In view of the fact that the existing person re-identification methods are difficult to avoid inaccurate feature extraction caused by environmental noise and are easily mistaken for person features, a person multi-feature fusion branch network based on dynamic convolution and attention mechanism is proposed. Firstly, due to the uncertain factors such as illumination change, human posture, object occlusion, etc., dynamic convolution is proposed to replace static convolution in ResNet50 to obtain a more robust Dy-ResNet50 model. Secondly, considering that there are great differences in the perspective of taking pictures of people and people are blocked by objects, it is proposed to embed the self-attention mechanism and the cross-attention mechanism into the backbone network. Finally, the cross entropy loss function and the difficult sample ternary loss function are used as the loss function together.Experiments are carried out on DukeMTMC-ReID、 Market-1501 and MSMT17 public datasets and compared with mainstream network models. The results show that the Rank-1 and mAP of the proposed model is 0.9 and 1.6 percentage points higher than that of the current mainstream model on the data set DukeMTMC-ReID.