Machine Learning and Artificial Neural Networks

Empiric-based optimization for deep learning - In this project, we take an empirically based perspective on optimization in Deep Learning.  Little is known about the shape of the typically very high-dimensional losses that are optimized in DL. Instead of basing our research on theoretical assumptions such as convexity, Lipschitz continuity, or interpolation, we directly infer real-world information from common optimization problems in DL.  On this basis, we try to gain new insights and improve optimization. (M. Mutschler)

Avalon - Assisting Search & Rescue with Object Detectors aboard a UAV. This project focuses on the efficient and robust detection of people and ships in case of sea accidents, missing people, maritime salvage and other scenarios. For that, new neural network architectures and machine learning methods are proposed that are able to learn robust representations. Access here PeopleOnGrass. (B. Kiefer, M. Messmer, L. Varga)

DeepStereoVision - In this BMBF-funded project, we study the problem of depth estimation via stereo matching using deep neural networks. (F. Shamsafar, R. Rahim)

Tri-camera Stereo Vision - With a tri-camera configuration, this project investigates depth perception using stereo vision technique from three viewpoints. (F. Shamsafar)

The TCML cluster - The Training Center for Machine Learning (TCML) is a BMBF funded project which consists of a large GPU cluster. (M. Mutschler)

iBinPick - This BMWi-funded project investigates the problem of bin picking from a bin full of small identical objects. (F. Shamsafar, T. Höfer, N. Benbarka)

Person Detection using weakly supervised localization - This project deals with detection of people in thermal images (Infrared) using Deep Convolutional Neural Networks (CNNs) (H. Riaz)

Autonomous Mobile Robots

Table Tennis Robot - A KUKA Agilus robot is used to play table tennis. The system includes ball trajectory and spin detection, prediction of the hitting point and robot stroke motion. (J. Tebbe, Y. Gao)

Object-based Semantic SLAM for UAVs - Research on object-based semantic simultaneous localization and mapping for flying robots. (Y. Wang)

cs::APEX - A framework for visual programming and data flow driven design that allows users to quickly prototype new algorithms by connecting inputs and outputs of computation nodes at run time. It currently is mainly used to prototype robot vision algorithms. (S. Buck, R. Hanten, A. Zwiener)

GeRoNa (Generic Robot Navigation) - A modular robot navigation framework, that bundles path planning and path following (including obstacle detection) and manages communication between the individual modules. It is designed to be easily extensible for new tasks and robot models. (S. Buck, G. Huskić)

3D Mapping using NDT - Research on robust and real-time capable mapping of multi-dimensional data, combining NDT and occupancy probabilities. (C. Schulz, R. Hanten)

Mobil Assist - Based on the electric wheeled walker beActive+e an intelligent wheeled walker is developed. It can support the users by avoiding obstacles, perform local and global navigation and predicting path traversability. (J. Jordan)

Mobile Manipulation - Research on save autonomous picking and placing of objects. Safety shall be enforced through tactile feedback by using joint torque sensing. (A. Zwiener)

MuSe-SMC - Research on Sequential Monte Carlo methods for state estimation and tracking. Currently 2D localization is supported. (R. Hanten, C. Schulz)

Control of Outdoor Robots - Research on a quickly moving outdoor robot, which should be able to follow a fast moving person, while driving on an uneven terrain and avoiding both static and dynamic obstacles on the way. (G. Huskić)

PATSY - (Person-recognizing Autonomous Transportation SYstem). Together with E&K Automation GmbH, Reutlingen, we develop the prototype of a transport system for medical containers in hospitals, which can recognize persons and other static and dynamic obstacles and plan/replan its route accordingly. (S. Buck, G. Rauscher, R. Hanten)

3D Person Detection - Research on robust and real-time capable person detection based on 3D data segmentation and visual feature classification. (R. Hanten, S. Buck, P. Kuhlmann, S. Otte)