Human Action Recognition Based on the Adaptive Weighted Dynamic Time Warping Algorithm

The skeleton information of human action could be extracted by Kinect sensor, and it was also a hot topic to identify the action based on them. So, an action recognition method based on adaptive weighted dynamic time warping algorithm was proposed. In the method, firstly, the skeleton joints' coordinates were obtained from the action by Kinect, and then they were normalized. Secondly, the weights were assigned to joint based on inter-intra class joints' variability and were introduced into the distance computation in the adaptive weighted dynamic time warping algorithm. Finally, voting method was used to recognize the action. Experimental results show that the presented method has a high performance for action recognition in the MSR Action3D database and it is efficient.