Research On Unsupervised Person Re-identification Method Based On Contrastive Learnin

Person re-identification refers to the identification and matching of the same person under different surveillance camera perspectives.This technology has broad application prospects and important research significance in the fields of public security,intelligent transportation,and marketing.In recent years,supervised person reidentification has made remarkable progress,but its reliance on labeled data limits the scalability of this technique for practical applications.In order to facilitate the widespread use of person re-identification techniques in practical applications,unsupervised person re-identification techniques that do not require labeled data have been proposed and have received increasing attention.In the field of unsupervised person re-identification,cluster-based methods have shown advanced performance,but the noisy pseudo-labels existing in the clustering process limit the further improvement of model performance.In addition,unsupervised methods are susceptible to background noise and loss of local details in deep features due to the lack of effective supervision from ground-truth labels.In response to the above problems,this thesis does the following research based on contrastive learning:⑴ In order to make full use of the relationship between global features and local features and further improve the quality of pseudo-labels,hybrid contrastive learning algorithm for unsupervised person re-identification is proposed in this thesis.In this method,the progressive feature compensation module compensates for the missing detail information in the deep features through detail guidance,so as to better identify different person with similar appearance.The pseudo-label refinement strategy based on multi-scale features considers the relationship between global features and local features,and selects reliable clusters by measuring the consistency of pseudo-labels among different feature spaces.Finally,multi-scale hybrid memory is used to store cluster-level and instance-level multi-scale features for hybrid contrastive learning to facilitate the model to mine discriminative information of person.⑵ In order to suppress the interference of background noise and generate reliable pseudo-labels,continuous epoch distance integration algorithm for unsupervised person re-identification is proposed in this thesis.In this method,the pseudo-label generation strategy of continuous epoch distance integration transfers the distance matrix of the previous epoch to the current epoch,and performs weighted fusion on the distance matrices of two continuous epochs,and uses the new distance matrix obtained after weighted fusion to generate a reliable pseudo-label.While eliminating noisy pseudo-labels,an attention module of spatial structure and channel dimension is proposed,which weights each element in the feature map to achieve the purpose of enhancing discriminative regions and suppressing background noise.