Human Action Recognition System based on Skeleton Data

In this paper, the proposed system aims to enhance human action recognition by using skeletal features from Kinect sensor to obtain discriminative features. Joints distance feature is used for feature extraction. Instead of using traditional (non-static) K-means, such feature is clustered based on static K-means algorithm which takes statically the initial defined centroids at the first estimates for the K centroids and reduces the randomized starting centroids at all time to increase the accuracy of postures selection. Each posture is labelled by using artificial Neural Network (ANN) which makes the system more intelligent. Recognition of human action is performed using hidden Markov Model (HMM) based on the sequence of known poses to improve performance and accuracy. The proposed system recognizes the fundamental actions (walking, sitting, standing, and bending) and evaluated on the public dataset UTKinect-Action3D. The experimental results show the better accuracy rate on the static K-means than the non-static K-means.