The focus of the team relates first of all to problems related to environment perception and understanding, to the research and development of related methods and algorithms. Let us give a brief overview on the main topics of the laboratory.
Development of Perception Algorithms:
The main objective is to develop neural network architectures (first of all for 3D object detection and end-to-end models) capable of learning from multiple type of data such as LiDAR, RADAR, cameras. Such networks are able to learn the synergies present between different type of data and thus may provide more reliable results, more robust behavior than networks operating on single type of data only. Besides the developent of algorithms for raw sensor fusion, we are working on the development of a so called central perception system capable of providing object detections by a cluster of infrastructural and vehicular sensors, fusing them together in real time and thus providing a digital twin of the traffic in the area covered by these sensors.
Measurement and Testing:
Besides testing our algorithms on publicly available datasets we perform also experiments under real circumstances. For this purpose we have built a measurement vehicle equipped with LiDARS, RADAR, Cameras, dGPS and V2X communication unit which allows us to test our algorithms under real circumstances in real-time. Here the calibration of sensors and that of the whole system is of key importance.
Verification of Models:
The risks we face in many safety-critical applications such as assisted and highly automated driving are rapidly changing because the internal logic of the employed deep learning perception methods is not humanly interpretable. This gives rise to new kinds of mistakes and more worryingly, new vectors of attack exploiting these new vulnerabilities. In our research we explore ways to quantify the robustness and verify guaranteed properties of machine learning systems. The goal is to design robust perception systems that come with mathematically certified properties. As a prerequisite for this, we have to develop more efficient verification methods and also find novel ways of formulating humanly interpretable specifications in ultra-high dimensional input domains (e.g. sensory data).