Recognition of Hand Gestures Tracked by a Dataglove: Exploiting Hidden Markov Models Discriminative Training and Environment Description to Improve Recognition Performance

Sequence classification based on Hidden Markov Models (HMMs) is widely employed in gesture recognition. Usually, HMMs are trained to recognize sequences by adapting their parameters through the Baum-Welch algorithm, based on Maximum Likelihood (ML). Several articles have pointed out that ML can lead to poor discriminative performances among gestures. This happens because ML is not optimal for this purpose until the modellized process is actually an HMM. In this paper we present a gesture recognition system featuring a discriminative training algorithm based on Maximal Mutual Information (MMI) and the integration of environment information. The environment is described through a set of fuzzy clauses, on the basis of which a priori probabilities are computed. Adaptive systems such as unsupervised neural networks are used to build a codebook of symbols representing the hand’s states. An experiment on a set of meaningful gestures performed during the interaction with a virtual environment is then used to evaluate the performance of this solution