Research On Multi-Camera Vehicle Recognition And Tracking Technology Under Overlapping Vision

Multi-camera multi-vehicle tracking is a key problem in intelligent transportation system(ITS).The problem can be divided into two parts:the tracking problem in overlapping view and the tracking problem in non overlapping view.This paper proposes a solution for the former.The scheme is mainly divided into two parts,single camera tracking(SCT)and multi-camera multi-target tracking(MTMC).In the aspect of single camera tracking,two key steps are target location and tracking.In the target location,this topic uses two algorithms:3D target detection algorithm and mapping from 3D detection frame to GPS coordinates.Then the single camera tracking is realized by combining the output GPS features of target positioning with the image-based features.Vehicle re-identification algorithm is the key to realize multi-camera tracking.This paper proposes a vehicle re-identification method based on image instance segmentation and metric learning,which effectively solves the problem of occlusion between vehicles and realizes multi-camera re-identification of vehicles.In this paper,the high-precision instance segmentation mask of the vehicle is obtained through the segmentation model,the vehicle is cut out from the original image,and the part irrelevant to the target vehicle is removed,and then the image features are extracted by inputting the vehicle re-identification model.For the vehicle re-identification model,this paper uses a baseline model in the field of pedestrian re-identification.According to the characteristics of road scene and vehicle target,this paper proposes a modification scheme,and proves its effectiveness through experiments.In the aspect of multi-camera multi-target tracking,this paper designs a multi-camera multi-vehicle tracking system based on spatiotemporal consistency and multi-level matching.The method integrates the re-identification feature,space-time feature and GPS trajectory feature,realizes the hierarchical matching of single camera tracking trajectory,and completes the continuous tracking of vehicles across cameras.The effectiveness of the multi-camera multi-vehicle tracking system is proved by experiments.Finally,combined with the experimental data set,the aerial view of the intersection is drawn,and the trajectory of the target is mapped to the aerial view for visualization.This operation can improve the efficiency of trajectory matching and the accuracy of target location.Experiments show that the proposed multi-camera multi-vehicle re-identification and tracking system is effective.