부산대학교 도서관

In this paper we present a deep learning scheme for classification of dance postures using Kinect II RGB data and Convolutional Neural Networks (CNN). This is achieved through the analysis of a data-set that includes three traditional Greek dances, where each dance was performed by 3 different dancers. The obtained data were processed and analyzed using a deep convolutional neural network, in order to identify the primitive postures that comprise the choreography. To enhance the classification performance, a background subtraction framework was utilized, while the CNN architecture was adapted to simulate a moving average behavior. The overall system can be used as an AI module for assessing the performance of users in a serious game for learning traditional dance choreographies