As a type of biometric technology, gait recognition makes use of the physical gestures made by a walking person to identify it. It has the benefit of a long recognition range, doesn’t require the subject’s cooperation, and has a lot of potential applications in public safety and security monitoring. Nevertheless, it is simple to overlook the interframe variations of gait sequences while extracting gait features in the temporal dimension, which may result in a decline in recognition accuracy. Consequently, one of the most difficult issues in the field of gait recognition is figuring out how to enhance the temporal feature extraction capacity of gait models. In this paper, we designed a network MST-Gait which makes full use of interframe details to realize gait recognition. MST-Gait extract the spatio-temporal aspects of gait by use of spatial graph convolution, Multi-scale Temporal Module (MSTM) and attention mechanism through the residual structure. To better utilize the information in the channel dimension, this paper introduces Batch Channel Normalization (BCN) in the network. Experimental results on the commonly used dataset CASIA-B show that the accuracy of the model in this paper gains some improvement compared to previous results.