Multi-class Vehicle Counting System for Multi-view Traffic Videos

Abstract

This paper presents a vehicle counting system using multi-class vehicle detection using YOLOX and multi-object tracking using ByteTrack. Counting is performed for each class of the vehicle including bus, car, motorcycle, pickup, truck, and van in a predefined region of interest (ROI). Our proposed system is designed to handle noisy and low contrast traffic videos of top and side view of the vehicles. In particular, side view videos show the occlusion of the two directional lanes which lead to vehicle occlusion problems. For object detection and classification, YOLOX shows promising mean average precision (mAP) approximately at 77.2 and 58.8 percent for top and side views, respectively, which outperforms YOLOv3 for both top and side view datasets. The counting results show that ByteTrack with YOLOX can handle vehicle occlusion which occurs in the side view videos. Counting the vehicles within an ROI can reduce fault detections for both top view and side view videos.