Research On Unsupervised Cross-Domain Person Re-Identification Based On Clustering

With the development of science and technology,video surveillance has been widely used in many fields.Video surveillance can not only protect the legitimate rights and interests of the public and prevent illegal crimes,but also play an important role in criminal investigation,finding evidence,and tracking missing persons.Person re-Identification(re-ID)is one of the important technologies used to realize these application scenarios.Person re-ID uses the existing image or video data of a target person to find out this person in other videos or images.Existing person re-ID technologies mainly include Supervised Person re-Identification and Unsupervised Person re-Identification.Supervised person re-identification is mainly carried out under the labeled dataset and achieves very high accuracy.However,annotating datasets requires a large amount of cost,so studying unsupervised person re-identification is becoming increasingly important for practical applications.Unsupervised person re-identification technology is carried out in the case that the source dataset is labeled and the target dataset is unlabeled.The two datasets are usually different.Therefore,unsupervised person re-identification is considered to be a special problem of the unsupervised domain adaptation(UDA).Two unsupervised cross-domain person re-identification methods are proposed in this thesis.The works are summarized as follows.(1)Unsupervised cross-domain person re-identification by deep clustering and instance learning(DCIL)Firstly,we carry out camera style migration for the target data.The final target dataset contains the Cam Style images and original target images.Then,the backbone network module extracts the features of source data and target data.The features of the two domains are used to train the model.The source images are input into the source domain branch network,the classification layer is used to predict the categories,and the cross-entropy loss function is used for optimization.At the same time,the features of the target data are inputted into the target domain branch network.We use DBSCAN clustering algorithm to obtain the pseudo labels of the images.The pseudo labels and features vector of the image are stored in the sample memory module.We use the clustering repelled loss function to optimize the module.Finally,the superior performance of this method is proved by experiments on two public datasets,DukeMTMC and Market1501.(2)Unsupervised cross-domain person re-identification by deep clustering and attention mechanism(DCAM)On the basis of the first work,in order to learn the more discriminative features of person and ignore the information differences between different domains,this thesis proposes DCAM method.In the same way,the target dataset is expanded with Cam Style images.The features are extracted using the backbone network.Then,the extracted features are inputted into the attention module to learn the features that contains more person features and ignore background features.Next,the features of the source data are inputted into the classification layer.We use the cross-entropy loss function with label smoothing to optimize the module.At the same time,the features of the target data are inputted into the K-means++ clustering module to obtain the pseudo labels.We use the Harder Mining triplet loss to optimize the network.Simultaneously,the maximum mean difference(MMD)between source domain and target domain is calculated as the loss function.Finally,the superior performance of DCAM method is proved by experiments on two public datasets,DukeMTMC and Market1501.