Research On Unsupervised Domain Adaptive Pedestrian Re-identification Based On Multi-label

Pedestrian re-identification is a hot issue in the field of computer vision in recent years.Its task is to retrieve pedestrians from multiple surveillance cameras in non overlapping areas.The pedestrian re-identification algorithm based on supervised learning achieves good performance on single domain data sets,but it performs generally on cross domain data sets.This means that in the new scenario,the method based on supervised learning needs a lot of manpower to annotate the data.In order to solve this problem,people began to study unsupervised pedestrian re-identification algorithm.Methods based on unsupervised learning can be divided into two categories: totally unsupervised learning and unsupervised domain adaptation.Compared with the unsupervised learning method which only uses the data of the target data set,the unsupervised domain adaptation method uses an auxiliary data set for domain adaptation,which has more practical significance.Therefore,this paper mainly studies the unsupervised domain adaptation(UDA).At present,most unsupervised domain adaptation methods only use a certain kind of loss or pseudo label generated by a certain method as the supervision information in the target domain to train the network,which fails to make full use of the supervision information,which makes the performance of the model general.Therefore,this paper proposes a multi label fusion algorithm(MLF)which combines clustering labels and soft labels.The labels generated by clustering are used to construct triplet loss and soft labels are used to construct cross entropy loss.In the aspect of soft label design,this paper designs a soft label,which allocates the probability value by calculating the distance between the features of each image in the target data set(domain)and the average pedestrian feature in the source data set(domain),so that the network has a certain fault tolerance ability.The experimental results show that the m AP of the multi label fusion method on the data sets Duke MTMC-reID?Market-1501 and Market-1501?Duke MTMC-reID is increased by 1.22% and 2.29%,and Rank1 is increased by 1.45% and 1.89%,respectively.This verifies the effectiveness and feasibility of the multi label fusion algorithm.Furthermore,considering that the multi label fusion algorithm only uses the last layer of network features to estimate the label,and there is a problem that the misestimated labels lead to the misclassification of the model,this paper also proposes an unsupervised domain adaptive pedestrian re-identification algorithm based on multi-stage and multi label,and designs a stage block with dual pooling.The multi-stage and multi label algorithm adds feature expression information and network restriction information,which changes the problem of adding restriction information only at the last layer of the network,and alleviates the problem of model deviation in the wrong direction due to wrong labels.The experimental results show that compared with the single stage block and global generalized average pooling method,the m AP of the multi stage multi label algorithm on the dataset Market-1501?Duke MTMC-reID and Duke MTMC-reID?Market-1501 are increased by8.37% and 17.98%,respectively,and the Rank1 is increased by 5.16% and 12.02%,respectively.The experimental results show that the accuracy of multi stage and multi label algorithm reaches the advanced level.