Research On Person Re-identification Based On Deep Learning

In a real video surveillance scene,in order to achieve the maximum coverage area,the camera is usually placed in a higher corner,which will seriously affect the quality of the captured image.In addition,any uncertain factors will cause the imbalance of pedestrian attributes,thereby affecting the performance of the recognition method.Pedestrian re-identification technology is currently a research hotspot in the field of video surveillance,and new theoretical research and methods have been continuously derived.However,mainstream pedestrian re-recognition algorithms need to be further developed in terms of studying the local characteristics of pedestrians.Therefore,the huge challenges faced by the pedestrian re-identification problem are summarized as follows:1)How to improve the quality of the data set to improve the accuracy of pedestrian recognition;2)How to improve the convolutional neural network structure to learn more refined pedestrian characteristics.The main research work of this article can be divided into the following two aspects:1.Person reidentification technology is affected by the camera shooting distance,angle,person density and lens occlusion,and the quality of the deep learning model is affected by the data set.These uncertain factors will lead to imbalance of person attributes,thereby reducing person Recognition performance for re-identification.Starting from the color of the data set,this paper proposes a person classification method based on the RGB-gray image feature of the convolutional neural network to classify the person in the image.This method first explains the color overfitting problem of the existing model.Secondly,using the complementarity of the RGB image and the gray image,the network model is used to learn the characteristics of the RGB and the gray image at the same time to better represent the person,and to a certain extent alleviate the color overfitting problem.The experimental results of the method in this article on the Market-1501,DukeMTMC-reID,and CUHK03 data sets are Rank-1=92.3%,Rank-1=85.2%,and Rank-1=73.3%,respectively.2.In order to be able to mine more refined person characteristics,this paper proposes a multi-scale feature model.First,the pyramid strategy is used to obtain person features of different scales.These features are distinguishable for different person but are easy to be ignored.The features are combined to form a feature that can better represent person,so as to improve the accuracy of person identification.The experimental results under the Market-1501,DukeMTMC-reID,and CUHK03 data sets are Rank-1=90.5%,Rank-1=82.6%,and Rank-1=66.8%,respectively.