Conflict extraction and characteristics analysis at signalized intersections using trajectory data

Abstract

Potential safety problems at signalized intersections can be recognized most effectively by identifying serious traffic conflict events and analyzing them in different scenarios. However, most studies extract conflicts using different threshold values, and lack thorough examinations and screening. An approach that may not reflect actual traffic conditions may result in the extraction of non-conflict events. Additionally, there lacks a of comprehensive conflict analysis that integrates diverse analyses across different scenarios and operational characteristics. Therefore, based on video data in Shanghai, China, this study provides a comprehensive method for extracting and analyzing serious conflicts. First, video recognition and trajectory reconstruction were conducted. Traffic conflict events were identified by integrating operational characteristics and road geometric design, and K-means++ was used to cluster the severity of these conflicts. Second, parameterized rear-end conflict, lane-changing conflict, and crossing conflict scenarios were reconstructed to analyze serious conflict events. These events were then analyzed separately from the perspectives of conflict distribution, path and turning modes, and vehicle types under three scenarios. The results show that the best clustering effect is achieved using jerk, longitudinal relative distance, and relative distance. Moreover, the validated time-to-collision (TTC) thresholds for classifying conflicts are 0.97 s for serious conflicts, 1.51 s for light conflicts, and 2.09 s for potential conflicts. The study also identifies rear-end conflicts, right-turn conflicts, and conflicts involving cars and trucks as the most serious. These findings support the extraction of key features from intersection scenarios, and facilitate the testing of hazardous scenarios for automated driving.