| Due to the rapid development of surveillance equipment in the security field in recent years,video surveillance analysis technology has attracted more and more attention.Person re-identification(Re-ID)is an important part in video surveillance analysis technology.The main goal of person Re-ID is to find matching images or videos from candidate person pictures or videos across perspectives quickly,which can be widely used in video tracking,pedestrian trajectory analysis,pedestrian retrieval and other fields.However,person Re-ID faces the influence of factors such as attitude changes,occlusion,illumination changes,background occlusion,etc.,making it difficult to obtain robust pedestrian faetures.Graph convolution aims to enhance the feature representation of vertices by using the relationship between different vertices.Therefore,graph convolution based person Re-ID is also developing rapidly.The existing related research work has many problems such as low performance and insufficient optimization.In order to solve the aforementioned problems,this thesis transform person Re-ID into graph representation learning,and the contributions of this paper are as follows:This thesis propose an Edge Attention Convolution Network(EACN)that can effectively aggregate pedestrian features.EACN can explicitly expand the semantic information of the vertex features by using the edge features between the vertices,and attentively aggregate the features on the graph through the self attention mechanism;at the same time,EACN can reduce the over-fitting phenomenon of the graph neural network by properly using the triplet loss constraint.A large number of experiments are conducted on the person re-identification benchmark datasets Market1501 and DuketMTMC-reID,which fully prove the effectiveness of the EACN framework.Secondly,on the basis of EACN framework,this thesis continue to develop a person Re-ID method combining graph contrastive learning and metric learning.In this method,dual view random transform is utilized to generate a data augmentation suitable for person Re-ID.Then,based on the random graph data transformation,the representation power of Edge Attention Convolution Network(EACN)is significantly improved through the implicit model ensemble of multiple graph convolution paradigms.Ablation study on Market-1501 and DuketMTMC-reID prove the effectiveness of the proposed method. |
