A Comparison of Dynamic Naive Bayesian Classifiers and Hidden Markov Models for Gesture Recognition

in this paper we present a study to assess the performance of dynamic naive bayesian classifiers (dnbcs) versus standard hidden markov models (hmms) for gesture recognition. dnbcs incorporate explicit conditional independence among gesture features given states into hmms. we show that this factorization offers competitive classification rates and error dispersion, it requires fewer parameters and it improves training time considerably in the presence of several attributes. we propose a set of qualitative and natural set of posture and motion attributes to describe gestures. we show that these posture-motion features increase recognition rates significantly in comparison to motion features. additionally, an adaptive skin detection approach to cope with multiple users and different lighting conditions is proposed. we performed one of the most extensive experimentation presented in the literature to date that considers gestures of a single user, multiple people and with variations on distance and rotation using a gesture database with 9441 examples of 9 different classes performed by 15 people. results show the effectiveness of the overall approach and the reliability of dnbcs in gesture recognition.