Research On Person Re-Identification Based On Pose Estimation And Feature Alignment And System Implementation

The purpose of Person Re-Identification(Re-ID)is to retrieve specific person targets from a collection of images captured by multiple cameras.With the widespread deployment of video surveillance systems,Re-ID is widely used in secure and smart cities.However,In the real-world surveillance system application,the accuracy of Re-ID still suffers from occlusion.To address this issue,we propose a Re-ID method based on Pose Estimation and Feature Alignment in this paper.It has the pose correction module(PCM),the feature reconstruction module(FRM),and the part align module(PAM).The specific contents are as follows:1.Research on pose correction for Person Re-IdentificationAiming at the problem of pedestrian information loss and adding meaningless features due to occlusion,a posture correction module PCM is designed.Before the feature matching stage,PCM builds the key-points confidence corrected mechanism of the self-adaptive weights learning before the part matching stage to strengthen part-level features in the non-occluded region and weaken ones in the occluded region.In addition,the corrected key-points confidence maps are used to guide the network to focus on the visible region and reduce occlusion interference.First,the Auto-Encoders is added to the Pose Estimation branch.Secondly,the heat map of key points output from the human posture estimation model and the loss function designed in this paper are used to determine whether there is occlusion of pedestrian targets.Then,the visibility weight of each part of the person adaptively adjusted according to the loss value of the joint loss function.Finally,the confidence values of key points of each part of the person are corrected according to the weight,and the feature extraction network is guided to learn pedestrian features to reduce the interference of occlusion noise.2.Research on feature reconstruction and part align for Person Re-IdentificationIn the feature matching stage,the network cannot eliminate interference caused by inconsistent background features and different scales in the matching area.The feature reconstruction module FRM and the component alignment module PAM are designed.A mechanism for separating inter-domain and class-domain features is proposed in FRM.By changing the feature distribution distance of the feature extraction network,the two feature extraction branches focus on the background and foreground pedestrian features respectively,there by suppressing background interference.A two-stage multiscale fusion strategy for human topological features and global semantic features is designed in PAM.In the first stage,key-points are used to dynamically divide feature boundaries,and then local and global features of different scales are extracted through outer product.Then,after global max pooling of key-points.a second stage of feature fusion is performed with global semantic features.The fused feature vector contains not only the semantic feature information of the target pedestrian,but also the topological information of the human body,eliminating noise interference caused by scale or posture inconsistencies,and enhancing the ability of the model to explore effective pedestrian feature information in regions of different granularity.3.Design and implement a Person Re-Identification systemBased on the proposed research model,an intelligent Person Re-Identification surveillance system is designed and implemented.The system design follows the principle of business layering,and implements five functional modules: Intelligent Processing,Human-Computer Interaction,Security Control,Media Service and Data Management,which meet the practical application requirements of the surveillance system under the occlusion scenario.