Unsupervised Domain-Adaptive Person Re-identification Based On Deep Learning

In recent years,pedestrian re-identification has become a key technology in the field of video security and surveillance,attracting widespread attention from both industry and academia.Due to the small scale of publicly available labeled pedestrian datasets,there is a bottleneck in the research of pedestrian re-identification in the supervised domain,and the performance is not satisfactory in cross-domain problems.In comparison,unsupervised domain adaptation pedestrian re-identification can significantly reduce the dependence on labeled data and enhance the model’s generalization ability in cross-domain problems.However,due to factors such as pedestrian posture,illumination,and camera imaging differences,images of different pedestrians may have a high similarity,while images of the same pedestrian under different cameras may have significant differences.The challenge lies in how to better distinguish different pedestrians with similar appearances in the inter-class dimension and recognize the same pedestrian with significant imaging differences in the intraclass dimension.Therefore,this thesis focuses on unsupervised domain adaptation pedestrian re-identification from both inter-class and intra-class dimensions.The main research content and innovations are as follows:(1)Unsupervised domain adaptation pedestrian re-identification based on occlusion of key areas.To address the issue of blurred boundaries between pedestrian classes,we propose a cluster-level contrastive learning framework,using the stability of pedestrian category center features to reduce the impact of clustering errors on model training.Moreover,considering that introducing hard samples in contrastive learning can make the model focus more on the edge areas of pedestrian categories,we design an Attention-Based Occlusion Module(ABOM)with plug-and-play capability,using it to generate hard positive sample of source domain data,to fully mine the supervision information in the source domain data,and to gradually transfer the ability to distinguish similar samples to the target domain through iterative training.(2)Unsupervised domain adaptation pedestrian re-identification based on camera perception.To address the issue of significant intra-class differences and reduce the imaging bias of the same pedestrian under different cameras,we propose a new Camera Consistency Metric(CCM)to more accurately represent the style differences between camera domains.By using CCM to recalculate the feature distance between pedestrians,we can reduce the impact of camera domain bias on clustering results and bring the features of the same category of pedestrians closer together.In order to further cover the effective information in hard samples,we improve the memory update strategy in Cluster Contrast by using hard samples to update memory features,enriching the feature representation within pedestrian classes and improving the model’s recognition accuracy for the same pedestrian under different cameras.