Research And Application Of Person Re-identification Of Non-overlapping Camera Based On Deep Learning

Person re-identification as one of the important research topics in security,it has received extensive attention in recent years.With the successful application of deep learning technology in this research,the technology has developed rapidly.In actual scenes,there are problems such as pedestrian occlusion and pedestrian posture change,which cause person to have large differences in the view of different cameras.In order to alleviate the above problems,this thesis focuses on the task under unsupervised conditions and proposes a local alignment feature algorithm to reduce the interference of extraneous regions.At the same time,a multi-scale perception pooling feature fusion method is proposed to reduce the lack of key information.This thesis designed a B /S architecture intelligent monitoring system to apply this task to actual scenarios.The core work contents are listed as follows:1.Aiming at the common occlusion misalignment in pedestrian images,this thesis adopts an unsupervised learning algorithm based on feature alignment.Firstly,an alignment algorithm is used to achieve local feature alignment to reduce the influence of the interference blocks of pedestrian images.Then,the global representation of the image is extracted,which can retain the expression characteristics of the pedestrian as a whole.Finally,the global characterization and local alignment characterization are combined,which not only retains the integrity of the entire features but also enhances the local expression ability of the image,thereby improving the performance of the benchmark method.2.Aiming at the problem that deep learning features are missing key representations,this thesis proposes a multi-perception scale pooling method.First,the representations are pooled at different scales as pyramid pooling in the middle-level layers to enhance the perception ability of the receptive field.Then the original representation map and the pooling results are cascaded and merged to reduce the lack of key information make the fea tures extracted by the algorithm more robust and effective.3.Aiming at the application practice and value prospect of person re-identification,this thesis builds a cross-view person re-identification intelligent monitoring system based on B/S architectu re,and embeds the algorithm into the actual monitoring system to solve the person re-identification problem in real scenes.