Abstract: Person re-identification is the task of cross-device retrieval in pedestrian images. It is a technology that has important application value in fields such as video surveillance and smart security. Due to the interference of environmental factors (such as lighting, angle, occlusion, etc.) that introduce noise, the difficulty of extracting and identifying pedestrian information features is increased. To this end, this paper proposes an attention-based multi-branch joint network structure to improve the model's recognition ability. The model uses Omni-Scale Network (OSNet) as the backbone network to achieve the fusion of full-scale features, and embeds serial channel attention mechanism and position attention mechanism to enhance the model's deep semantic expression. It employs multiple loss functions to jointly supervise the training of the model and achieve the global feature extraction and output capability of pedestrian features. Simulation experimental results show that the model’s comprehensive performance of pedestrian image information feature extraction on public datasets Market1501, DukeMTMC-reID and CUHK03-Labeled(Detected) is better than similar algorithms such as DRL-Net and DCAL. The model achieves high recognition accuracy with Rank-1 values of 95.3%, 90.1% and 80.4%(Labeled)/ 78.1%(Detected)and mAP values of 89.2%, 80.47% and 78.9%(Labeled)/75.4%(Detected) on Market1501, DukeMTMC-reID and CUHK03-Labeled(Detected) datasets respectively.