An autonomous vehicles' test case extraction method: Example of vehicle-to-pedestrian scenarios

Testing autonomous vehicles (AVs) in hazardous scenarios is a crucial technical approach to ensure their safety. A key aspect of this process is the generation of hazard scenarios. In general, such scenarios are generated through cluster analysis of traffic accident data. However, this approach may not fully capture the criticality of the generated scenarios, as it tends to emphasize the statistical characteristics of the data rather than its real-world applicability. This paper proposes a novel method to enhance scenario adaptation by integrating quantization weights with a new clustering algorithm. These weights, representing the correlation between scenario elements and the AV system, are calculated using fuzzy comprehensive evaluation (FCE). The proposed method is applied to 1044 pedestrian accident cases in China, resulting in the identification of nine categories of typical scenarios and corresponding test schemes for both the perception and decision-making systems of AVs. The results show that the new method increases the proportion of critical scenarios by 17.4 % and 13.6 %, respectively, compared to traditional methods. Overall, the critical scenarios generated in this paper can significantly improve the testing efficiency and safety of AVs.