Hidden markov model for human performance modeling
Jie Yang · 1994
Human performance is the actions and/or reactions of humans under specified circumstances. Actions reflect human skill in performing a task, and reactions reflect control strategy in response to the environment. Many human-computer interactions require modeling human performance. Human performance modeling, however, is challenging because of its stochastic nature and relative measure. In this dissertation, I propose a systematic methodology for modeling human performance based on the hidden Markov model (HMM) approach. To illustrate the concepts, procedures, and implementational issues, three case studies are carried out in great detail on gesture recognition, action learning, and control strategy learning. First, the problem of gesture recognition is investigated. A prototype of an HMM-based gesture recognition system has been developed to demonstrate the feasibility of the proposed method, to illustrate the procedures, and to address implementational issues. The system has achieved 99.78% accuracy for an isolated recognition task with nine gestures, and shows great potential for connected gesture recognition. The problem of action learning from observation is then studied. An HMM has been used to represent all the observed action data based on the most likely performance criterion and to acquire skill from these data. The HMM-based action learning approach has been successfully applied to the Self-Mobile Space Manipulator ($SM\sp2$) for learning the skill of exchanging an Orbit Replaceable Unit (ORU). Finally, the problem of control strategy learning is discussed. For a given system, the control strategy is partitioned into a set of decision patterns which are described by corresponding HMMs. Simulation results on a linear system and an inverted pendulum system are given. HMM provides a framework for modeling human actions and control strategy, to cope with stochastic nature of human performance. The concepts and systems developed in this dissertation are significant for various problems in man-machine interactions such as teleoperation, learning control, and skill learning. The approach is also a practical solution in developing a knowledge-based intelligent system to provide high quality performance. The technology presented here is extendible to a variety of other problems in man-machine interactions.