Analysis of autonomous vehicle accident severity factors based on hybrid model

To deeply investigate the impacts of interactions between various factors in autonomous vehicle accidents on the accident severity, a hybrid model of interpretable machine learning and tree-augmented naive (TAN) Bayes network was proposed considering the factor interactions. According to the summarized public dataset of autonomous vehicle accident reports in California from 2014 to 2023, 13 factors were determined to be the independent variables. The dependent variable (accident severity) was classified into three classes: no injury, minor, and moderate-to-major accidents. Four machine learning algorithms, namely, extreme gradient boosting, random forests, K-nearest neighbors, and support vector machines, were applied to establish a classification prediction model for autonomous vehicle accident severity. The classification prediction model with the optimal performance was selected as the base model. Then the base model was combined with the Shapley additive explanations (SHAP) algorithm to select the key factors and the TAN Bayes network was established to extract nine interaction factor pairs. Finally, the SHAP algorithm was employed to analyze the single factors as well as three groups of interaction factor pairs with significant interaction effects. Research results show that for no injury, minor and moderate-to-major accidents, the accident area of accident partner, vehicle driving mode, and subject vehicle pre-accident motion state have a significant impact on accident severity, respectively. After 2020, when the rear part of the autonomous vehicle is collided, the SHAP interaction values for no injury accidents are all positive, while those for moderate-to-major accidents are all negative. When the autonomous vehicle's accident area as the front and the object vehicle accident area as the rear, only 16.7% of the SHAP interaction values are positive in moderate-to-major accidents. If both the accident areas of the autonomous vehicle and object vehicle are the front, 89.1% of the SHAP interaction values are positive in minor accidents. Therefore, autonomous vehicles can timely change the running track to avoid vehicle frontal accidents and then reduce the severity of vehicle accidents.