.g. in real home
  improvement markets and guide people to a desired product.
  In such a scenario many robot internal tasks would benefit
  from the knowledge of knowing the number and positions of
  people in the vicinity. The navigation for example could
  treat them as dynamical moving objects and also predict
  their next motion directions in order to compute a much
  safer path. Or the robot could specifically approach
  customers and offer its services. This requires to detect a
  person or even a group of people in a reasonable range in
  front of the robot. Challenges of such a real-world task
  are e.g. changing lightning conditions, a dynamic
  environment and different people shapes. In this thesis a
  3D people detection approach based on point cloud data
  provided by the Microsoft Kinect is implemented and
  integrated on mobile service robot. A Top-Down/Bottom-Up
  segmentation is applied to increase the systems flexibility
  and provided the capability to the detect people even if
  they are partially occluded. A feature set is proposed to
  detect people in various pose configurations and motions
  using a machine learning technique. The system can detect
  people up to a distance of 5 meters. The experimental
  evaluation compared different machine learning techniques
  and showed that standing people can be detected with a rate
  of 87.29% and sitting people with 74.94% using a Random
  Forest classifier. Certain objects caused several false
  detections. To elimante those a verification is proposed
  which further evaluates the persons shape in the 2D space.
  The detection component has been implemented as s
  sequential (frame rate of 10 Hz) and a parallel application
  (frame rate of 16 Hz). Finally, the component has been
  embedded into complete people search task which explorates
  the environment, find all people and approach each detected
  person. },
  address  = {Grantham-Allee 20, 53757 St. Augustin, Germany},
  annote  = {[Winter term 2008] [BRSU] - [RoboCup@Home] [Ploeger],
  [Kraetzschmar], [Hochgeschwender] supervising},
  author  = {Frederik Hegger},
  month = {October},
  school  = {Bonn-Rhine-Sieg University of Applied Sciences},
  title = {3D People Detection in Domestic Environments},
  year = {2011}
  }

• M. U. Awais, “An adaptive search engine for Power Point objects,” Master Thesis, Grantham-Allee 20, 53757 St. Augustin, Germany, 2011.
  [BibTeX] [Abstract]
  

  In large scale organizations, with active information technology infrastructure, there is almost no concept of inception of a group discussion, without a formal presentation. In large organizations, volume of these presentations increases rapidly. In order to reuse these presentations, there is a need to develop a knowledge management application which intelligently manages this large amount of data. Intelligent searching should be one part of this application. This report presents a search methodology which is adaptive and intelligent. Using this search methodology a search engine is developed, to search Power Point objects. This engine may be used in conjunction with Microsoft Power Point and other softwares for preparing presentations. To make this search intelligent, two things are introduced. One is, latent semantic analysis, while second is, incorporating user feedback in a way that desires and expectations of users can vary the ranking of results. An innovative design is presented, which delicately inter relates both of these ideas, to present an adaptive behavior of the search system. While testing this application, a new idea for testing a search system is presented. The idea is to use artificial agents for testing, by automating user behaviors.

  @MastersThesis{ 2011awais,
  abstract  = {In large scale organizations, with active information
  technology infrastructure, there is almost no concept of
  inception of a group discussion, without a formal
  presentation. In large organizations, volume of these
  presentations increases rapidly. In order to reuse these
  presentations, there is a need to develop a knowledge
  management application which intelligently manages this
  large amount of data. Intelligent searching should be one
  part of this application. This report presents a search
  methodology which is adaptive and intelligent. Using this
  search methodology a search engine is developed, to search
  Power Point objects. This engine may be used in conjunction
  with Microsoft Power Point and other softwares for
  preparing presentations. To make this search intelligent,
  two things are introduced. One is, latent semantic
  analysis, while second is, incorporating user feedback in a
  way that desires and expectations of users can vary the
  ranking of results. An innovative design is presented,
  which delicately inter relates both of these ideas, to
  present an adaptive behavior of the search system. While
  testing this application, a new idea for testing a search
  system is presented. The idea is to use artificial agents
  for testing, by automating user behaviors.},
  address  = {Grantham-Allee 20, 53757 St. Augustin, Germany},
  annote  = {[2008] [Text mining] - [An adaptive search engine for
  Power Point objects] [Herpers], [Ploeger], [Willebrand]
  supervising},
  author  = {Muhammad Usman Awais},
  month = {March},
  school  = {Bonn-Rhein-Sieg University of Applied Sciences},
  title = {An adaptive search engine for Power Point objects},
  year = {2011}
  }

2010

• F. K. Heinstein, “Improving the Processing and Visualization of Ultrasound Imaging on GPU with CUDA,” Master Thesis, 2010.
  [BibTeX] [Abstract]
  

  Phased Array Ultrasound Technologies (PAUT) techniques is one the Non-Destructive Testing (NDT) methods where we have the possibility to perform inspections with ultrasonic beams of various angles and focal lengths using a single array of transducers. There are software application that precisely controlled the delay of both the emission pulse and the receive signal for each element in an array of transducers. However the processing of the receive beam forming requires a lot of computing power during the image reconstruction phase. The tremendous growth in Graphics Processors Unit (GPU) performance and flexibility has led to an increased interest in performing general-purpose computation on GPU. GPU provides a vast number of simple, data parallel, deeply multithreaded cores and high memory bandwidths. GPU architectures are becoming increasingly programmable, offering the potential for dramatic speedups for a variety of general-purpose applications compared to contemporary general-purpose processors (CPU). That is why with this master thesis we want to explore the area of general purpose computing on GPU, by looking at how the GPU can be utilized to accelerate the processing of ultrasound image. We will use CUDA (Compute Uni ed Device Architecture) which is a language from NVIDIA close to C programming language as a programming model for this GPU implementation. CUDA is a novel technology of general-purpose computing on the GPU allowing users to develop general GPU programs easily. In this thesis I will explore the e ffectiveness of GPU in Ultrasound Image and describes some specific coding idioms that improve their performance on the GPU. GPU performance will be compared to both single-thread version executed on the single-core CPU and multi-threaded OpenMP version executed on the multi-core CPU.

  @MastersThesis{ heinstein2010,
  abstract  = {Phased Array Ultrasound Technologies (PAUT) techniques is
  one the Non-Destructive Testing (NDT) methods where we have
  the possibility to perform inspections with ultrasonic
  beams of various angles and focal lengths using a single
  array of transducers. There are software application that
  precisely controlled the delay of both the emission pulse
  and the receive signal for each element in an array of
  transducers. However the processing of the receive beam
  forming requires a lot of computing power during the image
  reconstruction phase.
  The tremendous growth in Graphics Processors Unit (GPU)
  performance and flexibility has led to an increased
  interest in performing general-purpose computation on GPU.
  GPU provides a vast number of simple, data parallel, deeply
  multithreaded cores and high memory bandwidths. GPU
  architectures are becoming increasingly programmable,
  offering the potential for dramatic speedups for a variety
  of general-purpose applications compared to contemporary
  general-purpose processors (CPU). That is why with this
  master thesis we want to explore the area of general
  purpose computing on GPU, by looking at how the GPU can be
  utilized to accelerate the processing of ultrasound image.
  We will use CUDA (Compute Uni ed Device Architecture) which
  is a language from NVIDIA close to C programming language
  as a programming model for this GPU implementation. CUDA is
  a novel technology of general-purpose computing on the GPU
  allowing users to develop general GPU programs easily.
  In this thesis I will explore the e ffectiveness of GPU in
  Ultrasound Image and describes some specific coding idioms
  that improve their performance on the GPU. GPU performance
  will be compared to both single-thread version executed on
  the single-core CPU and multi-threaded OpenMP version
  executed on the multi-core CPU.},
  author  = {Heinstein, Fotso Kamgne},
  keywords  = {Non-Destructive Testing NDT, Phased Array Testing
  Ultrasound Technologies (PAUT), Ultrasound Images, High
  Performance Computing(HPC), General-Purpose Graphics
  Processing Units (GPGPU), CUDA, OpenMP, Parallel
  Programming, Multithreading, Multicore},
  month = {July},
  owner = {108012516},
  school  = {Hochschule Bonn-Rhein-Sieg},
  timestamp  = {2010.09.08},
  title = {Improving the Processing and Visualization of Ultrasound
  Imaging on GPU with CUDA},
  year = {2010}
  }

• S. Russell and P. Norvig, Artificial Intelligence: A Modern Approach, Third ed., Upper Saddle River, NJ: Prentice Hall, 2010.
  [BibTeX] [Download PDF]

  @Book{ norvig2010,
  address  = {Upper Saddle River, NJ},
  author  = {Russell, S. and Norvig, P.},
  edition  = {Third},
  publisher  = {Prentice Hall},
  series  = {Series in Artificial Intelligence},
  title = {{Artificial Intelligence: A Modern Approach}},
  url = {http://aima.cs.berkeley.edu/},
  year = 2010
  }

• C. Rougier, E. Auvinet, J. Rousseau, A. St-Arnaud, and J. Meunier, “Multiple cameras fall dataset – University of Montreal,” Technical report 1350, DIRO, p. 1–24, 2010.
  [BibTeX]

  @Article{ rougier2010,
  author  = {Rougier, Caroline and Auvinet, Edouard and Rousseau,
  Jacqueline and St-Arnaud, Alain and Meunier, Jean},
  journal  = {Technical report 1350, DIRO},
  title = {{Multiple cameras fall dataset - University of Montreal}},
  year = {2010},
  pages = {1--24}
  }

• R. Szeliski, Computer Vision:Algorithms and Applications, Springer London Ltd, 2010.
  [BibTeX]

  @Book{ szeliski2016,
  author  = {Richard Szeliski},
  title = {{Computer Vision:Algorithms and Applications}},
  publisher  = {Springer London Ltd},
  year = {2010}
  }

• U. Köckemann, “A Relational Robotics Workbench for Learning Experiments with Autonomous Robots,” Master Thesis, 2010.
  [BibTeX] [Abstract]
  

  Even though relational learning approaches, as Inductive Logic Programming and Statistical Relational Learning have great potential benefits in the area of robotic learning, there only have been few applications. We belief one of the reasons to be the large effort involved with performing relational learning experiments in real robotic domains. To resolve these issues and facilite research in this direction we propose the \textit{Relational Robotics Workbench}, a tool that aids researchers in conducting Inductive logic Programming and Statistical Relational Learning experiments in robotic domains. By maximizing re-usability of all components, as learning and inference components on the relational level, or algorithms used for robot control, the \textit{Relational Robotics Workbench} significantly reduces the amount of work in creating relational learning experiments. Furthermore, the learning experiments themselves are made explicit, allowing to store the robots setup, execution of plans (to gather training and test data), setup and execution of learning algorithms and evaluation of the final hypotheses in an easy-to-read configuration file. Functionality provided by the \textit{Workbench} includes: Inductive Logic Programming with Aleph, Statistical Relational Learning With Alchemy, Logic Programming Inference with Prolog, and Statistical Relational Inference with Alchemy. We will provide an analysis of requirements of the proposed \textit{Workbench} and derive a four layered architecture to fulfill them. Along with the description of the architecture, we will provide a full learning experiment from the XPERO project as a running example. To further demonstrate usefulness of the \textit{Workbench}, we then will perform another series of experiments in the ego-motion scenario of the XPERO project. The experiments include descriptions of some additional algorithms implemented in the Workbench, as for instance, a simple scheme to automatically extract learning examples to apply supervised learning methods in an unsupervised fashion. Results for all learning experiments will be presented and examples of the theories that were learned in each scenario will be discussed.

  @MastersThesis{ koeckemann10relational,
  author  = {K{\"o}ckemann, Uwe},
  title = {A Relational Robotics Workbench for Learning Experiments
  with Autonomous Robots},
  school  = {Bonn-Rhein-Sieg University of Applied Sciences},
  year = {2010},
  abstract  = {Even though relational learning approaches, as Inductive
  Logic Programming and Statistical Relational Learning have
  great potential benefits in the area of robotic learning,
  there only have been few applications. We belief one of the
  reasons to be the large effort involved with performing
  relational learning experiments in real robotic domains. To
  resolve these issues and facilite research in this
  direction we propose the \textit{Relational Robotics
  Workbench}, a tool that aids researchers in conducting
  Inductive logic Programming and Statistical Relational
  Learning experiments in robotic domains.
  By maximizing re-usability of all components, as learning
  and inference components on the relational level, or
  algorithms used for robot control, the \textit{Relational
  Robotics Workbench} significantly reduces the amount of
  work in creating relational learning experiments.
  Furthermore, the learning experiments themselves are made
  explicit, allowing to store the robots setup, execution of
  plans (to gather training and test data), setup and
  execution of learning algorithms and evaluation of the
  final hypotheses in an easy-to-read configuration file.
  Functionality provided by the \textit{Workbench} includes:
  Inductive Logic Programming with Aleph, Statistical
  Relational Learning With Alchemy, Logic Programming
  Inference with Prolog, and Statistical Relational Inference
  with Alchemy.
  We will provide an analysis of requirements of the proposed
  \textit{Workbench} and derive a four layered architecture
  to fulfill them. Along with the description of the
  architecture, we will provide a full learning experiment
  from the XPERO project as a running example. To further
  demonstrate usefulness of the \textit{Workbench}, we then
  will perform another series of experiments in the
  ego-motion scenario of the XPERO project. The experiments
  include descriptions of some additional algorithms
  implemented in the Workbench, as for instance, a simple
  scheme to automatically extract learning examples to apply
  supervised learning methods in an unsupervised fashion.
  Results for all learning experiments will be presented and
  examples of the theories that were learned in each scenario
  will be discussed. }
  }

• J. Kang, “Interactive Medical Image Segmentation via User-guided Adaptation of Implicit 3D Surfaces,” Master Thesis, Grantham Allee 20, 53757 St. Augustin, Germany, 2010.
  [BibTeX] [Abstract]
  

  Medical image segmentation plays a very important role in medical image analysis and is considered as a very challenging problem. In this work, a new interactive medical image segmentation framework via user-guided adaptation of implicit 3D surfaces is proposed. The surfaces are mathematically described by a radial basis function that interpolates some user-defined points. This framework allows two kinds of user interactions: Firstly, via mouse clicks the user can define points on the boundary of objects in 3D images. For each point, a 3D surface which interpolates the user-determined points is computed. The user can add or remove points with instant visual feedback until the desired accuracy is reached. Secondly, the user can start an automatic adaption. In this step, the surface computed in step one automatically evolves by maximizing a certain energy function with only external energy. Strong image edges are interpolated without losing the interpolation property of the user-defined points. Our framework is validated by segmenting lung tumors on CT data. Some good segmentation results by this framework show its potential for practical usage.

  @MastersThesis{ 2010kang,
  abstract  = {Medical image segmentation plays a very important role in
  medical image analysis and is considered as a very
  challenging problem. In this work, a new interactive
  medical image segmentation framework via user-guided
  adaptation of implicit 3D surfaces is proposed. The
  surfaces are mathematically described by a radial basis
  function that interpolates some user-defined points. This
  framework allows two kinds of user interactions: Firstly,
  via mouse clicks the user can define points on the boundary
  of objects in 3D images. For each point, a 3D surface which
  interpolates the user-determined points is computed. The
  user can add or remove points with instant visual feedback
  until the desired accuracy is reached. Secondly, the user
  can start an automatic adaption. In this step, the surface
  computed in step one automatically evolves by maximizing a
  certain energy function with only external energy. Strong
  image edges are interpolated without losing the
  interpolation property of the user-defined points. Our
  framework is validated by segmenting lung tumors on CT
  data. Some good segmentation results by this framework show
  its potential for practical usage.},
  address  = {Grantham Allee 20, 53757 St. Augustin, Germany},
  annote  = {WS06/07 Philips Hamburg Research Center Opfer,
  Kraetzschmar},
  author  = {Jingang Kang},
  month = {June},
  school  = {Bonn-Rhein-Sieg University of Applied Sciences},
  title = {Interactive Medical Image Segmentation via User-guided
  Adaptation of Implicit 3D Surfaces},
  year = {2010}
  }

• N. Kharecha, “Robotic Perception for Object Grasping: A ToF Camera based Approach for Environment Sensing,” Master Thesis, Sankt Augustin, Germany, 2010.
  [BibTeX] [Abstract]
  

  If 3D model of the grasping body is available, object surface can be evaluated for given gripper using force-closure criteria. In real-life house hold scenario, such models are not available. To determine grasp for any object, skill to interpret object geometry is required. As grasping of the object occur in 3D world, range sensors are considered better choice of perception. Different ranging devices provide range data but are either time consuming to capture (row-by-row laser scanning) or sparse (stereo camera). Such requirements of skill of grasp determination of model-free objects considering limitation of different ranging device, task of the thesis aim to develop Time-of-flight camera based robotic perception system to determine grasp for unknown object. Time-of-flight camera provides full scale range information at video frame rate but depth measurements are affected by different parameters. Different distance correction schemes are evaluated to suppress noise level. To interpret viewing scene, segmentation of table-top scenario is done to hypothesize different object segments. Single part grasping body is considered to be placed on top of the surface; elongated representation of the footprint of the object segment is presented. As an earlier stage of implementation, a very primitive grasp pose is computed for given elongated representation of the object segment.

  @MastersThesis{ 2010kharecha,
  author  = {Kharecha, Nimeshkumar},
  title = {Robotic Perception for Object Grasping: A ToF Camera based
  Approach for Environment Sensing},
  school  = {Bonn-Rhine-Sieg University of Applied Sciences},
  year = {2010},
  address  = {Sankt Augustin, Germany},
  month = {May},
  abstract  = {If 3D model of the grasping body is available, object
  surface can be evaluated for given gripper using
  force-closure criteria. In real-life house hold scenario,
  such models are not available. To determine grasp for any
  object, skill to interpret object geometry is required. As
  grasping of the object occur in 3D world, range sensors are
  considered better choice of perception. Different ranging
  devices provide range data but are either time consuming to
  capture (row-by-row laser scanning) or sparse (stereo
  camera).
  Such requirements of skill of grasp determination of
  model-free objects considering limitation of different
  ranging device, task of the thesis aim to develop
  Time-of-flight camera based robotic perception system to
  determine grasp for unknown object. Time-of-flight camera
  provides full scale range information at video frame rate
  but depth measurements are affected by different
  parameters. Different distance correction schemes are
  evaluated to suppress noise level. To interpret viewing
  scene, segmentation of table-top scenario is done to
  hypothesize different object segments. Single part grasping
  body is considered to be placed on top of the surface;
  elongated representation of the footprint of the object
  segment is presented. As an earlier stage of
  implementation, a very primitive grasp pose is computed for
  given elongated representation of the object segment.},
  keywords  = {ToF camera, computational geometry, grasp determination}
  }

• N. Khayat, “Monitoring and Analysis of Workflows in Learning Environments,” Master Thesis, Grantham Allee 20, 53757 St. Augustin, Germany, 2010.
  [BibTeX] [Abstract]
  

  With the prosperous spread of new technologies, Technology-Enabled Learning (TEL) Environments are playing a more signicant role, as a mean of delivering an intelligent learning process to the learners. In the context of the European SCY project, a collaborative, learners-centric TEL environment is being developed. Using the SCY system, the learners would execute a mission, represented as a workow, which could be considered as a provided plan for the learners to follow. In this thesis, we have developed a monitoring and analysis system, to be integrated with the SCY system. With the developed system, reference workows and workow executions will be analysed to extract meaningful patterns. The extracted patterns will be used to provide the teachers and pedagogical experts with knowledge insights about the missions executions. Meaningful behavioral attributes are extracted automatically from those patterns, which expresses the behavior model of the learner.

  @MastersThesis{ 2010khayat,
  abstract  = {With the prosperous spread of new technologies,
  Technology-Enabled Learning (TEL) Environments are playing
  a more signicant role, as a mean of delivering an
  intelligent learning process to the learners.
  In the context of the European SCY project, a
  collaborative, learners-centric TEL environment is being
  developed. Using the SCY system, the learners would execute
  a mission, represented as a workow, which could be
  considered as a provided plan for the learners to follow.
  In this thesis, we have developed a monitoring and analysis
  system, to be integrated with the SCY system. With the
  developed system, reference workows and workow executions
  will be analysed to extract meaningful patterns. The
  extracted patterns will be used to provide the teachers and
  pedagogical experts with knowledge insights about the
  missions executions. Meaningful behavioral attributes are
  extracted automatically from those patterns, which
  expresses the behavior model of the learner.},
  address  = {Grantham Allee 20, 53757 St. Augustin, Germany},
  annote  = {WS07/08 Fraunhofer IAIS - SCY "Science Created by You"
  Mock, Pl{\"o}ger supervising},
  author  = {Noury Khayat},
  month = {May},
  school  = {Bonn-Rhein-Sieg University of Applied Sciences},
  title = {Monitoring and Analysis of Workflows in Learning
  Environments},
  year = {2010}
  }

• U. Kayani, “ESNs with sparse output connections,” Master Thesis, Grantham Allee 20, 53757 St. Augustin, Germany, 2010.
  [BibTeX] [Abstract]
  

  This is your abstract.

  @MastersThesis{ 2010kayani,
  abstract  = {This is your abstract.},
  address  = {Grantham Allee 20, 53757 St. Augustin, Germany},
  annote  = {[Summer Semester 2006] [Project Affiliation] - [DEGENA]
  [Dr. Kobialka] and [Prof. Dr. Ploeger]},
  author  = {Umer Kayani},
  date-added  = {2016-09-04 11:54:23 +0000},
  date-modified  = {2016-09-04 11:54:23 +0000},
  month = {January},
  school  = {Bonn-Rhein-Sieg University of Applied Sciences},
  title = {ESNs with sparse output connections},
  year = {2010}
  }

2009

• J. Deng, W. Dong, R. Socher, L. Li, K. Li, and L. Fei-Fei, “ImageNet: A Large-Scale Hierarchical Image Database,” in 2009 IEEE Conference on Computer Vision and Pattern Recognition, 2009, pp. 248-255.
  [BibTeX]

  @InProceedings{ deng2009,
  author  = {Deng,Jia and Dong, Wei and Socher, Richard and Li, Li-Jia
  and Li, Kai and Fei-Fei, Li},
  booktitle  = {2009 IEEE Conference on Computer Vision and Pattern
  Recognition},
  title = {{ImageNet: A Large-Scale Hierarchical Image Database}},
  year = {2009},
  pages = {248-255}
  }

2008

• N. Noury, P. Rumeau, A. K. Bourke, G. ÓLaighin, and J. E. Lundy, “A Proposal for the Classification and Evaluation of Fall Detectors,” IRBM, vol. 29, p. 340 – 349, 2008.
  [BibTeX]

  @Article{ noury2008,
  author  = {Noury, N. and Rumeau, P. and Bourke, A. K. and
  {\'{O}}Laighin, G. and Lundy, J. E.},
  journal  = {IRBM},
  title = {{A Proposal for the Classification and Evaluation of Fall
  Detectors}},
  year = {2008},
  pages = {340 -- 349},
  volume  = {29}
  }

• A. Zakharov, “Robust navigation in everyday environment,” Master Thesis, Grantham Allee 20, 53757 St. Augustin, Germany, 2008.
  [BibTeX] [Abstract]
  

  Autonomous navigation is one of the challenging task in robotics. In order to navigate safely, a robot needs robust and reliable sensing mechanisms. Though there are a variety of range sensors for obstacle detection on the market, there is no straight forward solution, which sensor’s system and data processing mechanism apply in order to get reliable and low cost robot navigation for everyday environment. There is no ultimate range sensor, each type of sensors has it own advantages and disadvantages. The goal of this project is to design a navigation system, which is a combination of affordable range sensors (hardware) and data processing and enhancement algorithms (software), which would eliminate the sensors’ shortcomings.

  @MastersThesis{ 2008zakharov,
  abstract  = {Autonomous navigation is one of the challenging task in
  robotics. In order to navigate safely, a robot needs robust
  and reliable sensing mechanisms. Though there are a variety
  of range sensors for obstacle detection on the market,
  there is no straight forward solution, which sensor's
  system and data processing mechanism apply in order to get
  reliable and low cost robot navigation for everyday
  environment. There is no ultimate range sensor, each type
  of sensors has it own advantages and disadvantages. The
  goal of this project is to design a navigation system,
  which is a combination of affordable range sensors
  (hardware) and data processing and enhancement algorithms
  (software), which would eliminate the sensors'
  shortcomings.},
  address  = {Grantham Allee 20, 53757 St. Augustin, Germany},
  annote  = {SS06 [Prassler], [Shakhimardanov]},
  author  = {Alexey Zakharov},
  month = {June},
  school  = {Bonn-Rhein-Sieg University of Applied Sciences},
  title = {Robust navigation in everyday environment},
  year = {2008}
  }

2007

• C. Zach, T. Pock, and H. Bischof, “A Duality Based Approach for Realtime TV-L1 Optical Flow,” in Pattern Recognition, 2007, p. 214 – 223.
  [BibTeX]

  @InProceedings{ zach2007,
  author  = {Zach, C. and Pock, T. and Bischof, H.},
  booktitle  = {Pattern Recognition},
  publisher  = {Springer Berlin Heidelberg},
  title = {{A Duality Based Approach for Realtime TV-L1 Optical
  Flow}},
  year = {2007},
  volume  = {4713},
  pages = { 214 -- 223}
  }

2006

• T. Fawcett, “An introduction to ROC analysis,” Pattern Recognition Letters, vol. 27, p. 861 – 874, 2006.
  [BibTeX]

  @Article{ fawcett2006,
  title = {{An introduction to ROC analysis}},
  journal  = {Pattern Recognition Letters},
  volume  = {27},
  pages = {861 -- 874},
  year = {2006},
  author  = {Tom Fawcett}
  }

2005

• A. Bruhn, J. Weickert, and C. Schnoerr, “Lucas/Kanade meets Horn/Schunck: Combining local and global optic flow methods,” International Journal of Computer Vision, vol. 61, p. 211 – 231, 2005.
  [BibTeX]

  @Article{ bruhn2005,
  author  = {Bruhn, Andres and Weickert, Joachim and Schnoerr,
  Christoph},
  journal  = {International Journal of Computer Vision},
  title = {{Lucas/Kanade meets Horn/Schunck: Combining local and
  global optic flow methods}},
  year = {2005},
  pages = {211 -- 231},
  volume  = {61},
  issn = {0920-5691}
  }

2003

• G. Farnebäck, “Two-Frame Motion Estimation Based on Polynomial Expansion,” in Image Analysis, 2003, p. 363 – 370.
  [BibTeX]

  @InProceedings{ farneback,
  author  = {Farneb{\"a}ck, Gunnar},
  booktitle  = {Image Analysis},
  publisher  = {Springer Berlin Heidelberg},
  year = {2003},
  title = {{Two-Frame Motion Estimation Based on Polynomial
  Expansion}},
  pages = {363 -- 370},
  volume  = {2749}
  }

• A. Name, “Book Title,” Lecture Notes in Autonomous System, vol. 1001, p. 900–921, 2003.
  [BibTeX]

  @Article{ art1,
  author  = "Author Name",
  title = "Book Title",
  journal  = "Lecture Notes in Autonomous System",
  volume  = "1001",
  publisher  = "UFO",
  pages = "900--921",
  year = "2003",
  coden = "LNCSD9",
  issn = "0302-2345",
  bibdate  = "Sat Dec 7 10:05:42 MST 2003"
  }

1981

• B. K. P. Horn and B. G. Schunck, “Determining Optical Flow,” Artificial Intelligence, vol. 17, p. 185 – 203, 1981.
  [BibTeX]

  @Article{ horn1981,
  title = {{Determining Optical Flow}},
  journal  = {Artificial Intelligence},
  year = {1981},
  author  = {Horn, Berthold K.P. and Schunck, Brian G.},
  volume  = {17},
  pages = {185 -- 203},
  issn = {0004-3702}
  }