A Vision-based Framework for Intersection Monitoring and Signal Evaluation

This work presents a typical framework for traffic measurement estimation at intersections which can assist us with traffic signals evaluation through collaboration of the computer vision techniques and traffic simulators. The proposed system is a deep-visual tracking pipeline which generates traffic measurements at different time frames. The off-the-shelf YOLO object detection architecture variants are cascaded with a discriminative correlation filter (CSRT) within the system for tracking and producing long term trajectories of road users including vehicles and pedestrians. Experimental evaluation was performed on three different intersections and YOLOv4 demonstrated better performance in pedestrian detection and traffic measurement estimation. As an application of the proposed system, traffic measurements including frame level count, turning movement counts and waiting time can be further incorporated into a traffic simulator to investigate intersection signals with current timer settings and find an appropriate signal algorithm which minimizes waiting time.