Statistical Background Modelling for Tracking with a Virtual Camera.

Abstract

A method of robust feature-detection is proposed for visualtracking with a pan-tilt head. Even with good foreground models,the tracking process is liable to be disrupted by strong featuresin the background. Previous researchers have shown that thedisruption can be somewhat suppressed by the use ofimage-subtraction. Building on this idea, a more powerfulstatistical model of background intensity is proposed in which aGaussian mixture distribution is fitted to each of the pixels on a"virtual" image plane. A fitting algorithm of the"Expectation-Maximisation" type proves to be particularly effectivehere. Practical tests with contour tracking show markedimprovements over image subtraction methods. Since the burden ofcomputation is off-line, the online tracking process can run inreal-time, at video field rate. (Author abstract) 26 Refs.