AN ARCHITECTURE FOR VISION-BASED HUMAN-COMPUTER INTERACTION

摘要

Computer-vision as a mode of interaction with technical systems has found increasing interest in the past few years. In many of the proposed case studies and applications, the user's current pose is observed by the computer using cameras, and the computer's reaction is displayed to the user who changes the pose accordingly in order to reach a desired goal of interaction. A particular problem with many applications is that the constraints on the environment of interaction should be kept low, in particular concerning illumination, with the consequence that a reliable behavior of the system is hard to achieve by straightforward techniques. We present an architecture which uses time and space coherence, knowledge about key poses of interaction, and feed-back between different levels of the process of recognition, in order to eliminate errors caused by the natural environment. Furthermore, we report on experiences with its application.