Convolutional Neural Network Based Person Re-identification

In real-life video surveillance scenes,the camera had been generally choose to place it in a relatively high corner in order to be able to cover the largest area,which results in serious impact on the image quality of the captured image.Any uncertain factors will lead to the imbalance of pedestrian properties,thereby affecting the performance of the discriminant method.ReID technology has already been a hot topic in the field of video surveillance,while many methods and theoretical basis for person identification through video surveillance have been proposed or introduced.However,as a more common person identification method,it is a big challenge since most ReID methods usually fails to recognize the pedestrian parts feature.The problems of ReID technology can be described as the following aspects: 1)Improving pedestrian recognition accuracy by solving the label imbalance problem of pedestrian attribute.2)Learning more details of the pedestrian attribute by using improved convolutional neural network.3)Building the end to end ReID model by setting pedestrian parts cropping methods more effectively.To solve these problems,the main research work of this paper can be divided into two aspects:1.ReID technology is affected by factors such as camera distance,angle,pedestrian density and occlusion rate.The imbalance of pedestrian attributes will affect the recognition performance of ReID,since some uncertainties such as camera shooting distance or angle,pedestrian density and occlusion rate will lead to the imbalance.In this paper,a ReID on label Sensitive Convolutional Neural Network based Pedestrian Attribute Classification is proposed,by adjusting the existing multi label convolutional neural network structure,so that it can perceive more detailed pedestrian attributes,and updating the weight of the convolution neural network by error back propagation method,so that the network model can be sensitive to the weak labels,to reduce the influence of pedestrian label unbalance on ReID.2 An end-to-end ReID model is proposed,by adaptive learning pedestrian body segmentation parameters on neural network and integrating the label sensitive convolutional neural network learning pedestrian attributes,to find more detailed label feature of pedestrian.To verify the validity of the proposed ReID model,the comparative experiments are carried out on multiple video surveillance video sets.The experiment results show that the proposed method in this paper can handle the task of ReID well.