Deep Feature Representation For Person Re-identification

With the rapid development of computer vision technology and deep learning technology,person re-identification technology has great significance in the field of security and criminal tracking.However,monitoring videos with low resolution,partial occlusion,illumination differences,and pose changes limit the development of person re-identification.At present,the main research work is aimed at the above existing problems,focusing on obtaining the discriminative and robust feature descriptors of pedestrians.To this end,this thesis proposes two methods of feature learning to improve the accuracy of person re-identification.The main work of this thesis is as follows:(1)For local occlusion,illumination change and position mismatch in pedestrian images,existing researches have methods based on global features,methods based on local features and methods combining local and global features.In this thesis,local-global extraction unit(LGEU)is proposed for the shortcoming of these methods.The LGEU includes two processes,local selection and global feature learning.The local selection divides the pedestrian features into several parts,and performs local average pooling and convolution operations on the features of each part in turn,learning the importance of each part separately.In addition,global features are combined with the results of local selection to achieve learning of global features.LGEU model can fully exploit the connection between global and local features,and effectively utilize local features and channel information to help the learning of global features.After the comparison of experiments,it proves the excellent performance of LGEU.(2)In view of the change of pedestrian pose,it is inspired by PN-GAN algorithm.This thesis improves it and proposes the Pose-variation removing Generative Adversarial Network(PR-GAN),combining with Siamese network and Generative Adversarial Networks.PR-GAN learns not only the robustness feature related to pedestrian labels but also posture-independent,so that a model with strong robustness to posture changes can be obtained,and no additional posture information is needed in the testing phase.Experiments show that the change of posture has a great influence on the recognition of pedestrians.After removing the influence of posture,the performance is improved.In addition,PR-GAN proposed can not only be used to obtain the robustness feature related to the person label but not related to posture,and can also to generate the pedestrian image of arbitrary pose.