Research On Person Re-identification Algorithm Based On Deep Learning

Person re-identification is a technology that quickly recognizes and matches target pedestrians in an environment with multiple cameras.Recently,the number of surveillance cameras on the street,shopping malls and other places continues to increase,pedestrian re-identification technology has been widely used,playing an important role in criminal investigation,intelligent security and public safety protection.However,in practical applications,the pedestrian re-identification research has brought many challenges to the research of pedestrian re-identification due to the changeable poses of pedestrians,the susceptibility to occlusion of the target person,and the blurring of the photographed image.Therefore,how to accurately and effectively extract the feature information of the target pedestrian from the image is the focus and difficulty of this problem.The traditional person re-identification methods mainly judging the category information of person through clues such as shadows and textures.Although they are easy to implement,those methods are greatly limited by the external environment in the process of feature detection and matching,and are incapable in generalization.While the person re-identification based on deep learning methods make full use of the object structure and other information in the image,so these methods have robust feature representation and good generalization ability.In response to the above problems,this article uses deep learning methods to conduct the following research on the basis of previous studies:(1)A pedestrian re-identification network based on multi-scale feature fusion is proposed.In view of the difficulty of pedestrian feature extraction due to changes in pedestrian posture and location,this paper uses multi-scale feature fusion strategies at different stages of the network to effectively retain more high-level semantic information,thereby helping the network to extract more complete pedestrian feature representations;Then,the longdistance contextual semantic information is simultaneously aggregated from the horizontal and vertical directions to reduce the interference of the surrounding environment and improve the ability of the network to discriminate pedestrians.(2)Propose a network based on attention perception.In order to make more effective use of the global information in the network,we introduced a channel-dimensional attention mechanism to fully consider the connection between feature channels,thereby enhancing the network's representation of global information and improving the model's scene analysis capabilities.Considering that the hole convolution used in this article may reduce the continuity of feature information while obtaining a larger receptive field,this article fuses the multi-level cavity convolution structure with the complete feature image to improve the continuity of feature information.To increase the accuracy of the model ulteriorly.After a lot of experimental analysis on the data sets Market-1501,Duke MTMC-re ID and CUHK03,the results showed that the average accuracy value(m AP)reached 89.1%,81.5%and 64.4%,respectively.(3)Based on the pedestrian re-identification algorithm proposed in this paper,a pedestrian re-identification system is designed and implemented,and the task of pedestrian re-identification is integrated in the humancomputer interaction interface.By entering the target pedestrian to be queried,the results of pedestrian re-identification can be displayed intuitively,and all query results can be viewed.