| The purpose of person re-identification(Re-ID)is to retrieve the target images belonging to the same identity from different camera views,which can be widely used in social security and other fields.In the field of deep learning,supervised person re-identification tasks have achieved high accuracy by making full use of labeled person image information.However,the expensive labeling cost brought by the continuous growth of monitoring data makes the development of efficient and robust unsupervised person re-identification systems have great appeal in the academic community.Most existing unsupervised person re-identification methods usually use a clustering-based pipeline to generate pseudo-labels for supervised training.However,under such a framework,on the one hand,due to the differences between cameras in camera parameters,shooting angles,lighting conditions,pedestrian background,etc.,the same pedestrian image captured by different cameras may look very different,which will directly affect the training of the model due to the camera variance caused by the camera variation;On the other hand,in the unsupervised person re-identification task,the quality of pedestrian feature extracted by the model directly affects the performance of the model.Most of the existing methods only use a single clustering feature for contrastive learning,and the single clustering feature is often not able to fully mine the robustness features of the target pedestrian,which will lead to poor representation of pedestrian features extracted by the model and the robustness of the model is poor.To solve these problems,this thesis proposes an unsupervised person re-identification method based on camera-irrelevant feature consistent learning and an unsupervised person re-identification method based on multi-level feature contrastive learning.The main research contents and contributions of this thesis are as follows:1.To solve the problem that camera variance affects model training,this thesis proposes a camera-irrelevant feature learning framework.The query instances are compared with cluster features in the form of contrastive learning,and the model is gradually trained through network updating and clustering iteration.Specifically,in the training process,sampling images of other camera styles to be added to the training can help the model learn camera-irrelevvant pedestrian features,extract more robust pedestrian features,and enrich the cross-camera learning.In the clustering process,the camera-irrelevant distance matrix is calculated,and more accurate pseudo-labels are generated based on the matrix clustering,which makes it easier to match the images of the same pedestrian taken by different cameras to the same cluster,thus reducing the intra-class variance caused by camera variance.Demonstrate the effectiveness of the proposed method on a widely used person re-identification dataset.2.In view of the problem that single clustering feature often can not fully mine the robustness characteristics of target pedestrians,this thesis proposes an unsupervised person re-identification framework based on multi level feature contrastive learning.This framework comprehensively considers the cluster level and instance level feature information,and combines the farthest sample contrastive learning loss of intra-class relationship,the farthest sample contrastive learning loss of inter-class relationship and the cluster centroid contrastive learning loss for unsupervised person re-identification,so that the model can fully mine more robustness features of target pedestrians;At the same time,combined with the training based on the generated data image as the data enhancement data set input into the network,increase the diversity of samples and improve the robustness of the model.Experiments on multiple re-identification datasets verify the effectiveness of the proposed method. |
