A multi task gait recognition method based on multi-scale convolutional neural network and human posture estimation model is studied, explains the recognition results of neural network, and improves its recognition effect in the face of covariate changing scenes. In this method, the gait spatial features extracted by convolutional neural network and the human posture estimation model are fused with the temporal features of human joints for identity recognition. Experiments were carried out using normal and synthetic walking sequence data in gait dataset CASIA-B and TUM-GAID gait dataset. The results show that the recognition rates of the three scenes T1, T2 and T3 are 95.2%, 72.4% and 84.5% respectively. In the experiment of CASIA-B gait dataset, for normal walking sequences and two kinds of synthetic walking sequences, this method has good performance in recognition accuracy, which reflects that the model has strong robustness.