Research On Person Re-identification Algorithm Based On Pose-embedding Graph Representation

Person re-identification,as a key technology in the field of intelligent video surveillance,aims to correctly identify images belonging to the same person from the dataset collected by surveillance cameras under non-overlapping perspectives.In recent years,the development of deep learning has provided a strong technical basis for research on person re-identification.The methods based on part features can effectively improve the accuracy of person reidentification,but most of them ignore the potential relationship between part features.The influence of complex background noise and pose variants,as well as the existence of similar person appearances and partial occlusions between person images,bring great challenges to the research of person re-identification.In view of the above problems,this paper mainly starts from the perspective of the global relationship of the person,uses the basic convolutional neural network as the backbone,makes full use of the rich pose information of the person,and integrates the inherent skeletal structure information and fine-grained information of the person image,to effectively improve the performance of person reidentification.The main contributions of this paper are mainly in the following two aspects:(1)An Edge-Score-Embedding graph convolutional network(ESE-GCN)person reidentification network model is proposed.Considering that the existing methods based on part features ignore the potential relationship between the part features,and there are similar personal appearance exists between images,the proposed ESE-GCN constructs the topological structure of the human joint and skeleton by using the abundant pose information combined with the inherent structural relationship of the person.ESE-GCN explores the inherent skeleton structure information of the person(such as the skeleton length between the joints),and the skeleton structure information is learned through the constructed edge weight score predictor to obtain the adjacency matrix of the human body topology graph.Finally,the extracted part features of person key points features and the learned adjacency matrix as input,utilize the graph convolutional network to effectively integrate the potential relationship in the human joint and skeleton structure,to extract the robust feature representation of person images and improve the recognition accuracy of the model.On the public Market-1501 dataset,the accuracy rates of Rank-1 and m AP of our proposed method reached 96.3% and87.0% respectively,which were 7.8% and 12.9% higher than the existing state-of-the-art methods on average.(2)A part-feature-aware graph convolutional network person re-identification network model named ESE-TGCN is proposed.Considering that the deep person features extracted by ESEGCN using the convolutional neural network will lose some details information of the original image,these details information are also an important basis for identifying the similar appearance of different person images.To this end,the Transformer learning branch is added to the ESE-GCN to extract more discriminative feature representations of the person image.In the Transformer learning branch,the fine-grained semantic part features of the person images are constructed as input,and then learning the global relationship between the semantic part feature sequences through the cascaded Transformer encoder model,driving the model to focus on different part regions of the person images and further improving the recognition accuracy of the model.On the public Market-1501 dataset,the accuracy rates of Rank-1 and m AP of our proposed method reached 97.3% and 87.8% respectively,which were8.8% and 13.7% higher than the state-of-the-art methods on average.