Research On Person Re-identification Algorithm Based On Attention Mechanism And Local Feature Fusion

Person re-identification is an important part of criminal investigations,smart cities,and skynet projects,and aims to match and retrieve information on specific pedestrians under non-overlapping cameras.Complex background and changeable posture are the main challenges faced by person re-identification tasks.Extracting high-discrimination feature representations of persons is the key to improving the performance of person re-identification algorithms.This thesis focuses on the main body enhancement and fine-grained information learning in the process of feature learning through the attention mechanism and the method of extracting local features.The main work is as follows:Aiming at the problem of low pedestrian feature discrimination caused by interference information such as background and occlusion in the person detection image,the attention mechanism is used to enhance the representation of pedestrian features.According to the idea that the attention mechanism can obtain high-discrimination attention features by reweighting feature maps,this thesis proposes a cluster-based global attention module(Cluster-based Global Attention Module,CGAM),the attention weight learning process is reconsidered as the clustering center learning process,the spatial location points in the feature map are regarded as feature nodes,the clustering method is used to obtain a set of high-confidence attention weights to enhance the representation of pedestrian features.Aiming at the problem of CGAM-based algorithms that use global features to cause pedestrian details to be ignored,local features are used to learn fine-grained pedestrian information.In this thesis,the rigid horizontal segmentation method is used to obtain the local features.Specifically,the previously obtained attentional feature is divided into four rigid horizontal stripes,which are used to learn local features.Use the improved Resnet-50 as the backbone network and embedding the attention module and local branches,the global branch and the local branch are used.Triple loss and recognition loss are used for the global branch.Only recognition loss is used for local branches.In the test stage,the four local features are spliced to obtain the total local feature,and then the total local feature and the global feature are merged as the final person feature representation for classification.In order to verify the feasibility and efficiency of the algorithm,on the three popular datasets:Market-1501,Duke MTMC-re ID and CUHK03,multiple experiments were conducted with rank-1 and m AP as the evaluation criteria.Experimental results show that the person reidentification algorithm proposed in this thesis has excellent performance compared with other related methods.Finally,a brief summary of the research content of this article is made,the existing shortcomings are analyzed,and the research direction and goals of the next stage are given.