Investigating the factors affecting the severity of autonomous vehicle-related conflicts

AbstarctAutonomous Vehicles (AVs) are considered promising tool to enhance road safety. While some previous studies have investigated AV risk factors, this research is the first to leverage the combined power of Automated Machine Learning (AutoML) and SHAP explainability to analyze the complex, non-linear AV-road user interactions using real-world interaction data. The AutoML results indicated that the Gated Recurrent Unit model outperformed other models in predicting AV conflict severity for both road segments and intersections. Using SHAP explanations, the study pinpointed the average speed of vehicles, road user volume and the presence of bus stops as the primary factors determining conflict severity in road segments. While the percentage of cyclists, the number of lanes, and the presence of cycle lanes were not found to have a significant impact on conflict severity. For intersections, traffic control devices, road user volume, and protected left turn phases at signalized intersections were identified as the most impactful factors. Meanwhile, the time of conflict and the presence of cycle lanes were deemed to have little to no impact on conflict severity.