Abstract: A multi-granularity person re-identification model integrating attention mechanism is proposed to address the complex monitoring environment and the significant differences in image appearance of pedestrians under different conditions such as lighting changes, perspective changes, and occlusion, which make it difficult to extract detailed pedestrian features. The model extracts feature maps containing information at different scales by introducing a multi-branch structure. Combining multi-granularity segmentation modules and the attention mechanism, the local discriminative information of the feature map is further extracted to obtain diverse feature representations and achieve the coordination and unity of features. The model is supervised and trained by using federated learning to obtain a more comprehensive feature description. Excellent performance has been achieved on the mainstream person re-identification datasets Market-1501, DukeMTMC-reID, and CUHK03, with mAP reaching 88.42%, 78.86%, and 76.70%, respectively, demonstrating the effectiveness of the proposed model.