Research On Vehicle Target Detection And Tracking In Front Based On Deep Learning

With the increasing of domestic automobile ownership,the traffic safety problem of vehicle forward collision becomes more and more serious.Vehicle detection and tracking technology based on computer vision is one of the auxiliary technologies to prevent vehicle forward collision.It has become a hot spot and difficult point in vehicle anti-collision system to detect and track the vehicle in front of the road by using algorithm to process the image of the vehicle in front of the road quickly and accurately.The main points of this thesis are as follows:(1)Considering the limited computing resources of vehicular embedded devices,this paper makes a comparative analysis of YOLOV3 and YOLOV3-Tiny algorithms.The YOLOV3-Tiny detection algorithm with small number of parameters and high real-time performance is selected as the underlying module to design the vehicle target detection and tracking algorithm,which provides an assistant decision for the vehicle anti-collision system.(2)In this thesis,a lightweight Mobile Net network is used to optimize YOLOV3-Tiny to improve the accuracy of vehicle target detection on the premise of ensuring the detection speed.The Mobile Net V2 network combined with the backward residual structure and the deep separable convolutional network was used to replace the backbone network of Yolov3-Tiny.The network enhanced the capability of feature extraction while reducing the computational load of the network,so that the network could quickly and effectively extract the features of vehicle targets.The deep separable convolution is added into the model detection layer to further compress the model so as to achieve the purpose of efficient vehicle detection.The prior box selection was also optimized.K-means algorithm was used to re-cluster and the prior box size with a higher matching degree was selected to improve the detection ability of vehicle targets.(3)In this thesis,the optimized YOLOV3-Tiny target detection algorithm is combined with Deep Sort target tracking algorithm to design the vehicle target tracking algorithm in front of the road,which can effectively suppress the misdetection and missed detection of the upstream vehicle target detection.The Kalman filter is used to predict the vehicle motion state,and the Hungarian algorithm is used to carry out the optimal matching by combining the motion information and the improved vehicle appearance model information,and the matching results are optimized by cascading matching,so as to realize the first detection and then tracking of vehicle targets in front of the road.(4)The improved algorithm is trained and compared with the actual road vehicle data set.The experimental results show that the optimized vehicle target detection algorithm and tracking algorithm have good accuracy and real-time performance.The research work of this paper proves the feasibility of the optimized YOLOV3-Tiny target detection algorithm and Deep Sort target tracking algorithm,which provides a practical basis for vehicle collision prevention system.