Research And Implementation Of Cross Domain Person Re-Identification Algorithm Based On Deep Learning

With the development of intelligent security,smart city and other fields,person re-identification technology has attracted more and more attention.At present,most person re-identification methods use labeled datasets for training.These methods need a lot of manual annotation and lack expansion ability,so their performance will be significantly reduced on other datasets with domain gap.In real life,person will appear in cross domain scenes,so the research on cross domain person re-identification method is of great value.In this thesis,attention mechanism and pose estimation are used to improve the performance of cross domain person re-identification methods.The main contributions are concluded as follows:(1)To solve the problem that the backbone network lacks feature selection ability in cross domain person re-identification,the thesis proposes Multi-branch Selective Kernel Networks(MSKNet)based on multi-scale weight fusion.The thesis adds multi-scale pooling branches in the feature selection stage of Selective Kernel Networks(SKNet)to obtain detailed information and make the attention weight more reasonable.MSKNet has receptive field branches with different scales and shapes.And MSKNet adds multi-scale pooled branches in the selection stage of different receptive field to obtain detailed information and make the distribution of attention weight more reasonable.Then,MSKNet and Non-local Block(NL)are embedded into Res Ne Xt50 to design a more effective backbone network.MSKNet improves the ability to select features of different receptive fields.Non-local Block improves the ability to capture long-range feature dependence.Through the analysis of experimental results,it is verified that the backbone network proposed in this thesis has a good effect.(2)Considering the changeable background interference and occlusion in cross domain person re-identification,Multi-branch Pose-guided Networks(MPOSENet)is designed to make the model focus on human regional features.The key points of human posture are obtained by pose estimation,and then features guided by pose estimation are obtained.The irrelevant features such as background and occlusion are filtered.MPOSENet designs the partial feature branch guided by pose estimation,the fusion feature branch of global features and partial features guided by pose estimation,the upper partial feature branch guided by pose estimation and the bottom partial feature branch guided by pose estimation.Through multi-branch network,it not only captures the local detail information,but also pays attention to the information in the global features.The multi-branch network makes the features extracted by the model more comprehensive and effective.Through several groups of experiments,it is verified that MPOSENet can promote the generalization ability of the model.(3)In order to give play to the practical value of the model,the thesis designs a person re-identification system taking the cross domain person re-identification algorithm proposed above as the core.The system has been developed and tested.