Research On Vehicle Detection And Tracking System Based On Video Stream

With the progress of social economy and the improvement of people’s living standards,the increase of the number of vehicles constantly promotes the development of road traffic,but it also faces new challenges,such as road blockage,frequent traffic accidents and serious pollution.However,how to build a detection and tracking system with fast detection speed,high accuracy and good robustness to achieve real-time detection and tracking of vehicles to obtain accurate traffic flow information has gradually become an important research topic.Based on this,this paper proposes a vehicle detection and tracking system based on video stream,which can more accurately and efficiently monitor the vehicles in the surveillance video and count the traffic flow information of various models.The research of this paper includes the following main aspects:(1)Introduced the principle of YOLOv5 s target detection algorithm in detail,and improved it on the basis of analyzing the characteristics of its model.Firstly,the Swin Transformer network structure is integrated into YOLOv5 s,and the backbone network of the original algorithm is reconstructed to better capture context information and improve the detection accuracy of the detector.Then,by comparing different attention mechanisms,GAM attention module is introduced into Neck network to enhance the cross-dimensional interaction between channels and spatial dimensions of information,reduce information loss and enhance network performance.Finally,the original non-maximum suppression algorithm NMS was modified to DIo U-NMS to improve the fine tuning and inference effect of the model.(2)DeepSort multi-object tracking algorithm was used for vehicle tracking.Since the original DeepSort algorithm is mostly used for pedestrian detection and tracking,and the object of this paper is vehicle target,the original size is not applicable.Therefore,the feature extraction network structure of the algorithm was optimized,and the input image size was modified to 128(H)×128(W).Then the detection results obtained by the improved YOLOv5 s algorithm were combined with DeepSort to realize the continuous tracking of the target and reduce the occurrence of target identity jump in the tracking process.Experiments show that the accuracy of the improved YOLOv5 s model combined with the DeepSort tracking algorithm is improved by 2.05 percentage points compared with the original algorithm.(3)Based on the improved YOLOv5s+DeepSort algorithm,an optimal virtual detection line is selected at the appropriate position of the road,and the vehicle impact detection line is used to complete the traffic flow statistics.The video image data set was constructed,and the vehicle data was divided into Car,Bus,SUV,Truck-I(less than three axles)and TruckII(more than three axles).The vehicles were detected,tracked and counted under different lighting environments.Experimental results show that the accuracy of statistics of total vehicle flow in different lighting environments such as daytime,evening and night is higher than that of the original algorithm model,reaching 97.8%,95.8% and 86.1%,respectively.(4)Designed a simple,efficient,easy to operate and easy to deploy detection and tracking system.The vehicle detection,tracking and traffic flow statistics are integrated into a system to achieve man-machine interaction,so as to better grasp the vehicle information and road conditions,so as to solve the practical application needs.