Self-attention-Based Dual-Branch Person Re-identification

Person re-identification is a technique that retrieves a corresponding person from a person dataset. This technology addresses the issue of surveillance cameras being unable to capture clear facial images or failing to capture faces, relying instead on the overall physical features of a person for matching. However, in complex scenarios, such as most public places, a person's body may be partially occluded by buildings or other persons. To address the occlusion problem, this paper proposes a dual-branch method for local person re-identification, which introduces a self-attention mechanism aimed at improving the accuracy and robustness of person re-identification. This method first extracts the global and local features of a person through two parallel convolutional neural network branches. Then, it employs an attention mechanism to weigh the local features of a person, emphasizing regions which are more critical for identification. Finally, the global and weighted local features are fused to obtain the final representation of the person.