Variations on a Particle Filter

Geoff Gordon

Abstract

The particle filter is a popular algorithm for tracking the state of a mobile robot and its environment -- that is, for approximating the robot's posterior belief about the possible states of the world given the evidence from its sensors. Another popular algorithm for state tracking is the Kalman filter. Particle filters and Kalman filters have very different error properties: particle filters are good at tracking several distinct posterior modes in a low-dimensional space, while Kalman filters track only a single posterior mode but can handle higher-dimensional spaces. One might hope that there is a happy medium between the two: an algorithm which can handle a moderate number of posterior modes in a state space of moderate dimensionality. Many people have proposed candidates for this happy medium. I will explore some of these candidates, and hopefully provide experimental results on a simple people-tracking problem.


Back to the Main Page

Charles Rosenberg
Last modified: Tue Mar 12 18:00:28 EST 2002