Graph Laplacian distributed particle filtering

Michael G. Rabbat, Mark Coates, Stephane Blouin

Abstract: We address the problem of designing a distributed particle filter for tracking one or more targets using a sensor network. We propose a novel approach for reducing the communication overhead involved in the data fusion step. The approach uses graph-based signal processing to construct a transform of the joint log likelihood values of the particles. This transform is adaptive to particle locations and in many cases leads to a parsimonious representation, so that the the joint likelihood values of all particles can be accurately approximated using only a few transform coefficients. The proposed particle filter uses gossip to perform distributed, approximate computation of the transform coefficients. Numerical experiments highlight the potential of the proposed approach to provide accurate tracks with reduced communication overhead.