Research Of Person Re-identification Based On Deep Learning

In recent years,the rise of deep learning makes people come into contact with many technologies and products related to artificial intelligence in their daily life.China pays more and more attention to people's security issues,and the security industry has made great progress because of the deep learning,computer vision and other fields.Person re-identification(RelD)is one of the important research direction in the security industry and video surveillance.It plays a significant role in the field of smart cities.With the rapid development of deep neural networks and the growing demand for intelligent video surveillance,person re-identification has become a common concern of industry and academia.The main task of re-identification is to accurately identify specific targets from a large number of surveillance videos.Pedestrian re-identification can make up for a performance gap that cannot be used in the face recognition systems,such as back,side or blurred face.And the person re-identification performs cross-camera retrieval to realize the trajectory restoration and recognition of the target person.This thesis focuses on the field of person re-identification in smart cities and conducts research based on deep learning technology.It summarizes the current research status of person re-identification at domestic and foreign.Corresponding solutions are proposed for the problem of poor robustness caused by the complex background in person re-identification,the problem that traditional learning cannot be learned by deep learning,and the lack of feature judgment in attribute recognition.This thesis proposes multiple pedestrian re-identification methods from different perspectives of static images.The main research contents include the following parts:(1)An improved model based on the ResNet50 framework is proposed to solve the problems of background difference and human occlusion in the public dataset Market 1501,which uses the multi-layer features of ResNet to perform dimension reduction supervision and improves the loss function.The algorithm in this thesis outperforms the optimal open source algorithm of the same network frame type,AlignedReID++,with a Rank-1 accuracy of 5.4%and mAP accuracy of 10.4%improvements on the Market1501 dataset,with a Rank-1 accuracy of 5.4%and mAP accuracy of 10.4%improvements on the CUHK03 dataset,Similarly,it is higher than AlignedReID++with 5.4%Rank-1 accuracy and 1.6%mAP accuracy on the DukeMTMC-reID dataset(2)In order to solve the problem of color mismatch of person search results in the first part of the improved method,human body analysis was performed based on the LIP dataset.Improved on the basis of traditional adversarial networks,the semantic inconsistency of local features is reduced.Through the small receptive field,global and local features are mutually supervised to avoid the problem of poor convergence of high-resolution image loss functions.The top and bottom colors of the parsed person pictures are extracted,and the traditional manual features are used as prior conditions to screen out pictures that do not meet the color features,thereby improving the performance of person re-identification(3)An attention model was added on the basis of person attribute recognition,and experiments were performed on two large-scale attribute recognition datasets(Marketl501 and DukeMTMC-reID).The correlation between person re-identification and person attributes was analyzed.The results show that the improved algorithm has greatly improved the basic network of the inattentive model,and it can be verified that the improved person recognition model based on attribute recognition has higher accuracyThis thesis discusses the multi-scale pooling of network features,and experimentally verifies the feasibility of the proposed improved method for person re-identification,which can effectively improve the recognition accuracy of person re-identification;and to some extent,overcome the color differences in person pictures.The effective combination of person re-identification and person attribute recognition algorithms verified that the two can promote each other.Finally,this thesis discusses some possible future directions for person re-identification technology and provides recommendations for subsequent related research.