Influencing factors of vehicle accident severity between autonomous and human-driven vehicles

ObjectiveTo investigate the differences in influencing factors of accident severity between human-driven and autonomous vehicles, this study proposes a comparative analysis method for predicting accident severity and deciphering causal mechanisms across different driving modes, based on machine learning algorithms combined with SHAP.MethodThis study extracted accident data for human-driven vehicles and autonomous vehicles, based on US-Accidents and AVOID datasets.Four machine learning algorithms, i.e., XGBoost, random forest, support vector machine, and K-nearest neighbors, were utilized to establish accident prediction models with two driving modes respectively. The interpretable SHAP algorithm was then applied to analyze the factor effects and conduct comparative analysis.ResultThe extreme gradient boosting algorithm delivered the best overall performance in predicting human-driven vehicle accident severity, with an accuracy of 0.840 3, precision of 0.855 9, recall of 0.840 3, and F1-score of 0.824 4. While random forest algorithm achieved optimal performance in predicting autonomous vehicle accident severity, with an accuracy of 0.972 9, precision of 0.972 8, recall of 0.972 9, and F1-score of 0.972 8. Significant differences exist in the influencing mechanisms of accident severity across different driving modes, with the same factors exhibiting differential or even opposing effects in human-driven versus autonomous driving scenarios. Specifically, the incident scale contributed most significantly to predicting human-driven vehicle accident severity, whereas the spring season was the most influential factor for predicting autonomous vehicle accident severity. Furthermore, spring and peak hours exerted opposite effects on the accident severity with two vehicle types, while summer had a consistent effect on the accident severity of both driving modes.ConclusionThe findings can provide a scientific basis for seasonal and time period specific traffic management, as well as for the safety oversight of autonomous driving systems.