Accident Scenario Generation with Recurrent Neural Networks

There is a lack of approaches to derive accident scenarios automatically for the testing of Autonomous Drive systems. Current approaches that generate test scenarios do not scale due to the manual work required. Machine learning provides the possibility to automate such tasks. In this paper, an automated approach based on Recurrent Neural Networks to generate accident scenarios is presented. Based on a prototype, our approach is evaluated on temporal data from simulated in-vehicle and V2X data to automatically generate new accident scenarios. The results confirm that generated scenarios resemble the accidents that took place in an exclusive test set.