Performance Analysis of Human Action Recognition System between Static k-Means and Non-Static k-Means

In this paper, the human actions are classified in skeleton data from Kinect sensor based on joint distance features. To raise the accuracy rate of postures analysis, the proposed system uses the static k-means algorithm that it takes the static initial K centroids at the first estimates instead of using the non-static (traditional) k-means that it takes the randomized starting centroids at all time. Further, to improve the performance and accuracy, artificial neural network (ANN) is applied to label the classes of the human poses and discrete Hidden Markov Model (HMM) is also used to correctly recognize the human actions based on the sequence of known poses. Experiments with two different datasets (public dataset UTKinect and New dataset) show that the proposed approach produces good performance results and accuracy rate from the action class models.