In order to solve the problem that the feature extraction is easy to be affected by external factors and the computational complexity is high, the depth data was used for human action recognition, which is a more effective solution scheme. Using the joint data collected by Kinect, the human joint was divided into five regions. The vector angle of each region was discretized to describe different states, and then Baum-Welch algorithm was used to study multi-Hidden Markov Model (multi-HMM), meanwhile, forward algorithm was used to establish the generation region and action class probability matrix. On this basis, the region and action categories were intra-coupled and inter-coupled to analyze, thus expressing the interaction between the joints. Finally, the K-Nearest Neighbors (KNN) algorithm based on coupling was used to complete the action recognition. The experimental results show that the recognition rates of the five actions reach above 90%, and the comprehensive recognition rate is higher than that of the contrast methods such as 3D Trajecttories, which means that the proposed algorithm has obvious advantages.