A Computationally Efficient Real-Time Vehicle and Speed Detection System for Video Traffic Surveillance

Abstract

This research article presents a computationally efficient real-time vehicle and speed detection method for video traffic surveillance. It applies training and testing phases in parallel over pre-trained video surveillance dataset by employing YOLO algorithm. In order to enhance the computation time, dynamic background subtraction method has been applied in conjunction with YOLO algorithm. The aim is to reduce the total number of frames required to be considered for further processing. DBS has strong adaptability to changed background and can detect the movement in frame. In addition to vehicle detection, the proposed algorithm can detect the speed of detected vehicles using a centroid feature extraction method. Our proposed method is a dual combination of speed and accuracy. The performance of the proposed algorithm is analyzed over 2018 NVIDIA AI City challenge dataset and a comparative analysis is done with various state-of-the-art vehicle detection methods (CNN, R-CNN, Fast R-CNN, and FastYOLO).