Abstract: More and more researchers have begun to study the human action recognition based on depth information and skeleton information since the Kinect has been released. A method of human action recognition based on the skeleton feature of key frames is proposed in order to improve the accuracy and timeliness of the human action recognition, and reduce the computational complexity. The clustered data was obtained by using K-means clustering algorithm, and then the key frames were extracted by using the clustered data. Two features for human action recognition were extracted, one is the feature of the position of human joint, another is the feature of the skeleton angle between rigid body and rigid body. The sequence of action video was classified through the SVM classifier. This method leads to a more accurate recognition rate, and the real-time capability has been improved at the same time according to the result showed on the data set MSR-DailyActivity3D.

Key words: Kinect, human action recognition, key frames, k-means