Abstract: The key of person re-identification is to extract distinctive pedestrian features. Based on ResNet-50, this paper proposes a multi-branch network structure (MMNet) which uses multiple methods to extract features. By combining various methods in a clever way, the problem of lack of pedestrian discrimination feature information caused by pedestrian posture change and partial occlusion can be solved. In the first branch, the global features of pedestrians are extracted. In the second branch, the local features that person re-identification want to pay attention to are extracted by the channel attention module. In the third branch, the features extracted from backbone are evenly divided into different blocks, so as to extract the local features with different granularity. Then the model is trained by using the bath hard triple loss function and softmax loss function. Finally, the features extracted from different branches are concatenated as the final features. By complementing the global features and various local features of pedestrians, distinctive pedestrian feature is extracted. The map and rank 1 of the algorithm on Market-1501 and DukeMTMC-reID datasets are 87.7% and 95.9%, 79.9% and 89.2% respectively. The experimental results show that the features extracted by the multi-branch network are complementary, and map and rank1 are higher than that of most person re-identification algorithms.