Online action recognition based on skeleton motion distribution

Online action recognition which aims to jointly detect and recognize actions from video streams, is an essential step towards a comprehensive understanding of human behavior. However, it is challenging to accurately locate and recognize the occurrence of actions from noisy data streams. This paper proposes a skeleton motion distribution based method for effective online action recognition. Specifically, an adaptive density estimation function is built to calculate the density distribution of skeleton movements. Observing that each action has a unique motion distribution, we detect the occurrence of actions by identifying the transition of the motion distribution in a video stream. Once the starting point of an action is detected, a snippet-based classifier is proposed for online action recognition, which continuously identifies the most likely action class. Experimental results demonstrate that our method outperforms the state-of-the-art methods in terms of both detection accuracy and recognition precision.