A probabilistic approach to gestural recognition and dialogue management.

Abstract

This dissertation presents a new probabilistic approach to the handling of gestural recognition. The new Bayesian model (INCA) is introduced and various aspects of its performance are examined. This model is a meta-algorithmic approach and allows a variety of different gestural recognition techniques to be combined in such a way as to supplement one another's capabilities to produce better overall recognition rates. Results of testing on this model indicate that it can reduce system design time and can provide effective recognition algorithms in a relatively short design period. The dissertation also examines a new model for handling dialogue management in gestural interfaces. This model, Probabilistic Finite State Machines differs from traditional approaches to dialogue management in that it views the non-deterministic nature of gestural recognition as a normal aspect of the dialogue between user and system. Unlike traditional approaches, this model treats recognition errors as a natural, expected part of any dialogue and is designed to address such errors in a natural way.