Vehicle Detection And Tracking Based On Deep Learning

In recent years,with the continuous progress of computer hardware and the continuous improvement of computer computing power,the fields related to computer vision have developed rapidly.It has important application value in intelligent transportation,face recognition and intelligent driving.A direct result is that vehicle detection and tracking become more and more important in the field of transportation,It has received extensive attention.Vehicle detection is to detect the vehicle in the real-time detection image.Vehicle tracking is to solve the problem of target loss caused by frequent occlusion and light intensity change in target tracking by prediction based on vehicle detection.For the task of vehicle target detection and tracking,this paper was studied and analyzed the mainstream target detection algorithms in recent years.Then,a good performance yolov3 network model is deeply discussed,and its related network structure and some existing problems are analyzed in detail.On the basis of yolov3 network,this paper was proposed an improved yolov3 network based on deep learning,which is compared with the mainstream vehicle target tracking algorithm.Finally,based on the improved yolov3,combined with the vehicle target tracking algorithm,a complete vehicle target detection and tracking system is formed.The main contributions of this paper are as follows: 1)aiming at the problem of detection omission caused by detecting small targets in the original yoov3 network,this paper improves the part of darknet-53,the backbone network in yolov3,and puts forward new parameter settings and four branch output characteristic diagrams to improve the detection accuracy.2)Aiming at the problems of larger calculation,gradient elimination and gradient explosion caused by the larger improved model,this paper improves the residual network in the backbone network.By introducing the residual edge and fusing the residual branches,combined with the method of accumulating the residual units in the residual network once,the amount of calculation of the network is reduced.The improved algorithm and current typical target detection algorithm are evaluated on bdd100 k data set and Pascal voc-2007 data set,and PR curve,average detection accuracy and transmitted frames per second are used as evaluation criteria.In addition,the vehicle target detection and tracking system proposed in this paper is tested on mot16 data set,and compared and analyzed through visual effect display and running speed.The experimental results show that the improved yolov3 network based on deep learning proposed in this paper is better than the existing vehicle target detection algorithms on each data set.Compared with SSD and yolov3,the average detection accuracy is improved by 13.6% and 0.9% respectively.The network model in this paper has more advantages in detection effect,especially for small targets.