Abstract: Aimed at the problems that the recognition accuracy and model performance cannot be satisfied by multi-modal data fusion method for human interaction behavior recognition, a human interaction behavior recognition method based on improved spatial temporal graph convolutional network is proposed. The single-modal skeleton data was introduced into the cascaded densely spatial temporal graph convolutional block network to obtain rich spatial-temporal feature information and improve the feature reuse rate. An enhanced spatial temporal convolution network (EST-GCN) unit was designed to improve the information representation ability of the network between joints. A motion characteristic factor was introduced to measure the importance of different joints in the limbs to improve the model recognition effect. The experimental results on the Kinetics dataset and the case-handling area scene dataset show that the proposed method has certain advantages in the recognition effect, and the method is very competitive in model complexity and operating efficiency.