Novel Human-Posture Recognition System Based on Advanced Graph Convolutional Network Using Skeletal Data

Automatic human-posture or human-activity recognition is a very important research problem nowadays. In this work, we propose a novel human-posture recognition approach using the three-dimensional skeletal data acquired by the Kinect V2 sensor. The acquired skeletal data are first segmented using our recently proposed automatic-segmentation technique and each segment can be labeled with a particular kind of human-posture. We propose four different types of node feature matrices extracted from the segmented skeletal data, which can serve as the input features to the advanced graph convolutional network (GCN) for multi-classification. The realworld experimental results demonstrate that our proposed novel human-posture recognition system can reach a very high average classification-accuracy of 91.56%. In addition, the ablation study of the effect of skeletal-graph variations on the recognition performance is also presented. The average classification-accuracy further reaches up to 92.33% when four confusing joint-nodes are removed from the skeletal graph. Our proposed novel human-posture recognition approach can be very useful for practical applications such as human-computer interface, intelligent healthcare, robotics, etc.