They must prove conclusively that they do not pose a risk to anyone until autonomous vehicles engage in road traffic. New software developed at the Technical University of Munich (TUM) prevents accidents every millisecond by predicting various variants of a traffic situation.
An intersection is approached by a car. Another vehicle jets out of the lane, but it’s not yet clear if it’s going to turn left or right. At the same time, directly in front of the vehicle, a pedestrian steps into the lane and there is a bicycle on the other side of the street. In general , individuals with road traffic expertise can better determine the activities of other traffic participants.
“For autonomous vehicles powered by computer programs, these kinds of situations present an immense challenge,” says Matthias Althoff, TUM Professor of Cyber-Physical Systems. “But autonomous driving will only be approved by the general public if you can guarantee that vehicles do not put other road users at risk — no matter how complicated the traffic situation might be.”
Algorithms that peer into the future
The overall aim is to ensure that they do not cause accidents when designing applications for autonomous vehicles. Althoff, a member of TUM’s Munich School of Robotics and Artificial Intelligence, and his team have now created a software module that analyzes and forecasts events on a permanent basis while driving. Vehicle sensor data is collected every millisecond and analyzed. For any traffic participant, the program will measure all potential movements — provided they conform to the road traffic regulations — allowing the machine to look into the future for three to six seconds.
Streamlined models for swift calculations
Previously, this form of comprehensive traffic situation forecasting was considered too time-consuming and thus impractical. But now the theoretical feasibility of real-time data analysis with simultaneous simulation of potential traffic incidents has not only been demonstrated by the Munich research team: they have also shown that it provides accurate performance.
Simplified dynamic models make fast calculations possible. In order to measure possible future positions a car or a pedestrian might assume, so-called accessibility analysis is used. The estimates become prohibitively time-consuming when all characteristics of road users are taken into account. That is why, with simpler models, Althoff and his team work. In terms of their range of movement, these are superior to the true ones — and mathematically easier to manage. This increased freedom of movement enables the models to represent a greater number of possible positions, but
includes the sub-set of positions required for real road users.
Real traffic data for a virtual test environment
The computer scientists built a virtual model for their assessment based on real data that they had collected during test drives with an autonomous vehicle in Munich. This made it possible for them to construct a test environment that closely portrays regular traffic scenarios. “We were able to determine through the simulations that the safety module does not lead to any loss of performance in terms of driving actions, the predictive predictions are right, accidents are avoided, and the vehicle is demonstrably brought to a safe stop in emergency situations,” Althoff sums up.
The computer scientist stresses that the production of autonomous vehicles will be simplified by the new protection software since it can be integrated with all regular programs for motion control.
Materials provided by Technical University of Munich (TUM). Note: Content may be edited for style and length.
No CommentsLeave a comment Cancel