Person Re-identification With Deep Learning

The target pedestrian moves away from one camera view and is identified again in a nonoverlapping camera view.This process,known in computer vision as Person re-identification(reid),is widely used in real life.At present,the research on person re-identification focuses on the overall images.However,in the real scene,the images taken are not all complete due to the occlusion,and there may be only partial images of part of the body.Therefore,it is necessary to further study the partial Person re-identification.Compared with the traditional person re-identification,the partial person re-identification is also affected by the serious mismatch between the partial pedestrian image and the holistic pedestrian image and the lack of body information,which brings greater challenges to re-id.In this paper,a convolution network model,pose guided alignment network(PGAN),is proposed to solve the problem of obvious misalignment in partial person re-identification.On this basis,combined with the attention mechanism,the partial re-id based on mask learning is proposed to mitigate the effects of missing information.The innovation and contribution of this paper mainly include the following three points:(1)Pose information is introduced into the Pose-guided Spatial Transformation(PST)as an auxiliary information during training,and the alignment loss function is designed according to the symmetry of human body,so as to automatically align partial pedestrian images to the standard pose.The PGAN model with PST as the core can effectively solve the misalignment problem of partial re-id.(2)It is improved on the basis of the spatial transformation network and used to calculate the pose information corresponding to the affine transformed image,instead of directly using the pose estimation algorithm to obtain the required key point information.Therefore,it is not necessary to embed the pose estimation algorithm in the model,and no pose information is needed in the inference stage;(3)Combined with the attention mechanism,mask learning is performed using a PGAN model,and the mask image generated based on the skeleton information is introduced as guidance information to assist the attention mechanism to learn the feature mask to suppress interference information such as background and improve effective information.By paying attention to the effective information of body parts,the proportion of effective features is increased,thereby alleviating the problem of missing information.In this paper,the effectiveness of the proposed method is verified on two partial pedestrian datasets Partial-i LIDS and Partial-REID.