Research On Visible-infrared Person Re-identification Methods Based On Deep Feature And Metric Learning

With the popularization of intelligent surveillance video,the research on person reidentification has received more and more attention.Thereinto,person re-identification in the visible light field is one of the main research directions,but environmental conditions such as dim light or night will affect how the research results apply in real scenes.In response to this problem,considering the increasing application of dual-mode cameras,this thesis focuses on the visibleinfrared person re-identification,that is,recognizing visible(infrared)pedestrian pictures in infrared(visible)modality.The cross-modality discrepancies caused by the different reflection spectrum of visible light cameras and infrared cameras,and the intra-class variations caused by different camera viewpoint,pedestrian posture changes,complex backgrounds,illumination,etc.make visible-infrared person re-identification tasks more challenging,therefore,this thesis proposes two approaches based on deep feature and metric learning,and the main research contents are as follows:1.This thesis proposes a visible-infrared person re-identification method based on adversarial dual-path multi-stage fusion network(ADMF).The network framework of this method is divided into two parts,dual-path multi-stage fusion network and modality discriminator.The dual-path multi-stage fusion network generates the multi-stage fusion feature representations of the images,and the modality discriminator determines which mode the input feature representations come from.This method uses the idea of generative adversarial network for reference and optimizes the two parts through the minimax game.We also design a new dual-modality constraint loss(DCL)to guide dual-path network training and address the problem of cross-modality and intra-modality intra-class variations.2.This thesis puts forward a visible-infrared person re-identification method based on dualmodality hard mining triplet-center loss(DTCL).The DTCL separately learns a visible modality center and an infrared modality center for each class and selects online novel cross-modality triplets and intra-modality triplets for each sample.By shortening the distance between the samples and the centers of the same classes and pushing the distance between the samples and the centers of different classes,DTCL can supervise the network learning of modality invariant information and realize the reliable classification of pedestrians.In addition,we apply a dual-path part-based feature learning network(DPFLN)to learn the local feature representations of pedestrians.We carry out a lot of experiments on two commonly used datasets,RegDB and SYSU-MM01,and get better performance than the current state-of-the-art,which verifies the effectiveness of the proposed method.