Research And Implementation Of Target Detection And Tracking Based On Roadside Visible Light Images In Vehicle-road Collaboratio

Roadside perception in V2X(vehicle to everything)can reduce the hardware cost of single vehicle,increase the vision field of vehicle perception,make closed-loop policy with high-level automatic driving,and provide global optimal collaborative control capabilities for traffic participants and managers,which has important research meanings and broad application prospects.Compared with car scene,background area of roadside scene is often fixed and the perception ability is stronger.Compared with other sensors,visible images captured by camera have higher resolution and provide more object semantic information.Therefore,this paper focuses on the research of object detection and tracking based on roadside visible image in V2X.The main work and innovations are as follows:(1)A roadside visible image data collection scheme is designed,and the collected images are annotated with pedestrians,vehicles and other categories.A roadside object detection and multi-object tracking dataset is constructed.(2)This paper proposes an improved YOLOX roadside object detection method based on attention and focal loss.SE(Squeeze and Excitation)module and CBAM(Convolutional Block Attention Module)module are added to the FPN(Feature Pyramid Network)features to strengthen the channel and spatial relationship.Focal loss is used to replace the cross entropy loss of the object branch in order to solve the problem of imbalance between positive and negative samples.Experiments show that the proposed method can effectively improve the detection accuracy of roadside objects meanwhile the number of parameters and computation are barely changed.(3)This paper proposes a roadside object detection method based on background segmentation and multi-task learning.Due to the unchanged background of roadside scene,Gaussian mixture model is used for background modeling.The result is fused with annotated bounding boxes to remove noise and obtain the foreground image.With multi-task learning mechanism,a segmentation module is added and the foreground image is used as supervision target to train the network to automatically generate foreground background segmentation map.Non-local module is used as a self-attention mechanism,combined with segmentation map to reduce the response of background area and increase the response of foreground target.Experiments show that the proposed method can generate segmentation map in real time,and suppress background noise,strengthen features of interested foreground target,thus further improve the detection accuracy of roadside objects.(4)This paper proposes a roadside object tracking method based on ReID(re-identification)features and joint detection.A ReID branch is added to the detection head to extract ReID features as appearance information in order to improve the robustness of tracking.With the fusion of the mahalanobis distance computed on bounding boxes and the cosine distance computed on ReID features as similarity matrix,a two-stage matching strategy is used to associate objects and trajectories.A joint detection and tracking process is given.Experiments show that the proposed method can improve the detection accuracy,refine the tracking result during occlusion,thus realize stable tracking of roadside objects.