Densely Aligned 3D Part For Person Re-Identification

Person re-identification refers to the matching of pedestrians in the scene of the cross-camera.It is one of the hot research areas in computer vision.However,in practical application scenes,the variation of camera angle,occlusion of parts,and the variability of pose enable pedestrians misaligned in space,which brings a great challenge to the feature extraction stage.Therefore,how to obtain high discrimination features in cross-perspective scenes is the focus of person re-identification research.To solve the above problems,this thesis proposes a person re-identification method based on densely aligned 3D parts.The details are as follows:(1)Alignment of parts based on body posture estimation.A lot of work utilizes 2D pose estimation to obtain aligned pedestrian images.However,pose estimation errors commonly cause the framework to fail to accurately detect the joints,in which case a large number of parts are lost.Secondly,due to the limited information in 2D space,the same pedestrian may contain different part information in different images.Referred to the Dense Pose model,this thesis proposed a dense 3D part alignment algorithm with multi-frame completion to align pedestrians into the 3D dense space at the pixel level.Even if some parts cannot be detected,the problem of complete loss of parts or partial loss of part information can be alleviated by using the strategy of multi-frame completion.(2)Person re-identification based on densely aligned 3D part maps.To further improve the robustness of features of dense 3D parts,we designed a pixel-aligned multi-branch part reconstructing network to map the pixels on the pedestrian image to a unified densely aligned space.Specifically,we use a multi-tasking learning scheme.It consists of part reconstruction task and person re-identification task.In the reconstructing subnet,we assigned the task to reconstruct 24 densely aligned part maps,and at the same time,we carried out the feature representation learning task in the backbone network to learn the densely aligned features.The experiment result shows that our method has a good effect and precedes many state-of-the-art methods.Through the above research,this thesis settles the problem of pedestrian misalignment in the person re-identification task.Firstly,the Dense Pose model and multi-frame completion strategy are used to construct the densely aligned 3D part map.Then,a feature representation learning network is designed to extract densely aligned 3D part features.In the final experiment,our method has achieved a good result,which has the indicative price for the research of pedestrian misalignment in person re-identification.