Detecting Edge Cases from Trajectory Datasets Using Deep Learning Based Outlier Detection

The biggest challenge to overcome for automated vehicles is to prove their safety, as these vehicles are solely responsible for the passengers’ safety. The scenario-based testing approach promises an efficient safety validation procedure by only testing the safety in relevant scenarios. An open question is how to select the relevant scenarios for testing. So-called edge cases are frequently named in the automated driving domain to be important scenarios for testing automated vehicles. However, it is not an easy task to define what an edge case is and to find and validate them. In this work, we present a novel data-driven approach to finding edge cases in trajectory datasets using deep learning-based outlier detection. We develop a method that calculates embeddings for driving scenarios in a two-stage process. In the dimensionally reduced embedding space, outliers represent potential edge cases. We apply the approach to the exiD dataset and find potential edge cases. For validation, we present the found potential edge cases to a group of experts. The experts validate that the approach is capable of detecting edge cases in trajectory datasets.