Model Structure Selection and Training Algorithms for an HMM Gesture Recognition System

Abstract

Hidden Markov models using the fully-connected, left-right and left-right banded model structures are applied to the problem of alphabetical letter gesture recognition. We examine the effect of training techniques, in particular the Baum-Welch and Viterbi path counting techniques, on each of the model structures. We show that recognition rates improve when moving from a fully-connected model to a left-right model and a left-right banded 'staircase' model with peak recognition rates of 84.8%, 92.31% and 97.31% respectively. The left-right banded model in conjunction with the Viterbi path counting present the best performance. Direct calculation of model parameters from analysis of the physical system was also tested, yielding a peak recognition rate of 92%, but the simplicity and efficiency of this approach is of interest.