Arthur Pece and Anthony Worrall
Computational Vision Group, Department of Computer Science, The University of Reading, UK.
The tracking method presented in this paper is based on a pose-refinement algorithm which is a variant of active contours. The most important difference between our method and most active contour methods is that no "features" are detected at any stage: the evaluation function for the pose parameters is based on all the grey-level information extracted from the normals to the model contours, without thresholding. One advantage of the feature-free evaluation function is that it is a smooth function of both the pose parameters and the image grey levels. Another advantage is that it leads to a covariance estimate for the Bayesian evidence for the pose parameters. This covariance estimate is used for efficient pose optimisation by a Newton-like method and for proper weighting of the innovation in a Kalman filter. The method is demonstrated by tracking cars with 3-D models.