| description abstract | Real-time conflict prediction is an emerging research perspective of proactive road safety measures that can prevent potential risk situations. Previous research used macroscopic traffic flow data while underestimating the relationship between conflict events and conflict vehicle trajectories. This study introduced kinematics and status-related data of conflict vehicles to explore whether there would be a potential conflict in the near future. Prediction models were developed utilizing vehicle trajectories from the Shanxi Wuyu Freeway in China. Considering both rear-end and sideswipe conflicts, extended time-to-collision was proposed to categorize vehicle interactions into three distinct types: conflicts, normal interactions, and undisturbed passings. This classification formed the basis for three distinct prediction tasks. Sixteen predictor variables were derived from vehicle trajectories within a 15-second time window before conflict events. Predictor variables, which captured vehicle velocity and surrounding traffic conditions, were used to build conflict prediction models based on random forest, support vector machine, and artificial neural network. An undersampling algorithm was employed to solve the sample imbalance. Twelve conflict prediction models were compared in terms of precision, recall, specificity, F1-score, and area under the curve. The results indicated that binary conflict prediction models based on support vector machines presented the best prediction performance, achieving higher scores across multiple evaluation metrics. Among the predictor variables, the mean of longitudinal velocity emerged as the most significant factor, while the hour of conflict occurrence was also identified as an essential state variable. The exploration of different time windows indicated that shorter timeframes before conflict events enhance the performance of real-time conflict prediction models. Finally, the proposed real-time conflict prediction models and contributing factors provide a novel approach for simplifying conflict estimation by eliminating the need to analyze complex kinematic relationships among vehicles. It also contributes to designing further proactive safety systems for conflict warnings implemented on vehicle dashboards. | |
