Arm-hand Action Recognition Based on 3D Skeleton Joints

Abstract

In this paper, a simple and reliable method for arm-hand action recognition is proposed based on the representation of 3D skeleton joints. It mainly consists of two parts, feature extraction and classifier design. First, through the whole 3D skeleton joints collected by Kinect sensor, a hand joint, which is selected to represent the arm-hand action, is fed to a k-means cluster to decrease the number of features. Meanwhile, a data preprocessing of rotation is employed to deal with the multi-view problem. Then, the bag of words model is applied to form codebooks for each kind of arm-hand action. Finally, arm-hand action recognition is implemented by training SVMs using different kernel functions and parameters. The method was tested and compared with some state-of-the-art approaches on a self-built dataset for action recognition. The results manifested the effectiveness of the proposed method.