Research On Person Re-Identification Technology Based On Deep Learning

With the development and popularization of smart cities,computer vision technology has a very wide range of application scenarios,such as criminal investigation,intelligent security,image retrieval,human-computer interaction and so on.Compare with the traditional manual identification,intelligent monitoring and management technology can greatly reduce costs and improve the detection efficiency.Cross-camera retrieval is a characteristics of person re-identification.The evaluation index is to retrieve the same person under different cameras.In a real scene,due to uncontrollable factors such as illumination,shooting angle,and person posture changes,it is difficult to identify the retrieval target,which makes the research of re-identification a challenging issue.In recent years,with the rapid development of deep learning in multiple fields,this paper research on the person re-identification technology based on deep learning.The main work and contributions of this paper are as follows:Most of the existing methods are based on the global features of persons or the local features of simple division.However,feature extraction is an important part of person re-identification technology.In order to extract more discriminative feature information on person images and reduce the influence of background environment on person feature extraction,a convolutional neural network framework based on person skeleton division and multi feature embedding is proposed.In this paper,we extract skeleton points of person,and used the optimization division criteria to divide the local area of person image.The convolutional neural network is used to train local areas and global area respectively,and then extracted global and local features are fused.The fused features retain both richer local informations and complete global information.In the similarity distance calculation,the cross-matching principle of local information is used to expand the distance between classes and reduce the distance within classes.Finally,a new ranking list is obtained.In the research of person re-identification technology,the camera visual changes and cross-domain tracking are the difficulties,which include different light intensity,background differences and changes in person posture.Due to the scale-invariant and rotation-invariant of feature points,we proposed a person re-identification method based on hybrid structure and multiple feature points matching.Combined with the advantages of deep learning,the multi task convolution neural network is used to train and extract the features of person images.The obtained person features are sorted by traditional measurement methods.By obtaining the feature points of the image in the original ranking table,and matching them with the feature points of the target sample.The number of correctly matched feature points is taken as the measure of similarity distance to obtain a more accurate ranking list.In this paper,experiments are carried on the public datasets Market1501,Duke,and CUHK03 to verify these methods proposed in this paper.The experimental results show higher accuracy in Rank-1 and mAP evaluation indexes.It shows that the improvement of the algorithm proposed in this paper on feature extraction and similarity distance can improve the retrieval accuracy of the person re-identification algorithm.