CAPTURING PEDESTRIAN-VEHICLE CONFLICTS USING COMPUTER VISION: PREDICTING THE SEVERITY OF CONFLICTS AND EXAMINING THE EFFECTS OF PEDESTRIAN, VEHICLE AND SIGNAL TIMING-RELATED FACTORS

According to crash statistics, the United States witnessed 6,205 pedestrian fatalities and 76,000 injuries on its roads. These numbers are still unacceptably high and urge the need for proactive measures to mitigate pedestrian-vehicle conflicts and strive toward achieving a crash-free society. This research focuses on object detection and tracking algorithms, specifically YOLOv4 and Deep SORT, to examine pedestrian safety at a signalized intersection with a fixed cycle time and an intersection controlled by rectangular rapid flashing beacons (RRFBs). Long short-term memory (LSTM) neural network and adjacent-category models were developed for both intersections to predict the severity of pedestrian-vehicle conflicts and examine the effects of pedestrian, vehicle, and signal timing-related factors. The system can warn drivers 2s ahead about a potential conflict with a pedestrian, fostering a proactive approach to mitigating conflicts and enhancing overall road safety. The findings also provided evidence that increasing the yellow time and the RRBF flashing time significantly lowered the severity of pedestrian-vehicle conflicts at both intersections, emphasizing the importance of these two signal timing factors as integral measures for enhancing pedestrian safety and minimizing potential conflicts with vehicles.