Particle Filter with Efficient Importance Sampling and Mode Tracking (PF-EIS-MT) and its Application to Landmark Shape Tracking

We develop a practically implementable particle filtering (PF) method called "PF-EIS-MT" for tracking on large dimensional state spaces. Its application to tracking the shape change of a large number of "landmark" (feature) points from image sequences is shown. Two issues common to most large dimensional problems are (a) observation likelihood is often multimodal and the state transition prior is often broad in at least some dimensions and (b) direct application of PF requires an impractically large number of particles. PF-EIS-MT combines the advantages of two recently proposed ideas which address both of these issues. Improved performance of PF-EIS and PF-EIS-MT over existing PF algorithms is demonstrated for landmark shape tracking.