Fusion of Human Posture Features for Continuous Action Recognition

This paper presents a real-time or online system for continuous recognition of human actions. The system recognizes actions such as walking, bending, jumping, waving, and falling and relies on spatial features computed to characterize human posture. The paper evaluates the utility of these features based on its joint or independent treatment within the context of the Hidden Markov Model (HMM) framework. A baseline approach wherein disparate spatial features are treated as an input vector to trained HMMs is used to compare three different independent feature models. In addition, an action transition constraints is introduced to stabilize the developed models and allow for continuity in recognized actions. The system is evaluated across a dataset of videos and results reported in terms of frame error rate, the frame delay in recognizing an action, action recognition rate, and the missed and false recognition rates. Experimental results shows the effectiveness of the proposed treatment of input features and the corresponding HMM formulations.