Abstract: Aiming at the problems of low signal-to-noise ratio and insufficient feature information of current bone data, this paper applies a human action recognition method combining most informative joints with spatiotemporal pyramid. It extracted most informative joints of the human skeleton to construct spatial domain pyramid features, and used multi-level superimposed covariance to construct temporal pyramid features, which not only preserved the spatial structure of the skeleton hinge system, but also solved the problem of preprocessing the length of the video sequence. Experimental results on the MSR-Action3D and UTKinect dataset show that the method has high accuracy and good real-time performance, and can be widely applied in various fields of behavior recognition.