Research On Person Re-identification Based On Global Feature And Local Feature Jointing

Person re-identification is a technology that uses computer vision to identify and retrieve specific person across cameras and scenes.It is mainly used in criminal investigation,security monitoring,intelligent commerce and so on.The current person re-identification methods mainly focus on two aspects: feature extraction and metric learning.Feature extraction determines the upper limit of model performance,which can directly affect the subsequent metric learning.Because of the problems such as the change of view angle,the low resolution,the dim light and the occlusion of the pedestrian image taken by the surveillance camera,the feature extraction of the person image has brought great challenges,so it is still a very challenging task to research a distinguishing person feature extraction method.This paper studies and improves the network framework of pedestrian recognition based on deep learning from global and local feature extraction.The main research contents are as follows:(1)For global feature extraction,we proposed a person re-identification method based on multi-layer global feature jointing.Firstly,the multi-layer global feature of the network is extracted by convolution and pooling layer,which improves the performance of the network from the spatial dimension.The multi-layer global feature after jointing is used as the global feature attribute of the person image.Secondly,a batch normalization is added after the multi-layer global feature jointing.At the same time,the label smoothing loss and triple loss are used to train the model.(2)For local feature extraction,we proposed a person re-identification method based on multi-granularity local feature jointing.It is a network with multi-granularity local feature and multi-layer global feature splicing.The network firstly adopts multi-layer global feature jointing network structure to extract the global feature of person image;secondly,it is divided into two branches to cut the person feature horizontally.Finally,global and local features are jointed as person feature.Multi loss function strategy is used to constrain the model,label smoothing loss and rank list loss are used for global features,softmax loss is only used for local features,and the generalization ability of the model is improved through the targeted selection of loss function.Finally,we experiment on the Market1501,DukeMTMC-reID,CUHK03 and MSMT17 datasets.The results show the advantages of the proposed methods,especially the multi-granularity local feature stitching method.