Combining global features and local features is one way to improve person re-identification accuracy. Existing algorithms usually extract features from specific semantic regions of the human body. Since the human body structure is not taken into account, the learning difficulty is increased, and the efficiency and robustness are poor in the scenes with large differences. In order to solve the above problems, this paper proposes a person re-identification algorithm based on multi granularity feature fusion, which combines global feature, local feature and human structure feature. The algorithm does not introduce any prior knowledge of human body structure. In terms of feature extraction, the average pooling and maximum pooling are used to weight the feature map to obtain strong global features. Local features are obtained by slicing the feature map. Based on the original local features, local relative features are introduced as human structural features. In terms of metrics, a multi-level supervision mechanism with triple loss and ID loss at different scales is used. Experiments on Market1501 and DukeMTMC-reID show that the Rank-1 index of the algorithm is 1.3% and 3.9% higher than Part-based Convolutional Baseline(PCB) method, and the mean Average Precision(mAP) is 5.1% and 9.8% higher than PCB method.