Head parallax tracking for control of a virtual space: a comparison of algorithms

Abstract

Describes and compares five techniques for vision-based tracking of a subject's head in front of a desktop PC. These algorithms are based on color, motion magnitude, background subtraction, face detection, and a probabilistic fusion of the first three algorithms. Head tracking is used to determine the subject's position with respect to the monitor noninvasively. This information is used, in turn, to control a 3D virtual space in which the viewpoint changes according to the user's head position. All of the algorithms are analyzed in terms of their qualitative strengths and weaknesses. Results compared with ground-truth collected from a Polhemus tracking device verify our analyses. It comes as no surprise that probabilistic fusion of simple elements is shown to outperform any single method, but at the cost of additional computational expense. For environments that can be constrained, color-based tracking or motion-based tracking tends to provide the most stable and most accurate state estimates of subject head position.