Tracking using explanation-based modeling

We study the problem of tracking, namely, estimating the states of physical objects with time, from streams of noisy and unreliable observations. The most common model for the tracking problem is the generative model, which is the basis of solutions such as the Kalman filter and particle filters. In this paper, we consider a different formulation -- an {\em explanatory} framework -- for tracking, and we provide a tracking algorithm based on our formulation. We provide experimental results which compare our algorithm to particle filters on simulated data, and finally, we describe an implementation of our algorithm in a real-world scenario, namely, tracking faces in video segments. One surprising outcome of the simulations is that the new algorithm outperforms particle filters in high noise situations even when the particle filter is based on the correct likelihood function.