Abstract: In order to extract human motion information efficiently and improve the accuracy of action rec- ognition from videos, an approach for action recognition based on human pose spatio-temporal features is proposed. Firstly, with the joint positions of human body in each frame of the video acquired, we extracted pose information by handcrafted features. Specifically, the positions of joints and relatives in the spatial di- mension, as well as the change of that in the temporal dimension were calculated. The two together consti- tuted human pose spatiotemporal feature descriptors. Then the Fisher Vector model was utilized to compute fixed dimension Fisher vector for each descriptor separately. Lastly, features were weighted to fusion for classification. Experimental results show that the proposed algorithm can effectively improve action recog- nition performance.