Gesture Recognition using Hidden Markov Models, Dynamic Time Warping, and Geometric Template Matching.

摘要

Gesture recognition is useful in many applications, including human-computer interaction, automated sign language recognition, medical applications, and many more. The main focus of this thesis is to improve the isolated gesture recognition accuracy of Hidden Markov Models (HMMs) and to provide a comparison to Dynamic Time Warping and Geometric Template Matching. These techniques are compared with single-path gestures, such as the gestures created by one hand, and different coding techniques for multi-path gestures, such as gestures created with both arms. Subsequence duration structures of user-defined gestures are important for accurate recognition, and modelling these structures has been shown to increase the recognition accuracies of HMMs. It is hypothesized that Vector Quantization is responsible for the superior performance of HMMs under specific circumstances. Other contributions of this thesis include an analysis of user-defined full-body 3D gestures, several modification to increase the accuracy of HMM models, and a multi-path template matcher.