Person Re-identification Based On Pose Embedded Multi-scale Feature

Person re-identification(RE-ID)is a technology that uses computer vision technology to determine whether there is a specific pedestrian in an image or video sequence.It is generally considered as a sub-problem of image retrieval.In recent years,the continuous development of deep learning has made great success of pedestrian re-rccognition technology,which has also been widely used in intelligent security and intelligent monitoring and other fields.Since pedestrians arc vulnerable to changes in attitude,changes in perspective,complex background and occlusion,etc.,they arc difficult to be recognized and retrieved.Currently,there arc many problems to be solved:(1)How to solve the impact of occlusion and complex background by focusing on the significant area of pedestrian image.(2)Pedestrian pose and background image constantly change with pedestrian movement.How to reduce the problem of misalignment in pedestrian matching by learning the change of Person pose?(3)Feature extraclion and distance mcasurement cannot completely climinate the effects of different interferences.How to extract more robust features and improve pedestrian matching to eliminate the effects of different interferences?In view of the above problems,this paper studies the Person re-identification method,and its main contributions can be summarized as the following two aspects:(1)In order to solve the problem of the interference of sheltering and complex background,this paper proposes a model based on significant characteristics of the depth of the polymerization methods,first using the particularity of the pedestrian image significant characteristics of each pedestrian image are extracted,the significance of using significant algorithm to extract the image body area to filter the background complex part.Then the deep learning method is used to learn the global features of the image,and the original image is extracted by convolution feature to capture the global information of the image.Finally,a feature weighting scheme is designed to effectively fuse the extracted features and use the fused features to learn.Through this strategy,the most information-rich region of each image is enhanced,so that the extracted features are more robust.The experimental results show that the proposed method can focus on some salient areas of pedestrians adaptively according to the characteristics of the image itself,and effectively reduce the influence of complex background and occlusion.(2)In order to solve the problem of misalignment caused by the change of attitude estimation,a multi-scale convolution feature fusion model based on pose embedding is proposed in this paper.Firstly,the original image is preprocessed by the random erasure image preprocessing method to improve the performance of the baseline network.At the same time,the triple loss function is used to improve the generalization ability of the baseline network and to extract the deep global features that can effectively capture the global information.Then,the pose learning subnetwork was embedded with the significance model and the attitude estimation model.Through the embedded pose learning subnetwork,the high-order features and topological information of the image were learned,and the semantic local features which paid more attention to the local details were extracted.Finally,considering the complementary advantages among different scale feature vectors,the depth global feature and the local semantic feature were weighted by depth aggregation,and the similarity measurement was improved by reordering strategy.The experimental results show that the proposed method can learn the attitude estimation adaptively,and effectively reduce the misalignment problem of pedestrian image through local spatial learning and global information capture.