A Comprehensive Evaluation Model of Lane-based Risk for Urban Intersections Based on Bayesian Inference and XGBoost

This study addresses the challenges of inadequate researches on lane-scale risk evaluation and the uncertainties in complex interactions at urban signal-controlled intersections. To this end, a comprehensive risk evaluation model is developed based on Bayesian inference and XGBoost. Specifically, this study is based on traffic video data from three intersections in Xi'an: Jixiang Village, Mingguang Road, and Qingsong Road. Two innovative risk-evaluation sets are constructed from the dimensions of temporal and spatial proximities, in which key indicators, including post-entrainment time, maximum speed, distance difference and speed difference are selected to capture dynamic risk characteristics of intersections. Further, Bayesian inference is used to develop a probabilistic evaluation method to address uncertainties in complex interactions at intersections. Next, SHAP value theory of the XGBoost model and Logistic regression are applied to analyze the significance and importance of factors influencing lane risk levels. The results show that: ①The proposed model outperforms baseline models in identifying medium and high-risk interactions. It also more accurately assesses extreme danger interactions, avoiding the overestimation observed in baseline models. ②Among the typical interactions, that between motor vehicle-bicycle, pedestrian-motor vehicle, and pedestrian-bicycle, only a small portion are classified as extreme risk, though medium-risk interactions account for 29.7%, 20.8%, and 34.3%, respectively. ③There are significant differences regarding the risk level across different lanes, with the first lane being more prone to traffic conflicts compared to the second, third, and fourth lanes. ④For all three interaction types, lane risk is mainly influenced by speed, acceleration, and traffic volume. In motor vehicle-bicycle interactions, the highest risk occurs in the first lane and on roads with narrow buffer zones, particularly during morning rush hours and on right-turn lanes. Pedestrian-motor vehicle interactions are primarily influenced by speed and traffic volume, with higher risks in the first lane. For pedestrian-bicycle interactions, narrower bicycle lanes increase the risk of conflicts.