| Person Re-Identification(ReID)aims to retrieve a target pedestrian from the pedestrian images captured by cameras in different scenes.It is considered as a sub-problem in the field of image retrieval.ReID can be applied in multiple fields such as intelligent security,intelligent supermarkets,album clustering,human-machine interaction,etc.Therefore,it has practical research and application value.Due to light changes,pedestrian posture changes,different shooting angles,occlusion,low camera resolution and other problems,ReID has become a challenging task.Most of the existing ReID studies focus on how to extract the more discriminative feature expressions,while ignoring the potential relationship between multiple pedestrian images.A single image is prone to lose some information due to occlusion,blurring and other problems.ReID datasets usually contains multiple images of the same pedestrian,so the missing information of a single image is expected to be compensated from other images.Therefore,constructing connections between multiple pedestrian images may be effective.From this point of view,this thesis designs an Adaptive Graph Attention Model and a Deformable Transformer Based Model.We use graph attention networks and sampling points based attention mechanism to construct relationships between multiple images,in order to obtain more discriminative pedestrian feature expressions.Numerous experiments have shown that both models can effectively fuse features from multiple pedestrian images and improve recognition accuracy.The main research contents of this thesis are as follows:1.This thesis proposes an Adaptive Graph Attention Network.It takes the features of a batch of input pedestrian images as the vertex set of the graph.The edge set can be determined based on the similarity relationship between the vertices.On this basis,two graph attention mechanisms are designed and added to the graph to construct a complete adaptive graph attention model.Experiments have shown that this model can effectively aggregate features from multiple pedestrian images and improve the discrimination of pedestrian features.2.In view of the effectiveness of local feature based methods for ReID task,this thesis proposes a Two-Branch Attention Mechanism based on Transformer structure,including a global branch and a local branch.The global branch is used to extract the global feature representation of a single pedestrian image.And the Masked Cross-Attention Mechanism is designed in the local branch to extract several local features.In order to separate the local features from each other as much as possible while ensuring the discrimination,so as to focus on different areas of the pedestrian image,local loss is introduced to constrain the model.3.On the basis of the Two-Branch Attention Mechanism,this thesis designs a Deformable Transformer structure,which uses the multi-head self-attention mechanism based on sampling points proposed in this thesis.It can efficiently fuse the global and local features of multiple pedestrian images,and obtain more discriminative pedestrian feature representations.4.In this thesis,a large number of ablation and contrast experiments have been made on several large public ReID datasets such as DukeMTMC-ReID,and the results fully demonstrate the effectiveness of the algorithms.Compared with some of the most advanced ReID algorithms known to us at present,the models proposed in this thesis have better performance. |
