University of Bremen: Important advances through cognitive neuroinformatics
In a project with the automotive supplier Continental, the Cognitive Neuroinformatics working group has now contributed important research successes to the development of advanced driver assistance systems. Complex traffic situations are better recognized with the help of artificial intelligence.
PRORETA 5 is the name of the research project that the automotive supplier recently completed with its scientific cooperation partners – in addition to the University of Bremen, the TU Darmstadt and the TU Iași (Romania) were involved. “At the end there was a driving demonstration in Darmstadt. There we presented autonomous driving functions that we worked on intensively,” says Professor Kerstin Schill, head of the Cognitive Neuroinformatics working group at the University of Bremen. “The research vehicle was able to autonomously follow the course of the road with a predefined destination and to react to other road users – pedestrians, cyclists and other vehicles. In the case of a simulated sensor failure, which resulted in the missing detection of an object,
Algorithms should make driving decisions
The goal of the PRORETA research project was the development of algorithms. They should use sensor data to derive correct driving decisions that are comparable to those made by humans. At an unregulated intersection, for example, it is a challenge to interpret all objects relevant to the planned direction of travel. It’s about their direction of movement, intention and priority in traffic. Artificial intelligence (AI) should be able to make reliable decisions without human intervention. “The great advantage of AI: After a training phase, it is able to draw the right conclusions based on what it has learned, even in unfamiliar situations,” explains the computer science professor. “One part of the project was to observe the human drivers how they themselves reduce and evaluate the complexity of the environment. The adaptive algorithms are now trained according to similar principles.”
In the project, the Cognitive Neuroinformatics working group investigated AI methods for environmental perception – objects and obstacles should be recognized in the environment. In addition, new methods for human attention control (Human Attention Modeling) based on camera data were developed. In the process, conspicuousness maps are created that determine relevant areas in the image in which, for example, other road users or signs appear. In addition, new mathematical models were developed that correctly represent the position, orientation, speed or size of other road users and describe complex vehicle geometries.
Tasks are now solved more efficiently, more robustly and more securely
Finally, object tracking was implemented, which is able to perceive other road users in the surveillance area and to estimate their condition over time. “These methods ensure that the corresponding tasks can be solved more efficiently, more robustly and more securely. They therefore make an important contribution to highly automated and autonomous driving,” says Kerstin Schill. “The project is an ideal example of how profitable cooperation between university and commercial research can work. Projects like PRORETA strengthen Germany as a business location on both levels.”
The research contributions of the working group in detail:
Jaime Maldonado worked on modeling human attention (Human Attention Modeling) in the context of autonomous driving. In particular, an attention-driven pipeline consisting of two components was developed. On the one hand, relevant areas in camera images are determined using so-called saliency maps. On the other hand, the driver’s gaze is projected into the image to expand the relevant area. As a result, relevant and irrelevant regions in the image can be distinguished and processed more efficiently by subsequent algorithms.
Andreas Serov implemented an object tracking that detects relevant objects in the vehicle’s surveillance area and determines their position, speed, orientation and size in real time. A list of tracked objects is made available to the following modules (prediction, planning and control) for further processing. Object tracking is based on RADAR and LIDAR data. The state of each object is estimated with a probabilistic filter, where the state is processed on a manifold.
Lino Giefer examined theoretical principles for state estimation and representation in autonomous driving. In particular, he set up new models to describe articulated vehicles – such as buses, trams or vehicles with trailers – correctly in mathematical terms. He also examined state and measurement uncertainties for localization and object tracking.
Razieh Khamseh-Ashari researched a multimodal object detection based on LIDAR and camera data using AI methods. Through an early fusion of the sensor inputs, a high-precision localization of objects in the surveillance area is achieved.