The Research Of Person Re-identification Based On Deep Learning

With the development of smart cities and the increasing demand for public security,the number of cameras installed is increasing,the application fields of intelligent video analysis technology are expanding,and the technical requirements are constantly improving.An important technology in intelligent video analysis technology is person re-identification.Person re-identification is a cross-view retrieval task and aims at identifying a specified pedestrian in other fields of view,given one or more photo person-of-interest.Person re-identification is currently unable to meet the standard of commercial application with the influence of low resolution of video,the diversity of pedestrian posture,pedestrian shielding,complex monitoring environment,etc.This thesis studies the person re-identification based on deep learning.The main research work is summarized as follows:1.For the problem of representation learning focuses on learning the robust feature of pedestrian and metric learning focusing on processing the similarity of pedestrian,this thesis combines two methods,designing and implementing a global feature re-identification model with classification loss and metric loss.Compared with the classified network,rank-1 and mAP increased by 1%and 1.1%on Market1501,1.4%and 1.2%on DukeMTMC-reID,1.7%and 1.6%on CUHK03.2.In order to explore the accuracy of person re-identification with global feature and local features,this thesis proposes a new method to obtain local features based on the global feature model.The specific method is to perform principal component analysis on all the column vectors on the tensor feature map outputted by the last convolutional layer of the convolutional neural network,and use the principal component as the pedestrian feature,and propose the minimum distance method to measure the similarity between pedestrians.The number of principal component features is determined by experiment.The experimental results on multiple datasets show that the feature column vectors selected by principal component analysis is an effective pedestrian feature description,and the local feature re-identification method is higher in accuracy than the global feature re-identification method.3.For the problem of the accuracy of the global feature model based on representation learning and metric learning is still low,this thesis makes modifications on the basis of Part-based Convolutional Baseline and adds the triplet loss constraint,combining the two kind of loss to supervise the local features learning.For the phenomenon of the missing detection of the sample with the same identity and false detection of the sample with the different identity in the pedestrian retrieval method of this thesis,this thesis proposes the local features reweighting to improve the fusion algorithm,optimizing the distance between the samples again.Comparing to the reproduction PCB,rank-1 and mAP of improved algorithm finally increase by 13%and 3%,2.0%and 2.9%,2.5%and 2.6%in Marketl501,DukeMTMC-reID,and CUHK03.