Auxiliary Particle Filtering With Multitudinous Lookahead Sampling for Accurate Target Tracking

Abstract

The auxiliary particle filter, which is the popular extension of the standard bootstrap particle filter, is known to assist in drawing particles from regions of high probability mass of the posterior density by leveraging the incoming measurement information in the sampling process. The filter accomplishes this by looking ahead in time to determine those particles that become important when propagated forward, retract, and then propagate those particles forward in time. The key problem with this approach is that a particle determined to be important may not fall in regions of importance when actually propagated forward, either because of a large diffusion of the state transition kernel and/or a highly informative measurement, thus defeating the entire purpose of the filter. This problem leads to degeneracy. This paper proposes a method of sampling a multitude of particles for each particle to make such a decision. The key idea here is to use multiple disturbances, instead of one as does the auxiliary particle filter, as lookahead means to guide particles to regions of high probability in the posterior probability density. Through evaluation, we show that the proposed idea overcomes the said problem and exhibits less degeneracy and high tracking accuracy.