Analysing Pedestrian Safety at Intersections Using Video and AI-Based Tracking

This study aims to evaluate pedestrian safety at a signalised intersection involving left-turning vehicles, using computer vision techniques. YOLO, DeepSORT, and Neural Networks were employed for object detection and tracking, as well as for calculating the Surrogate Safety Measure (SSM), Post-Encroachment Time (PET). The object classifier achieved an average precision of 0.890, a recall of 0.770, and an average accuracy of 0.830. PET values in this study ranged from a minimum of 0.52 seconds to a maximum of 8.72 seconds. Among the 29 observations, 8 indicated a high risk of collision, and 17 showed dangerous conflicts between pedestrians and vehicles. Overall, the PET analysis reveals that high-risk interactions between left-turning vehicles and pedestrians, both on the nearside and far-side crosswalks, occur frequently at the intersection. The integration of a deep learning-based tracking system significantly improves the accuracy of PET estimation by enabling precise identification and monitoring of pedestrian and vehicle movements.