Research On Cross-Modality Person Re-Identification Based On Deep Features

The massive video information brought about by the rapid development of the video surveillance network brings a huge challenge to the analysis of traditional artificial video surveillance.Person re-identification is an important issue in the analysis of video content under large-scale video surveillance networks.It has important scientific significance and great application value.In a video surveillance network,pedestrians are affected by factors such as the angle of view,posture change,background noise,lighting conditions,camera settings,etc.,resulting in a person having a large difference in the same camera at different times or in different cameras.In particular,under different light conditions during the day and at night,the cameras often work in visible light and infrared modes,which poses a great challenge to persons for recognizing problems.This paper investigates the challenging issue of infrared-visible light cross-modality person re-identification under different lighting conditions during the day and night.The main tasks are as follows:1.A person re-identification method based on cross-modality generation training is proposed.In this paper,we design a cross-modality to generate confrontation networks and perform end-to-end training to learn the cross-modal discriminative infrared-visible person characteristic representation to achieve person re-identification.This method is based on the generation of an antagonistic training method.A discriminator is designed and a generator network is designed.The discriminator judges the modality to which a given image belongs.The generator combines the mixed triplets and the classification loss function.Through confrontation training,the algorithm projects infrared image pedestrian features and visible light pedestrian features into a common subspace,learns discriminant infrared-visible pedestrian features,and realizes cross-modality person re-identification.The method proposed in this paper validates the effectiveness of the infrared-visible person re-identification dataset and achieves the best performance at present.2.A complete cross-modality person re-identification system is designed.This paper not only realizes the person search from infrared image to visible light image,but also integrates the person search of visible light image to visible light image,thus forming a complete cross-modality person re-identification system.The research in this paper focuses on cross-modality person re-identification technology.The cross-modality person re-identification method for cross-modality training is used to solve the problem of heavy-to-neighbor recognition in different lighting conditions during the day and night.It has a higher theoretical and practical value on research and application of person re-identification technology.