Gesture modelling and recognition by integrating declarative models and pattern recognition algorithms

Abstract

Gesture recognition approaches based on computer vision and machine learning mainly focus on recognition accuracy and robustness. Research on user interface development focuses instead on the orthogonal problem of providing guidance for performing and discovering interactive gestures, through compositional approaches that provide information on gesture sub-parts. We make a first step toward combining the advantages of both approaches. We introduce DEICTIC, a compositional and declarative gesture description model which uses basic Hidden Markov Models (HMMs) to recognize meaningful pre-defined primitives (gesture sub-parts), and uses a composition of basic HMMs to recognize complex gestures. Preliminary empirical results show that DEICTIC exhibits a similar recognition performance as “monolithic” HMMs used in state-of-the-art vision-based approaches, retaining at the same time the advantages of declarative approaches.