Research On Vehicle Detection And Tracking Method Based On Video

Obtaining accurate and effective vehicle location and trajectory information has become an important task in the extraction of traffic information since the vehicles are the major players in traffic behavior.The wide application of traffic monitoring equipment provides a good platform for the application of computer vision under complex traffic scenes.This paper proposes a vehicle detection and tracking algorithm based on Faster R-CNN and adaptive Kernelized Correlation Filters through the research of traffic video processing technology based on Convolutional Neural Network,which improves the efficiency of vehicle detection and tracking.The main contents of this paper are R-CNN as follows.(1)The basic model of video-based vehicle detection and tracking in city roads is proposed based on image characteristics and processing requirements of traffic scenes through the analysis of detection algorithm based on convolutional neural network and common tracking algorithms.(2)The vehicle target tracking initialization system based on Faster R-CNN vehicle detection is proposed.Then,the formatted KITTI vehicle datasets are trained and verified.The network infrastructure that meets the needs of real-time tracking initialization in the city road environment is selected through the comparative analysis of the accuracy and speed of the different network.(3)An adaptive model updating single-target kernelized correlation filter algorithm is adopted.The result of the deep learning detection algorithm is used as initialization and kernelized correlation filter tracers are generated.Then,a multi-targets tracking algorithm for complex traffic scenes is designed by combining spatial location data association algorithms,which can overcome the problem of model update accumulation error largely when the target is occluded by the traditional algorithm.The traffic videos under the conditions of sunny,fog and night on the city roads are collected in this paper,and the data sets are created by interval frame extraction from the videos for the training and testing of the Faster R-CNN detection network.Finally,the trajectory information of the vehicle is obtained with the tracking algorithm and fusion detection results.The method in this paper increases the distance accuracy of 14.7% by day and 12.9% by night with the error threshold of 20 pixels and the curve area of the success rate is increased by 18.0% and 12.5% compared to traditional KCF in the environment of the Caffe framework and Opencv 2.4.13 which built on the platform of HPZ640.The average processing speed reaches 37.2 fps for multi vehicle monitoring and tracking in sunny,haze and night conditions,which can achieve real-time multi-targets detection and tracking.