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

In recent years,with the popularity of video surveillance cameras,it has become a hot issue to help the development of video surveillance field with the help of artificial intelligence technology.The person re-identification(Re-ID)is a very important basic application research in the field of intelligent security monitoring.Its mission is to retrieve the pedestrian image with the same identity as the query image but taken by different cameras in the gallery.The person Re-ID has a wide range of application requirements in tracking crime,anti-terrorism and smart city.With the development of deep learning technology,the person Re-ID model with supervised learning method has achieved great performance on the labeled data set.However,supervised learning methods have high requirements for the quality and quantity of the manually labeled identity labels,which is too expensive in the real application scenarios.Therefore,the cross-domain transfer learning method based on self-supervised training is more and more concerned,which achieves relatively good retrieval effect by exploring the potential label estimation in the target domain.The keys of transfer learning method based on selfsupervised training are how to improve the accuracy and difference of pedestrian feature expression and the reliability of generated pseudo labels.Firstly,in order to enhance the accuracy and difference of person feature expression,this thesis proposes a metric learning method module based on high-order information statistics,which obtain the low-order features and high-order features of person image by calculating the high-order information statistics of person image coarse features.Based on the pedestrian features of different orders,a multi-level and efficient end-to-end pedestrian feature extraction network is designed,which is trained by metric learning to improve the discrimination ability of pedestrian feature expression.Secondly,in order to improve the reliability of pseudo label generation,this thesis proposes a hierarchical clustering guided Re-ID cross-domain deep learning algorithm,which uses the advantages of hierarchical clustering algorithm and the average distance measurement criterion to improve the accuracy of pedestrian feature clustering,so as to improve the reliability of pseudo label generation of person images in the target domain,and provide the basis for the joint supervised learning strategy to transfer the Re-ID model to the target domain.At the same time,in the process of self-supervised training,this thesis leverages the global and local information of pedestrian image to enrich the content of pedestrian pseudo labels in the target domain,which improves the iterative training effect of person Re-ID model and enhances the discriminant ability of pedestrian feature expression extracted from the target domain.Finally,the experimental results show that the proposed cross-domain deep learning algorithm is feasible,which can effectively improve the pedestrian image retrieval performance of the person Re-ID model in cross-domain application scenarios.It has a certain theoretical significance for the in-depth research and industrial application of person Re-ID technology.