Identification of High-risk AV Accident Scenarios Based on Machine Learning and Association Rules

Abstract

To identify high-risk accident scenarios of autonomous vehicles and their influencing mechanisms, this paper proposes a hybrid model that integrates interpretable machine learning and association rules. Based on accident data from California, USA, three technical features are added to traditional variables: vehicle model year, sensor type, and driving mode. By establishing a ROS-LightGBM model combined with the SHAP algorithm, 10 key risk factors are identified. Association rules are then used to mine the multi-factor coupling effects in high-risk accidents, and high-risk scenario profiles are depicted accordingly. The results show that feature combinations such as the subject vehicle reversing or stopping, the collision object going straight, the collision object being a small vehicle, and the subject vehicle being rear-ended exhibit strong correlations, which can increase the probability of high-risk accidents by a factor of 1.88 to 1.92. Based on these results, three typical high-risk scenarios are summarized: rear-end collision scenarios when the subject vehicle is reversing or stopped, narrow road conflict scenarios under nighttime autonomous driving mode, and collision scenarios involving early model vehicles that are reversing or stopped.