Michael Isard and Andrew Blake
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.
In Proc 5th European Conf. Computer Vision