Video-Based Surrogate Safety Analysis for Conflict Detection at Intersections Using CCTV Footage and Extreme Value Theory

Traditional crash-based safety analysis is reactive and limited by underreporting and delayed insights. This study proposes a proactive conflict analysis framework using traffic CCTV footage and advanced video analytics to detect and classify intersection conflicts. A Bird’s Eye View (BEV) transformation and YOLOv8-OBB detection model were employed to extract accurate vehicle trajectories, followed by conflict detection using Post-Encroachment Time (PET) and Time to Collision (TTC). Peak over Threshold (POT) univariate Extreme Value Theory (EVT) and clustering approach was applied separately to PET and TTC to classify conflicts into low, medium, high and very high severity classes. Manual conflict data was collected by identifying event with PET ≤ 4s which was used to validate the accuracy of the automatic conflict detection framework. The automatic conflict detection framework achieved high accuracy of 99% in identifying different conflict types. Conflict severity analysis revealed that PET and TTC values exceeding 1 second were consistently associated with low-severity outcomes, suggesting that even a minimal buffer of 1 second often provides drivers with sufficient time to execute evasive maneuvers and avoid critical interactions. PET proved to be a more reliable conflict indicator for CCTV-based analysis, as it depends only on spatial overlap, while TTC’s reliance on speed and distance makes it more prone to estimation errors. Overall, this framework provides a scalable, plug-and-play solution for proactive safety monitoring using existing camera infrastructure. It addresses key gaps in previous studies by integrating detection, severity modeling, and behavior analysis, supporting data-driven safety improvements aligned with Vision Zero goals.