Feature Representation And Re-identification Of Person

With the large-scale deployment and application of video surveillance networks in recent years,how to effectively analyze and process massive amounts of surveillance data has become an urgent problem.Person re-identification has received widespread attention as one of the key technologies,and has made significant progress.The goal of person re-identification is to identify the target person from other scenes according to the specified pedestrian image or sequence data under the condition of cross-scene.This article mainly focuses on three major aspects:person re-identification based on image data,person re-identification based on sequence data,and unsupervised person re-identification.The main research contents and innovations include:In the study of person re-identification based on image data,a person representation algorithm is proposed,which combines the middle-level semantic features and part-based features to obtain the discriminative features containing multi-level semantic information.In addition,a multi-scale saliency-based feature enhancement algorithm is designed to improve the representation ability of pedestrian features.In the study of person re-identification based on sequence data,a bidirectional frame feature enhancement algorithm is proposed,which restores the missing information of the image while suppressing noise,and realizes the enhancement of pedestrian features.In addition,a hybrid feature fusion strategy is proposed.Through the hierarchical correlation fusion algorithm and 3D convolution,the effective extraction of pedestrian spatio-temporal features is realized.In the study of unsupervised person re-identification,an unsupervised re-identification framework based on the enhancement of pedestrian key point information is designed.The adaptive cluster diversity regularization effectively is designed to improve the generalization performance of cross-scene person re-identification.The performance of the three algorithms is compared with the current representative algorithms on multiple public person datasets,and the results verify the effectiveness of the proposed algorithm.