Corner Case Dataset for Autonomous Vehicle Testing Based on Naturalistic Driving Data

The safe and reliable operation of autonomous vehicles is contingent on comprehensive testing. However, the operational scenarios are inexhaustible. Corner cases, which critically influence autonomous vehicle safety, occur at an extremely low probability and follow a long-tail distribution. Corner cases can be defined as combinations of driving task and scenario elements. These scenarios are characterized by low probability, high risk, and a tendency to reveal functional limitations inherent to autonomous driving systems, triggering anomalous behavior. This study constructs a novel corner case dataset using naturalistic driving data, specifically tailored for autonomous vehicle testing. A scenario marginality quantification method is designed to analyze multi-source naturalistic driving data, enabling efficient extraction of corner cases. Heterogeneous scenarios are systematically transformed, resulting in a dataset characterized by diverse interaction behaviors and standardized formatting. The results indicate that the scenario marginality of the dataset constructed in this study is 2.78 times that of mainstream naturalistic driving datasets, and the scenarios exhibit considerable diversity. The trajectory and velocity fluctuations, quantified at 0.013 m and 0.021 m/s, respectively, are consistent with the kinematic characteristics of real-world driving scenarios. These results collectively demonstrate the dataset’s high marginality, diversity, and applicability.