Research On Vehicle Object Detection And Cross-Camera Tracking In Traffic Scenes

The traffic scene monitoring video under multiple cameras is used to construct the vehicle running state of the whole scene,which can grasp the road traffic information and analyze the traffic situation in time.However,there are still problems in realizing accurate and fast vehicle detection and cross-camera tracking in traffic scenes,such as the high calculation amount caused by large numbers of video data,and the time error between cameras.This paper focuses on the detection and tracking of vehicle in traffic scenes,as well as the trajectory association of the same vehicle under multiple cameras,to obtain the complete trajectory and spatio-temporal map of vehicle operation,so as to analyze traffic parameters and events,and escort traffic safety.The main research content is divided into the following four points:(1)Research on vehicle object detection in traffic scenes.Based on the YOLOv4 object detection network with high accuracy and good real-time performance,this paper proposed an improved vehicle object detection strategy for the problems existing in traffic scenes.Secondly,for the three traffic scenes of high-speed day,high-speed night,and tunnel,a vehicle object dataset containing 45052 pictures was produced.Finally,the dataset was augmented and trained to get the object detection model.The experimental results showed that the m AP of the improved detection algorithm is 90% high-speed day scene,and the average detection speed is73 FPS.(2)Research on multi-object tracking in traffic scenes.According to the results of object detection,a object position prediction method based on camera calibration was proposed,which used speed to predict the possible position of the object.On this basis,a fast multi-target tracking method based on frame skipping detection was achieved.A multi-object tracking dataset is made for traffic scenes for experimental verification.The results showed that the vehicle object tracking was basically correct in the case of jumping 8 frames,the MOTA average value was higher than 85%,and the tracking speed reached 50 FPS,which significantly improved the tracking efficiency.(3)Research on cross-camera vehicle trajectory association.On the basis of the vehicle tracking trajectory under single camera,this paper proposed an association algorithm based on the linear fitting of the vehicle trajectory and the cosine similarity of the trajectory node images,aimed to complete the trajectory association of the same vehicle under cross-camera.Finally,experiments proved the effectiveness of the algorithm,which can associate multiple trajectories in multiple cameras at the same time to achieve continuous tracking of vehicles in a large range.(4)Traffic incident analysis based on trajectory data.Through the single-camera vehicle trajectory and the cross-camera trajectory association,a spatio-temporal map of the vehicle trajectory is generated,and a traffic video intelligent monitoring system is proposed to achieve the acquisition of traffic parameters and the detection of traffic incidents.Through experimental verification,it was found that the algorithm in this paper can accurately count traffic parameters,detect traffic events,complete the perception of vehicle operating status in traffic scenes,and provide a basis for traffic control and analysis.