Effect of initial HMM choices in multiple sequence training for gesture recognition

We present several ways to initialize and train hidden Markov models (HMMs) for gesture recognition. These include using a single initial model for training (re-estimation), multiple random initial models, and initial models directly computed from physical considerations. Each of the initial models is trained on multiple observation sequences using both Baum-Welch and the Viterbi path counting algorithm on three different model structures: fully connected (or ergodic), left-right, and left-right banded. After performing many recognition trials on our video database of 780 letter gestures, results show that a) the simpler the structure is, the less the effect of the initial model, b) the direct computation method for designing the initial model is effective and provides insight into HMM learning, and c) Viterbi path counting performs best overall and depends much less on the initial model than does Baum-Welch training.