Automatically Improving Scenario Descriptions Derived From Recorded Traffic

In scenario-based testing of autonomous vehicles, scenario types serve as input for generating test cases. Towards a comprehensive catalog of scenario types, expert-defined scenario types are complemented with types derived from real traffic data. Recorded traffic contains instances of different scenario types: By clustering these instances, a scenario type can be associated with each cluster. This involves adding semantics to recorded scenario instances, our clusters' elements, in the form of maneuver-based descriptions of the behavior of scenario participants. Several works offer approaches to derive such instance descriptions. These approaches are often automated. The quality of the generated descriptions matters, given that subsequent steps build upon them. How to derive good descriptions automatically has not been addressed until now. In our work, (1) we approximate a description's quality by measuring the Fréchet distance between the trajectories obtained by simulating a description and the original trajectories in the dataset; (2) we introduce a quality threshold as to what makes a <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">good</i> description and, based on this threshold, we identify inadequate descriptions; (3) we build upon these evaluation results to modify identified <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">inadequate</i> descriptions automatically. In experiments, we extract and describe a set of 843 scenario instances from the inD intersection dataset: the average reconstruction quality in terms of a description's aggregated Fréchet distance of 0.67 and a ratio of 54% (459) of adequate descriptions improves to 0.34, and 97% (819). The average aggregated distance is reduced by half, i.e., the reconstruction quality is doubled.