In this paper, we propose an efficient method for tracking 3D modelled
objects in cluttered scenes, Rather than tracking objects in the imag
e, our approach relies on the object recognition aspect of tracking. P
ossible matches between image and model features define volumes within
a transformation space. The model is best aligned with the image in v
olumes satisfying the greatest number of correspondences. Object motio
n defines a trajectory in the transformation space. We propose an effi
cient algorithm to compute these transformations. (C) 1998 Elsevier Sc
ience B.V. All rights reserved.