Attractor-directed particle filtering with potential fields

In this paper we describe a particle filter algorithm that allows incorporation of prior knowledge about future states. Incorporation of such knowledge can significantly reduce the uncertainty in the estimation of future state predictions. Estimation of the state is based on a transition model where the current state is not only conditioned on the previous state but also on an attractive potential field, exercising an attractive force on the particles, causing them to travel a certain path towards an attractor point. With the help of dynamic Bayesian networks we can show that the attractor-directed particle filter algorithm is correct if the number of particles used for the estimation approaches infinity. Through an experiment we show how prior knowledge of future states can successfully be used to improve the estimation performance of the attractor-directed particle filter algorithm.