| Summary: | 碩士 === 國立清華大學 === 資訊工程學系 === 101 === In this thesis, a context-dependent action interpreter for interactive storytelling is proposed. At first, a regular expression for actions, named Action Regular Expression (ARE), is proposed; in fact, ARE can be thought as an extension of traditional regular expression, new operators are defined to deal with the properties of actions. To acquire and recognize user’s body joints, a Microsoft Kinect sensor is used; then eight actions are defined and recognized via ARE. The experiment runs on six subjects and the obtained total average accuracy is 86%. To interpret user’s actions, an action interpreter is defined which contains a plan library and the classifier. The plan library with 75 similar plans is built, and due to the ability of maximizing data margins, support vector machine is applied to learn the plan library. To get the optimal parameters for the support vector machine, a set of 110 combinations are tried; at final, a trained model with 94.74% accuracy is obtained while choosing the optimal parameters.
|
|---|
