Untethered human motion recognition for a multimodal interface

Abstract

This thesis used machine learning techniques to extract useful information about human body articulations. First, it presents a learning approach to model non-linear constraints; a support vector classifier is trained from motion capture data to model the boundary of the space of valid poses. Next, it proposes a system that incorporates body tracking and gesture recognition for an untethered human-computer interface. The detection step utilizes an SVM to identify periods of gesture activity. The classification step uses gesture-specific Hidden Markov Models (HMMs) to determine which gesture was performed at any time period, and to extract the parameters of those gestures. Several experiments were performed to verify the effectiveness of these techniques with encouraging results.