부산대학교 도서관

This paper proposes a novel automatic posture recognition approach using the skeletal data of human subjects acquired from the Kinect sensors. The acquired skeletal data are used as the input features for training the artificial-intelligence driven recognizer. In this work, we formulate the underlying human-posture recognition problem as the classical multi-classification problem. The graph convolutional network (GCN) is trained to identify the human postures by successive frames through an activity using the Kinect skeletal data (three-dimensional skeletal coordinates). Experimental results using realworld data demonstrate that our proposed GCN leads to a promising classification-accuracy of 92.2% for automatic human-posture recognition. As a result, our proposed novel GCN-based human-posture recognizer greatly outperforms other existing schemes.