Person Re-identification Based On Pose-guided Generative Adversarial Network

Person re-identification is a challenging and valuable research topic in the field of computer vision.It needs to match person images with the same identity in multiple camera systems.In recent years,face verification technology has become more and more mature and has achieved great success in building "smart city" and "safe city".However,in realistic video surveillance systems,cross-cameras can not guarantee that the visually recognizable images of human faces can be captured in complex and changeable situations.Therefore,it is particularly important to use the whole body information to locate and identify pedestrians.The convolutional neural network is used to extract the overall pedestrian features to achieve cross-camera tracking of pedestrians.As a result,the field of computer vision began to "person re-identification" technology research work.Person re-identification can be simply understood as an image retrieval problem,that is,given a picture of a person,one or more images of the person are found from the pictures taken by multiple cameras.When the video surveillance system recognizes the same target,the problems of viewpoints,illumination intensity,posture variety and partial occlusion will bring huge difficulties for person re-identification across cameras.In order to mitigate the influence of pose variations on person re-identification,the paper proposes a method based on a pose-guided generative adversarial network(PG-GAN),which can be used to learn identity-sensitive and pose-insensitive features.The algorithm is composed of a Siamese convolutional neural network(S-CNN)and generative adversarial networks(GANs).S-CNN is a symmetric structure with Dense Net-121.GANs contain multiple pose discriminators and identity discriminators,as well as incorporate pose loss,which requires appearance of generated images with same identity to be similar.In order to improve the generalization ability of PG-GAN to new images,we use PG-GAN to learn pedestrian features which are sensitive to identity and insensitive to posture.During test,for the purpose of improving retrieval accuracy,this paper also uses the re-ranking method based on k-reciprocal nearest neighbors to reorder pictures to be retrieved from the gallery.The proposed method has reached a high level of recognition on the three public datasets,and also plays a good role in the actual video surveillance system,which fully shows the advancement of the algorithm presented in the paper.