June 5, 2006
Allen Tannenbaum
Title: Statistical models for contour tracking
We consider the problem of sequentially segmenting an object(s) or more generally a "region of interest" (ROI) from a sequence of images. This is formulated as the problem of tracking (causally estimating) the boundary contour of a moving and deforming object(s). The observed image is assumed to be a noisy and possibly nonlinear function of the contour. The image likelihood given the contour ("observation likelihood") is often multimodal (due to multiple objects or background clutter or partial occlusions) or heavy tailed (due to outliers). One practical way to deal with multimodal observation likelihoods is to use a particle filter. If the contour is represented as a continuous curve, contour deformation forms an infinite (in practice, very large), dimensional space. Particle filtering from such a large dimensional space is impractical. But often, it is fair to assume that for a certain time period, ``most of the contour deformation" occurs in a small number of dimensions. This ``effective basis" for contour deformation can be assumed to be fixed or slowly time varying. We have proposed practically implementable particle filtering algorithms under both these assumptions. This formulation leads to a large number of new practical problems of what is a good finite basis for a given application, how to detect the need to change the basis dimension or basis directions, how to estimate the new basis and how to deal with errors in basis estimation.