Research On Domain Adaptive Person Re-ID Algorithm For Surveillance Scene

With the implementation of security projects such as "safe city" and "three-dimensional urban security control",the importance of pedestrian re-identification as a core technology of intelligent security is increasing in daily life.It is widely used in the fields of public security,criminal investigation and intelligent business,and is one of the research hotspots in the direction of artificial intelligence.Current pedestrian re-id techniques with supervised scenes have achieved extremely high accuracy,however,in practical applications,the requirement that the dataset belongs to the same camera network greatly limits the generalisation capability of the network.Therefore,domain adaptive Re-ID that migrates networks trained on the source domain to the target domain has become a focus of current research.However,data captured by different cameras with different backgrounds,resolutions and illuminations lead to blurred extraction of pedestrian representations and incorrect identity classification.In addition,there are large style differences and sharp distribution of intra-domain feature differences between the trained source and target domains,leading to insufficient ability of the network to extract multi-domain features.To address the above problems,this thesis conducts an in-depth study.The main work is as follows:(1)In order to better mine the salient features of different channels while avoiding pedestrian identity overfitting,this thesis proposes a Person Re-identification Network based on a Multi-dimensional Attention Mechanism(MDAN).Specifically,MDAN consists of a backbone network,a Multi-dimensional Attention Block(MAB)and an Optimisation Loss Module(BIM).MAB is used to guide the backbone network to extract salient spatial information in different channels,while BIM avoids overfitting of pedestrian identities by reducing constraints on the Re-ID loss function for network training.With the addition of these two modules,the training process of the network learns more effective pedestrian representation information and speeds up the convergence of the loss function,increasing the training speed of the network.(2)In order to mitigate the significant distribution differences that exist in both crossdomain and intra-domain features,this thesis proposes a Multi-Domain Feature Differentiation based Domain Adaptive Network(MDFUDA-Net).Specifically,the MDFUDA-Net architecture consists of a backbone network,an Intra-Domain Normalization Block(IDNB),and an Inter-Domain Feature Fusion Module(IFMB).The IDNB aims to enhance the network’s feature extraction capability for different domains by removing domain-specific style features while also attempting to retain the information lost due to the instance normalization module.The IFMB aims to reduce the performance impact caused by domain gaps by achieving deeper domain fusion of source and target domain features.By leveraging these two techniques,MDFUDA-Net effectively weakens both inter-domain and intra-domain style differences,resulting in enhanced network transferability.(3)To validate the efficacy and superiority of MDAN and MDFUDA-Net,this thesis conducted extensive experimental comparisons and ablation studies on three publicly available datasets,namely Market-1501,Duke MTMC-Re ID,and MSMT17.The experimental results show that the MDAN and MDFUDA-Net designed in this paper achieve high performance in both m AP and Rank-1 accuracy compared to current domain adaptive pedestrian reidentification methods.