The thesis introduces the components of a hand gesture recognition system comprising hand segmentation, feature extraction and gesture recognition by Hidden Markov Models (HMMs). The major research focusses on the applications of HMMs to Gesture Recognition.
Two applications are developed. They are Six gesture variation and the 26 letter input system. Both perform well and achieve high recognition performance.
Each part of the system is described, including hand segmentation, feature computation, tracking methods, active statistical models, Principal Component Analysis, hand edge detection with fingertip landmarks, and vector Quantization for discrete output.
The emphasis is on the analysis of Hidden Markov Models. There are three types of HMM model structures examined — Fully Connected, Left-Right and Left-Right Banded. Two training methods are compared — the traditional Baum-Welch and the more recent Viterbi Path Counting method. The effect of initial model parameters on the system is also investigated, and a direct computation method for the HMM models parameters is explored. Two special shape gestures (Triangle and Square) are used to enhance the understanding of the learning and recognizing process of the Hidden Markov Models.
Finally, some extensions of HMMs from one-dimension to two- dimensions are examined.