Research On Key Technologies Of Multi-view Person Re-identification

With the increasingly serious safety problems of autonomous vehicles,the intelligent vehicle infrastructure cooperative system has received extensive attention from academia.Person re-identification(Re-ID)has attracted much attention as one of the core technologies of the intelligent vehicle infrastructure cooperative system.Existing multiview Re-ID methods are mostly supervised.And main problems of them are poor crossdomain performance,high production costs and excessive reliance on training sets.The emergence of unsupervised Re-ID has made up for these shortcomings.Considering the problems of current unsupervised Re-ID,such as low accuracy,insufficient utilization of data,poor realtime and so on,this thesis proposes a improved unsupervised Re-ID framework.Considering the problem of low accuracy and insufficient utilization of data in vehicle infrastructure cooperative scenarios,a improved unsupervised Re-ID framework is proposed.In order to increase the initial accuracy of the visual classifier,an improved visual classifier model is proposed,which is improved on feature extraction,metric learning and person matching.In order to improve the problem of insufficient data mining,utilize camera number and shooting time to establish a multi-camera topology estimation model to provide semantic information for the visual classifier,and enhance the discrimination ability of the visual classifier.Finally,in order to improve the problem of insufficient utilization of data,utilize ranking results to establish the incremental optimization model based on the hierarchical fusion,so that the visual classifier can be incrementally optimized.Experiments validate the effectiveness of the proposed method.Considering that the proposed model has a poor performance on realtime,a new structure of visual classifier is proposed.Through analysis,it is found that the calculation time mainly comes from the comparison process.Based on this,a visual classifier based on Q-learning is proposed.Although the realtime problem has been solved to a certain extent,it also causes the significant drop in accuracy.For this reason,the person attributes are introduced to the existing feature extraction network to establish a new feature extraction network based on person attributes and features.This network can provide visual classifier with more discernible details and improve the problem of greatly reduced accuracy.The experimental results show that,the optimized model achieves the goal of realtime processing while guaranteeing the Rank-1 accuracy.