Research On Person Re-identification Based On Unsupervised Domain Adaptation

With the continuous development of intelligent video surveillance,person re-identification has been widely concerned by academia and industry.Person re-identification aims to complete the task of associating person images with the same identity under different cameras.At present,most research work focuses on supervised scenarios given training data has identity labeled information.However,these methods rely on large amounts of expensive labeled data,and require that training and test data have to be drawn from the same camera network or the same dataset.When these person re-identification models are directly applied to different datasets,the performance will be severely degraded.This largely limits its generalization to real world tasks.Therefore,person re-identification introduces the unsupervised domain adaptation method by migrating the trained model from the labeled source domain to the unlabeled target domain.Compared with traditional unsupervised domain adaptation methods,unsupervised domain adaptation person re-identification is a more challenging open set problem.Source domain and target domain contain their own class space.The current methods based on feature distribution alignment are not suitable for the application scenarios of open sets.In addition,most of the methods only consider the distributed migration of the whole data domain,but ignore the changes in the target domain.Each data set is photographed by multiple cameras of different styles.Therefore,the data distribution varies largely due to camera viewpoint,illumination and background,which brings great challenges to cross-camera pedestrian retrieval in the target domain.This paper conducts in-depth research on the distribution migration between source domain and target domain and the drastic changes within target domain in unsupervised domain adaptation person re-identification.The main research contents of this paper are as follows:(1)An unsupervised domain adaptation algorithm for the loss of patch-based Feature Disentangling(PFD)based on part model is proposed.Firstly,Patch Generation Network(PGN)is introduced to generate discriminant patch features for each pedestrian image.For the unsupervised domain adaptation person re-identification task,the part model has more superiority than the whole image model.By pulling similar local patches closer,the potential discrimination information of local patches can be mined,and multiple local patches can be combined to identify the pedestrian identity.Then,PFD loss is proposed based on the principle of representation learning.It assumes that samples with closer feature distances share the same identity labels,and gradually analyzes positive and negative pedestrian samples with high confidence by constructing an entanglement space for each pedestrian local patch.Finally,the local patches with high similarity are pulled closer and the local patches with low similarity are pushed away in the feature space,so that the model can adapt to the target domain well without identity labels.(2)An unsupervised domain adaptation algorithm based on Tracklet and Camera Discrepancy Distribution Aligning(TC-DDA)is proposed to improve the data distribution migration between source and target domains.By transforming the common feature representation space into the discrepancy feature representation space,the alignment of the inter-domain discrepancy feature distribution is performed.Even though source domain and target domain are different in feature space due to different class space,the new discrepancy feature space only has intra-class and inter-class clusters.Since person re-identification is essentially a cross-camera pedestrian retrieval task,and both source and target domains contain multiple camera with different styles.Based on tracklet information and camera information,the distribution of differential features can be marked as intra-class/inter-class and intracamera/inter-camera distribution,and Maximum mean discrepancy(MMD)can be introduced to align the corresponding distribution of the source domain and the target domain.(3)An unsupervised domain adaptation algorithm based on Camera Aware Neighbor Mining(CANM)is proposed to alleviate the drastic variations in the target domain.In addition to the open set attribution,person re-identification dataset has another characteristic,which is the domain hierarchy.Both source domain and target domain can be further divided into multiple camera subdomains.The pedestrian similarity between cameras is much smaller than the pedestrian similarity within cameras.Therefore,Intra-camera matching is more likely to occupy the top of the ranking list,regardless of positive or negative samples.First,neighbor mining is constrained within and between cameras.In addition,due to the differences in camera styles,it is more difficult to search neighbors between cameras,and inter-camera neighbor mining is more likely to lead to biased searches than intra-camera neighbor mining.Therefore,the disentangling strategy is further utilized in inter-camera neighbor mining,and positive and negative samples of high confidence matching between cameras are mined to further optimize the matching results in the target domain.To sum up,PGN is used in this paper to propose PFD,TC-DDA and CANM algorithms to improve unsupervised domain adaptation person re-identification.Experimental results on multiple datasets verify the excellent performance of the proposed method.At the same time,the algorithm proposed can not only ensure the accuracy of recognition,but also meet the real world generalization requirements.