Mesh-Based Consensus Distributed Particle Filtering for Sensor Networks
Yang Liu, Matthew Coombes, Cunjia Liu
Abstract
Following the Bayesian inference framework, this paper investigates the problem of distributed particle filtering over a sensor network to achieve consensus. The objective of the posterior-consensus strategy is to fuse the posterior probability distribution functions (PDFs) at different sensor nodes, so that an agreement of belief can be established in terms of the Kullback-Leibler average (KLA). To facilitate the consensus process and reduce the communication load, the local PDFs are approximated with weighted meshes and transmitted between neighboring nodes. The mesh representations are constructed by resorting to a grid partition of the state space, such that the PDF can be approximated by a linear combination of indicator functions. To derive a particle representation of the fused PDFs, a novel importance density function is designed to draw particles with respect to the information from all neighboring nodes. The weights of the particles are calculated via the recursive solution of the KLA. The effectiveness of the proposed filtering approach is demonstrated through two target tracking examples.