Research On Pedestrian Re-identification Method Based On Convolutional Neural Network

With widely used of the video surveillance technology in public security,it is difficult for a single camera to satisfy multi-angle and multi-direction surveillance.So the technology of person re-identification based on multi-camera has become a hot issue in the field of video surveillance.The core of person re-identification is the correlation matching of the same target with different cameras and monitoring periods.However,due to the influence of camera angle and illumination changes,there are some problems in pedestrian images,such as pose variations,dress similarity and partial occlusions,which bring great challenges to person re-identification.In this paper,the methods of person re-identification based on convolution neural network are investigated.Different structures of network models are used to solve the problems of variations of illumination and viewpoints,information loss and mismatch in person re-identification,so as to improve the robustness and accuracy of the algorithms on VIPeR,CUHK01,CUHK03 and Market-1501 datasets.The main research contents of this paper include the following aspects:(1)This paper proposes the SVM-CNN model for person re-identification which is solved as ranking-like comparison issue.The joint representation learning method is used to combine the traditional two inputs of network as a single input.The Softmax classification function is used at the loss layer of the network,instead of the traditional Softmax classification function to reduce the complexity of network and the loss of effective information,and increase the discriminability of features.Experiment results on the datasets of VIPeR and CUHK01 demonstrate that the identification performance of the SVM-CNN model is significantly improved by more than 2% and 10% compared with KLFDA,MFA and other advanced algorithms.(2)Since the features extracted from the network are not robust to the variations of illumination and viewpoints,this paper proposed the SVM-CNN model with multi-feature fusion.This method combines the three features of SCNCD,ELF and SILTP,which are robust to the variations of illumination and viewpoints.It retains the local features of pedestrian images to the greatest extend,and can be better applied to the scences with great variations of illumination and viewpoints.The experiment results on the datasets of VIPeR and CUHK01 demonstrate that the identification accuracy of the network is improved by more than 3% compared with the advanced feature representation algorithms such as SCNCD and LOMO+XQDA,and it is superior to the SVM-CNN model proposed in chapter 2.(3)The multi-channel pyramid matching SVM-CNN model(MC-SPSN)is proposed to sovle the mismatch problem which can easily occur in the person re-identification.This method adds the pyramid matching module consisting of three multi-scale atrous convolution layers to the network,and reduces the occurrence of the image mismatch without increasing the parameters of the network by using the image semantic features.It can be applied to complex background environment and partial occlusions.The experiment results on the datasets of VIPeR,CUHK01 and CUHK03 demonstrate that the identification accuracy of MC-SPSN model is significantly improved by 3%,18% and 15% compared with DCSL,ImprovedDL,LOMO+LSTM and other advanced deep learning methods,and it is superior to the SVM-CNN model with multi-feature fusion proposed in chapter 3.(4)In order to solve the efficiency problem of person re-identification in large amounts of data,QDH-SPN model based on quadruplet deep hashing is proposed in this paper.In this method,hash technology is introduced into the research of person re-identification,which can improve the efficiency of network as well as reducing the loss of effective information in recognition,so that it can be better used in person re-identification with large-scale data.The experiment results on the datasets of Market-1501 and CUHK03 datasets demonstrate that the efficiency of QDH-SPN model is improved with a large amount of data compared with WARCA,GateSiamese and other advanced person-identification algorithms,and the identification accuracy is improved by about 7%,which is better than the MC-SPSN model proposed in chapter 4.