Adaptive Models for the Recognition of Human Gesture

Tomorrow's ubiquitous computing environments will go beyond the keyboard, mouse and monitor paradigm of interaction and will require the automatic interpretation of human motion using a variety of sensors including video cameras. I present several techniques for human motion recognition that are inspired by observations on human gesture, the class of communicative human movement. Typically, gesture recognition systems are unable to handle systematic variation in the input signal, and so are too brittle to be applied successfully in many real-world situations. To address this problem, I present modeling and recognition techniques to adapt gesture models to the situation at hand. A number of systems and frameworks that use adaptive gesture models are presented. First, the parametric hidden Markov model (PHMM) addresses the representation and recognition of gesture families, to extract how a gesture is executed. Second, strong temporal models drawn from natural gesture theory are exploited to segment two kinds of natural gestures from video sequences. Third, a realtime computer vision system learns gesture models online from time-varying context. Fourth, a realtime computer vision system employs hybrid Bayesian networks to unify and extend the previous approaches, as well as point the way for future work. Thesis Supervisor: Aaron F. Bobick Title: Associate Professor College of Computing, Georgia Institute of Technology Thesis Supervisor: Bruce M. Blumberg Title: Assistant Professor of Media Arts and Sciences Program in Media Arts and Sciences Adaptive Models for the Recognition of Human Gesture by Andrew David Wilson The following people served as readers for this thesis: