Comparative Analysis of YOLO Algorithms for Intelligent Traffic Monitoring

Abstract

The growing traffic congestion is becoming a major challenge in cities. The research aims are to develop a control framework that can do intelligent scheduling of traffic light according to instantaneous traffic density feedback from traffic cameras at signalized crossroads. The research begins by doing a comparative performance analysis of YOLO (You Only Look Once), object detection algorithm, and its different evolved versions (v1, v2, v3, v4) for the real-time traffic scenarios. The Microsoft COCO (Common Object in Context) dataset was used to assess the performance of various algorithms, as well as their strengths and limitations, using common criteria like frames per second (FPS) and mean average precision (mAP), across all the implementations on real video sequences of road traffic. Once we have detected a vehicle, it is assigned an id then it needs to be tracked over different frames using an efficient Real-time Tracking algorithm like SORT (Simple Online and Real-time Tracking). The phases of traffic signals can then be optimized based on the data collected, specifically, queue density and waiting time per vehicle, in order to let as many vehicles to pass safely with the least amount of waiting time. The findings of the project show that YOLO v4 had a considerable edge in detection speed, with FPS and mAP as compared to other YOLO versions for real-time vehicle detection.