Research On Cross-Camera Vehicle Re-identification Method Integrating Spatial-Temporal Information

Traffic surveillance cameras are everywhere in modern cities,so there is an urgent need for intelligent analysis of massive videos.Among them,vehicle re-identification and tracking have become research hotspots in the field of computer vision.Vehicle reidentification refers to retrieve the images of the same vehicle identity from multiple nonoverlapping cameras,and also can be viewed as a preprocessing step of cross-camera vehicle tracking.Since the region occlusion and visual similarities among different vehicles,it is difficult to stably distinguish their identities in the long term only by the license plates or visual features,and affect the subsequent cross-camera identity association.Therefore,this thesis focused on how to integrate the spatial-temporal information into the vehicle reidentification and tracking algorithms,and improve identity retrieval accuracy and identity association stability.The work of these methods can be helpful in the intelligent transportation system and smart city application.In general,this thesis systematically summarized the research advances of both vehicle re-identification and cross-camera tracking algorithms and outlined the main approaches of both tasks.Based on these analyses,this thesis proposed a vehicle re-identification algorithm that integrated the visual features and the spatial-temporal information.Furthermore,this thesis implemented cross-camera vehicle tracking based on the trajectory linking method.Finally,these two methods are evaluated on public datasets and campus videos.The main work and contributions could be summarized as follows:1.This thesis proposed a track clustering aided vehicle re-identification model,tc Re ID,that leverages both spatial-temporal information and visual similarities between image pairs.The visual stream uses the Res Net50 network as the deep learning backbone to evaluate the visual similarity;the track stream firstly clusters the image candidate set based on the spatial-temporal information,and then retrieves vehicle images according to the time feasibility in the clustered trajectories.Then,a joint metric is introduced to combine the results of two streams to get a unified score.Tests on the Ve Ri dataset showed the proposed tc Re ID model achieves 92.82% in m AP,outperforming the previous state-of-the-art methods.2.This thesis implemented a cross-camera vehicle tracking method in the campus videos.The idea of the track clustering is introduced into this method,and cross-camera tracking is realized by trajectory linking.In the AI City Challenge vehicle tracking competition which was held by the 2020 IEEE Conference on Computer Vision and Pattern Recognition(CVPR),this method ranked the 4th place.