Action recognition using weighted three-state Hidden Markov Model

Hidden Markov Model (HMM) based human action recognition (HAR) has been broadly adopted by HAR community. However, existing works donpsilat pay attention to the relationship between the layout of the model and the property of human action. In this paper, a novel HAR method is proposed based on the assumption that human action can be essentially recognized by three key postures located around the initial, middle and terminal action period. Rested on this hypothesis, we improve the HMM-based HAR method. The main contributions are twofold: (1) human action is modeled as non-return three-state HMM; (2) output probabilities in each state are weighted in terms of the key postures position in an action period. Experiments on publicly available human action database show the approach constitutes a suggestive proof for the close relationship between human action property and the design of HMM for HAR task.