Automatic Extraction of Semantic Action Features

With the development of the technology like 3D specialized markers, we could capture the moving signals from marker joints and create a huge set of 3D action MOCAP data. The more we understand the human action, the better we could apply it to applications, e.g., action recognition (security), animation (sport, 3D cartoon movies, and virtual world), analysis of sports, game etc. In order to find the semantically representative features of human actions, we propose the semantic annotation approach of the human motion capture data and use the relational feature concept to extract automatically a set of action features as spatial information. For each action class, we propose a statistical method to further extract the common sets as temporal sequences of above from spatial information. The final knowledge extracted is used to recognize the action. In our experiments, we show that the knowledge extracted by this method achieves very high accuracy in recognizing actions on testing data-set with only few of training samples.