In order to protect the stable development of ethnic minority dance art, this paper proposes a method of human behavior recognition based on human skeletal joint. Extracting spatial information of video frames through key joint node models, this method extracts temporal features of video sequences using temporal covariance pyramid model, effectively removes useless noise data and retains key information, thereby improving the recognition rate of actions. The experimental results show that the application effect of the method in this paper is good, the recognition results of typical ethnic minority movements are in line with the actual situation, and the recognition accuracy is relatively high. Moreover, it realizes a complete process from dance movement data collection to movement recognition, which lays the foundation for accurate and effective analysis of dance training movements.