Urban high-risk scenarios for automated vehicle safety testing: A generation and generalization method based on...

To improve the effectiveness of automated vehicles (AVs) safety testing and address key issues including distribution bias in road data collection and insufficient coverage of high-risk events, this study adopts a generation and generalization method to establish high-risk scenarios based on the AV Testing - High-Risk City Accident Dataset (AVT-HRCAD). Firstly, Cramer's V coefficient and eta squared coefficient are employed to identify risk variables that significantly affect accident severity. Secondly, the K-medoids clustering algorithm, iteratively optimized based on Gower distance, generates baseline risk scenarios. Ultimately, a Risk Index (RI) measures risk levels, while the NRPE criterion-assessing Number, Risk, P-value, and Effect-is intended to evaluate and generalize test situations. Principal findings indicate: Nine, seven, and seven key risk variables were identified for expressways, intersections, and road segments, respectively. Ego behavior, target type, collision angle, and lighting conditions consistently emerged as consistently significant risk factors across all three road types. Scenario generalization effectively addressed low-sample/high-severity variables (e.g., three-wheelers), broadening 18 baseline risk scenarios into general-risk, high-frequency-risk, and long-tail high-risk scenarios. A total of 93 urban high-risk test scenarios were established to assess AV capabilities in risk avoidance (across different vehicle types and collision angles), safety distance determination, and distance maintenance. This method provides a more authentic and valuable testing platform for comprehensive AV safety evaluation.