Research On Key Technologies Of Person Re-identification Based On Privacy Protection And Fine-grained Feature Fusion

With the progress of society and the continuous development of science and technology,the application of video surveillance systems in the field of urban management,public safety and traffic has gradually become an indispensable part.Person re-identification technology as an important part of the video surveillance system,how to efficiently and accurately carry out largescale pedestrian retrieval has gradually become a concern.In the development of person reidentification technology,the privacy protection issue is of great concern.How to improve the retrieval accuracy while effectively protecting individual privacy has become a core issue that needs to be urgently addressed in research and practice.Although scholars have proposed many effective person re-identification methods,the problem of person re-identification for open environments is still not adequately addressed.The main challenges faced by existing person re-identification techniques are:(1)In large-scale person re-identification datasets with attribute annotations,incomplete pedestrian attribute annotations lead to a bottleneck in model performance,and the separation of the existing camera style augmentation and person re-identification process leads to insufficient adaptation of the model to the differences in camera styles,which affects the accuracy and robustness of person re-identification.(2)Some approaches based on supervised learning for person re-identification are constrained by the need for a large amount of labeled data,thus limiting their applicability in large-scale person re-identification,and thus domain adaptation is becoming popular.However,in the process of domain adaptive person re-identification,the source domain labeled sample data occupies an important position,which also raises the issue of pedestrian privacy protection.(3)Source free domain adaptive person re-identification utilizes source pre-trained models to achieve adaptation to the target domain,and these methods do not need to access the source data,which solves the pedestrian privacy problem to a certain extent.However,the absence of source data leads to domain disparities that are challenging to effectively quantify and mitigate.Consequently,source free domain adaptive person re-identification in open environments encounters significant challenge.(4)For source free domain adaptive person re-identification,the source data privacy is protected to some extent,but the target data also has the challenge of privacy leakage.Due to the specificity of pedestrian identity information,the full-data privacy protection problem becomes a key issue to be solved in person re-identification.Aiming at the problems of supervised learning,conventional domain adaptation and source free domain adaptation for person re-identification,this thesis carries out the research of key technologies of person re-identification based on privacy preservation and fine-grained feature fusion on how to carry out person re-identification efficiently and safely,and the main results are summarized as follows:(1)Aiming at the problem of incomplete attribute annotation and the separation of camera style augmentation and person re-identification process leading to model performance bottleneck,this thesis proposes an attribute-aware style adaptation for person re-identification.The method takes advantage of the complementary nature of camera style information and attribute annotation information by converting attribute features with commonality into uniform factors and developing an attribute-aware module to associate image style generation.Experimental results on the person re-identification dataset with incomplete attribute annotation show that the method has high person re-identification performance and attribute recognition accuracy.(2)Aiming at the privacy protection problem caused by the need to access the source domain data during model training in the domain adaptive person re-identification,this thesis proposes a Source-free Style-diversity Adversarial Domain Adaptation with Privacy-preservation(S2ADAP).The method employs a domain style diversity enhancement module based on a generative adversarial network to address the problem of inter-domain pedestrian appearance style differences,while the challenge of intra-domain personalized style misalignment is addressed by an adversarial mutual mean-teaching model.Experimental results show that without accessing any source domain data,the method both protects source domain data privacy and achieves high person re-identification performance.(3)Aiming at divergence-agnostic problem due to the lack of source data in the source free domain adaptation,this thesis proposes an Instance-level Adversarial Mutual Teaching(IAMT)framework.The method utilizes a variance-based division module for instance-level separation including source-similar and source-dissimilar subsets to implicitly measure the domain divergence,and introduces a dynamic adversarial alignment strategy to narrow the domain differences by using an adversarial instance confusion discriminator for the target domain data.Experimental results show that the method achieves domain adaptation while considering source data privacy.(4)Aiming at the problem that the source-free domain adaptive method only protects the privacy of source data while ignoring the privacy of target data,this thesis proposes a full-data privacy-preserving domain adaptive(Secure DA)method based on adversarial attack.The method utilizes a fine-grained multi-view adversarial attack to encrypt pedestrian images in the target domain,and achieves full-data privacy protection by transferring the original samples to source-style adversarial samples with quasi-imperceptible perturbations prior to publication through reliable complementary relationships of pre-trained source models.Experimental results show that the method can achieve full-data privacy preservation and obtains optimal person reidentification performance.