Optimized YOLOv2 based vehicle classification and tracking for intelligent transportation system

Abstract

The prevailing real-time system used for vehicle detection and classification using deep learning techniques accuracy diminishes due to its background, illumination variation, occlusion, and variation of vehicle sizes in a scene. Hence the real-time system is proposed to enhance the accuracy level on detection and classification of vehicles for a multi-view surveillance video using an optimized YOLOv2 deep learning algorithm. The improvisation in optimized YOLOv2 has two major stages, firstly to eliminate the influence of background and illumination variation on Focal Loss is improved. Secondly, Rk-means++ clustering discovers the best prior, and normalization to bounding boxes loss to find length and width. Additionally, the network is made more accurate in detecting and classifying vehicles. The experimental results are tested on the COCO-2017 dataset and show that the mean Average Precision (mAP) is 97.85% underneath the premise of substantiating real-time performance. Since there exists a pre-trained model and GPU increases the speed processing to 40 frames per second.