Optimizing test-set diversity: Trajectory clustering for scenario-based testing of automated driving systems

The developments in automated driving have raised questions concerning the safety of automated vehicles. The reliable behavior of automated driving functions has to be proven by a testing process. While real-world testing is the evident method for verification, there are significant concerns regarding test scalability. Simulative methods are economical and safe alternatives to assess driving functions' quality. A catalog with scenarios recorded in real-world traffic for resimulation must be assembled, defining a set of requirements the driving function must pass before release. Therefore, to ensure safe behavior, the test catalog must include many diverse scenarios while reducing repetitive scenarios, as they do not pose additional challenges to the virtual vehicle and would unbalance the performance metrics. This work introduces a method to cluster and characterize vehicle behaviors based on trajectory data. The first step of the two-step bottom-up approach assigns individual coordinates of trajectories to clusters. Then, trajectory analysis is applied to histograms representing the distribution of retrieved cluster-IDs. Furthermore, we performed experiments, demonstrating the performance of our algorithm on various real-world data-sets. Using 1608 traffic scenarios, the method reduced the data-set by 61.0 %, eliminating redundant scenarios while upholding scenario space coverage. This algorithm serves as a basis for traffic scenario understanding to optimize test-set diversity.