Vehicle detection in Indian traffic using an anchor-free object detector

Abstract

Traffic camera video feeds are helpful in implementing intelligent vehicle detection and classification (IVDC). It has various applications in the transportation engineering domain, such as queue length estimation, vehicle tracking, traffic parameters estimation etc. However, in the Indian traffic wide variety of vehicles (motorbikes, auto-rickshaws, cycle-rickshaws, minitrucks, trucks etc.) travel on the road. They do not follow lane disciplined and occluded each other, making vehicle detection very challenging. This work presented an anchor-free object detection model (YOLOX) on the Indian traffic dataset (ITD) and compared it with the existing object detection models. It achieves 88% mean average precision (mAP) and 37 frames per second (FPS) on ITD.