Pre-classification based hidden Markov model for quick and accurate gesture recognition using a finger-worn device

Hidden Markov Model (HMM)-based recognition methods are very commonly used for some applications and can be highly accurate. However, they have a high computational complexity that creates problems when they are used for gesture recognition on resource-constrained wearable devices. In this paper, we propose a pre-classification method to reduce recognition complexity by dividing gesture vocabularies into groups, and maintain, even improve, the recognition accuracy by adaptively adjusting the HMMs for different groups. The technique consists of three tasks: gesture grouping, group modeling, and gesture modeling. Gesture grouping is performed using a K-means++ algorithm; the groups are modeled using a table-based method; and the gestures are modeled using an HMM-based approach. We evaluated the pre-classification method using real data collected by a tiny finger-worn device called a Magic Ring. The complexity of our method is much less than the standard Hidden Markov Model, without any loss of accuracy.