Fast visual tracking by temporal consensus

At the heart of every model-based visual tracker lies a pose estimation routine. Recent work has emphasized the use of least-squares techniques which employ all the available data to estimate the pose. Such techniques are, however, susceptible to the sort of spurious measurements produced by visual feature detectors, often resulting in an unrecoverable tracking failure. This paper investigates an alternative approach, where a minimal subset of the data provides the pose estimate, and a robust regression scheme selects the best subset. Bayesian inference in the regression stage combines measurements taken in one frame with predictions from previous frames, eliminating the need to further filter the pose estimates. The resulting tracker performs very well on the difficult task of tracking a human face, even when the face is partially occluded. Since the tracker is tolerant of noisy, computationally cheap feature detectors, frame-rate operation is comfortably achieved on standard hardware.