Research On Multi-target Detection And Tracking Technology For Traffic Scenes Based On DETR Network

In the traffic monitoring system,vehicle detection and tracking results which are accurate and real-time can provide important data support for traffic management.There are factors such as fast vehicle movement speed,large vehicle scale changes,and occlusion in the video,so there are many problems in the vehicle detection and tracking results in highway scenes such as low accuracy and slow speed.In response to the above questions,this paper constructs a traffic target detection and tracking dataset based on the video data of highways and tunnels.This paper proposes a tracking algorithm based on DETR and an integrated detection and tracking network,which improve the speed and accuracy of vehicle detection and tracking.The main research contents of the paper are as follows:(1)Construct a traffic object detection and tracking dataset.The basis of this paper is surveillance video on highways and tunnels.In this paper,the video screening strategy,the labeling rules of traffic objects,and the classification method based on vehicle features are designed.This paper constructs a traffic data set in highway and tunnel scenarios and expands the data set through data augmentation.Compared with the current public target detection and tracking datasets,this dataset focuses on highway and tunnel scenarios.The data has the characteristics of rich data,complex and changeable scenarios,and strong practical value.(2)Design a tracking algorithm based on DETR.Aiming at the problem that traffic targets are prone to occlusion during tracking,this paper adopts the strategy of matching low-resolution confidence frames in the Byte Track algorithm,combined with the DETR object detection network,to realize the continuous tracking of vehicle targets.The experimental results in the dataset show that the algorithm has a MOTA of 84.82% and an FPS of 25.58,and can handle common target occlusion problems with good performance.(3)Design an integrated detection and tracking network based on DETR.Aiming at the problem of error accumulation in the detection-based tracking algorithm,based on the traffic target detection and tracking dataset,this paper constructs a DETR-based detection and tracking integrated network.The algorithm is based on the DETR network,and a time series 3D convolutional layer is designed to obtain feature and time series information.In order to deal with the problem of complete occlusion of the target in a short period,the network improves the Hungarian algorithm to achieve target matching and data association.The experimental results show that compared with the DETR-based tracking algorithm,the MOTA and MOTP of the DETR-based detection and tracking integrated network are 86.30% and 86.02%,which are1.48% and 1.39%,respectively.This paper studies the multi-target detection and tracking technology based on the DETR network,which can be better applied to vehicle target detection and tracking in actual traffic scenes.This technology can lay an important data foundation for traffic parameter extraction,traffic incident detection,and vehicle operating status analysis in real traffic scenarios.