Sway detection in human daily actions using Hidden Markov Models

This paper proposes a novel method for computer vision-based, marker-less analysis of daily human actions for detecting motion irregularities (sway). Sway occurs due to a temporary loss in balance and is an important indicator of decay in motor skills. One should note that the purpose of the proposed approach is not to recognize the performed activity (which is a controlled variable in our experimental design), but to detect irregularities in the performance of this activity. The proposed motion model is based on population Hidden Markov Models. This model has been trained and tested on a custom-designed database involving multiple daily actions. Experimental results demonstrate its robustness with respect to subject and speed variability in training sequences, as well as its ability to capture sway-type motion irregularities.