A smoothing filter for Condensation
Michael Isard and Andrew Blake
Proc 5th European Conf. Computer Vision, Vol. 1 767-781, 1998
Abstract
Condensation, recently introduced in the computer vision literature, is a
particle filtering algorithm which represents a tracked object's
state using an entire probability distribution. Clutter can cause
the distribution to split temporarily into multiple peaks, each
representing a different hypothesis about the object configuration.
When measurements become unambiguous again, all but one peak,
corresponding to the true object position, die out. While several
peaks persist estimating the object position is problematic.
`Smoothing' in this context is the statistical technique of
conditioning the state distribution on both past and future
measurements once tracking is complete. After smoothing, peaks
corresponding to clutter are reduced, since their trajectories
eventually die out. The result can be a much improved
state-estimate during ambiguous time-steps. This paper implements
two algorithms to smooth the output of a Condensation filter. The
techniques are derived from the work of Kitagawa, reinterpreted in
the Condensation framework, and considerably simplified.
Compressed (gzip) postscript
version of the paper.
misard@robots.ox.ac.uk