Research On Person Re-identification Method Based On Deep Feature Metric

Recently,Cameras and surveillance has played a crucial role in the construction of smart city,and person re-identification(Re-ID)is an important content of the cameras and surveillance.Therefore,more and more researchers have begun to focus on person re-identification.Person re-identification refers to identifying whether the pedestrian images captured by different cameras are the same person in a multi-camera and surveillance system.,and it can automatically track and retrieve targets such as criminal suspects in the cameras and surveillance network.However,in reality pedestrian images captured generally have problems such as blurred pictures,background interference,and irregular changes in pedestrian posture,which cause large differences in the appearance of the same pedestrian,and these problems bring huge challenges to person reidentification.Earlier research on Re-ID mainly used traditional feature extraction methods(such as color,shape and local descriptors)to extract the shallow invariant features of pedestrians.As deep learning methods mature in computer vision,researchers have used deep convolutional neural networks to solve Re-ID.Firstly,this process uses a deep learning network model to learn the representative pedestrian feature,and then uses a distance function to measure the similarity of the pedestrian feature vector to determine whether the pedestrian images are the same person.Therefore,designing a distinguishing and stable feature indicator to represent pedestrian images is the most important part of person re-identification.Based on the metric learning method,this paper proposes an improvement scheme for the feature network and the loss function,which efficiently solves this type of problems while achieving good known results.The main research contents are as follows:1.The basic constitution and basic method of person re-identification are introduced in detail.For person re-identification,the basic processing flow is systematically introduced,and then several important research methods for person re-identification are introduced,including feature representation methods and metric learning methods,and the characteristics of various methods are systematically introduced.Finally,the evaluation criteria and public data sets of person re-identification are described.2.Inspired by the idea of residual network and attention mechanism,to solve the problem that the appearance of the same pedestrian under different cameras is quite different,the ordinary network model cannot extract representative pedestrian features,so it is proposed an enhanced deep feature converged network model: Firstly,in the feature learning stage,the invariant features of pedestrians are mainly extracted.Divided into three branches,one branch is used to learn the global features of pedestrians,and another branch is used to learn local features.In the global branch,the spatial attention and channel attention are combined to propose an embedded attention network called Spatial and Channel attention network(SC-Net),and then the SC-Net network is embedded into Res Net50;In the local branch,a long-short-term memory change network(LSTM-CN)is proposed,which uses LSTM-CN to extract more representative local features.Finally,the global features and local features are fused,and then the network model is trained by means of the joint loss function.The excellent results of this method are verified on multiple public datasets.3.In view of the situation that similar categories in person re-identification tasks are easy to be misclassified,if we only extract pedestrian feature information from a single scale of pedestrians,it will bring limitations.Pedestrian features extracted at a single scale are poorly discernible.therefore,this paper proposes a multi-scale network model(MSN).This network model decomposes pedestrian images from a multi-scale perspective,and the feature vectors extracted from each branch are sent to loss function to train the network model.The loss function abandons the traditional triple loss function.In order to consider the impact of specific samples,we use a combination of two methods,cross entropy loss function and label smoothing regularization(LSR),to effectively solve the arbitrary nature of the training samples.The MSN model has achieved good experimental results on the public Re-ID datasets.At the same time,the validity and correctness of the method is verified by using the baseline model.