Abstract:

In order to address the requirements of high-risk testing environments for validating intelligent vehicle collision avoidance systems, while simultaneously to enrich the content and methods for evaluating autonomous driving scenarios involving vulnerable road users (VRU). This study collected and systematically analyzed traffic accident cases in Guilin City, Guangxi Province, from 2016 to 2020. A total of 1 429 vehicle-VRU collision accident data were screened. Based on accident investigation experience, 13 risk factors were identified, and 10 typical vehicle-VRU collision scenarios applicable to urban traffic conditions in China were constructed using self-organizing K-means clustering analysis. An evaluation model for the complexity of VRU scenarios was established utilizing information entropy theory. The state of variables and the weight of each dimension were determined through a combination of logistic regression models and back propagation (BP) neural networks, and the complexity of various scenarios was calculated. Additionally, the Gaussian mixture model was employed to cluster the complexity levels, resulting in four distinct scene complexity categories. The results show that on roads with a speed limit of 30 km/h, the nighttime side collision between a straight-moving car and an electric bicycle crossing the road outside a pedestrian crossing area is the most complex scenario. The findings in this study provide an experimental scenario reflective of urban road characteristics in China for intelligent vehicle safety testing and offer a basis for the formulation of external VRU collision avoidance strategies and decision-making.

Key words: vulnerable road users (VRU), intelligent vehicles, typical hazardous scenarios, self-organizing, k-means clustering analysis