Using Computer Vision to Analyze the Sequence of Vehicles Passing Through Regulated Intersections

Abstract

Many papers on traffic management have dealt with optimizing traffic light signals with the assumption that the traffic flow (TF) speed is fixed or follows a given distribution. In our study, we focused on determining vehicle speed in real time and assessing its impact on the delay of vehicles. A convolutional neural network (YOLOv3) is used to detect vehicles and determine their speed through the real-time processing of video streams from traffic surveillance cameras. The developed system can identify and classify 11 traffic flow types and track the trajectory and speed of vehicles passing through a regulated intersection. When analyzing the obtained data, we identified two important factors contributing to the formation of vehicle queues at intersections during a red light. We revealed the nature and statistically significant measure of reducing free vehicle movement speed depending on the queue size, and determined the maximum vehicle queue size which does not significantly affect the dynamics of passing through an intersection. The obtained data allow us to optimize adaptive regulation and synchronization of traffic lights based on the recommended traffic flow speed.