Dynamic Gesture Recognition using Transformation Invariant Hand Shape Recognition

In this thesis a detailed framework is presented for accurate real time gesture recognition. Our approach to develop a hand-shape classifier, trained using computer animation, along with its application in dynamic gesture recognition is described. The system developed operates in real time and provides accurate gesture recognition. It operates using a single low resolution camera and operates in Matlab on a conventional PC running Windows XP. The hand shape classifier outlined in this thesis uses transformation invariant subspaces created using Principal Component Analysis (PCA). These subspaces are created from a large vocabulary created in a systematic maimer using computer animation. In recognising dynamic gestures we utilise both hand shape and hand position information; these are two o f the main features used by humans in distinguishing gestures. Hidden Markov Models (HMMs) are trained and employed to recognise this combination o f hand shape and hand position features. During the course o f this thesis we have described in detail the inspiration and motivation behind our research and its possible applications. In this work our emphasis is on achieving a high speed system that works in real time with high accuracy. Glossary of Acronyms ASL American Sign Language BSL British Sign Language CHMM Continuous Hidden Markov Models CSL Chinese Sign Language DHMM Dynamic Hidden Markov Models EM Expectation-Maximization IICl Human Computer Interaction HMM Hidden Markov Models ISL Irish Sign Language LBP Linear Binary Patterns MDA Multiple Discriminant Analysis PCA Principle Component Analysis CHAPTER