We do research in the following fields:
We do research on artificial neural network algorithms, but with a strong emphasis on applications of neural networks in computer vision, mobile robot vision and some robot manipulation.
Theoretical neural network research: On the algorithmic side we developed the PAL Optimizer using a parabolic approximation to estimate step sizes for fast line searching (M. Mutschler). In addition, we worked on automating the design of efficient network models using neural architecture search (K. Laube). With FourierNet, we presented Fourier series-based differentiable shape decoders for instance segmentation (H. Riaz, N. Benbarka).
Application-oriented neural network research: Convolutional neural networks (CNNs) show excellent performance in several application scenarios. In the Avalon project, CNNs are used for maritime search (B. Kiefer, M. Messmer, L. Varga). Depth estimation is the focus of the DeepStereoVision project (F. Shamsafar, R. Rahim). Detection of small objects is performed for the bin-picking project iBinPick (F. Shamsafar, T. Höfer). Deep learning is also used in a person detection project with IR cameras (H. Riaz) and via reinforcement learning methods on our table tennis robot (J. Tebbe, Y. Gao). Hyperspectral images are analysed to predict the ripeness of fruits (L. Varga) and infestations on fields (K. Laube).
The chair is also operating a medium-sized GPU cluster funded by the project TCML with 4 racks of 10 nodes with 4 Nvidia GTX1080Ti, i.e. 160 GPU nodes, plus some smaller GPU servers for individual projects.
Our robotics research aims at robot vision, sensor fusion, navigation, path planning and control of mobile robots. We operate 3 robotics labs and are doing research on wheeled robots, aerial robots and robot manipulator arms.
Wheeled robots research: In the project PATSY (S. Buck, G. Rauscher, R. Hanten) a person recognizing autonomous transport system was developed. It had a 3D person detection based on low resolution time-of-flight camera data, and in the context of this research also the system CS::APEX, a framework for visual programming and data flow driven design of robot algorithms (S. Buck, R. Hanten, A. Zwiener) was developed. Path planning and path following of robots with different drive systems can be done with GeRoNa (S. Buck, G. Huskić). It especially is used for control of outdoor robots (G. Huskic). An electric wheeled walker has been turned into an autonomous mobile robot in the project MobilAssist (J. Jordan). Research on sequential Monte-Carlo methods has turned into the framework MuSE-SMC (R. Hanten, C. Schulz), and 3D Mapping using NDTs (C. Schulz) improves normal distribution maps. In a new project SAFE-AI (F. Engmann, Y. Quan, T. Höfer) safety issues of industrial vehicles, like tractors and communal vehicles in autonomous mode are investigated.
Aerial mobile robots research: In object-based SLAM for UAVs (Y. Wang) CNNs are used to classify objects in SLAM algorithms and to determine relative localization of UAVs in front of buildings (C. Yang). Aerial robots with hyperspectral cameras have been used in the project FarmingIOS (K. Laube, C. Geckeler, S. Rajappa), to detect plant pests early for precision farming.
Robot manipulator research: Our table tennis robot (J. Tebbe, Y Gao) is one of the first worldwide, which can deal with spin balls, and has lead to cooperation projects with KUKA and with Sony Research, the latter on event-based vision (J. Tebbe, A. Ziegler). We did research on tactile feedback by using joint torque sensing in Mobile Manipulation (A. Zwiener) and are currently pursuing a new class of robot arm motion planning algorithms based on geodesics (M. Laux).