Description

This tech report describes a new approach to the estimation theoretic form of decentralised data fusion (DDF). The previous state-of-the-art for optimal linear-Gaussian estimation is based on channel filters, which is very limiting. This report describes a delayed-state approach, which makes possible optimal DDF without channel filters or topology constraints. The report presents the essential mathematics and, hopefully, indicates the rich possibilities of this approach. It also presents a simulation to demonstrate a simple implementation; but the reader must realise that this implementation is a very preliminary form of DS-DDF - merely a proof of concept that DS-DDF is possible over ad hoc networks - and much more efficient and elegant variations are possible.

The key concept of the report is as follows. With delayed-states the information-form Gaussian is additive and separable for state augmentation and fusion operations. This allows pieces of the joint state to be cut up, communicated between nodes, and reconnected like a jigsaw puzzle. Alterations to the joint, by fusion or augmentation, can occur locally and be passed throughout the network. With appropriate bookkeeping and judicious use of marginalisation, the possible variations and applications of DS-DDF are large and exciting indeed.

Abstract

Delayed-state decentralised data fusion (DS-DDF) is proposed as a general methodology for consistent DDF, which does not impose constraints on network topology. The resulting estimates, although lagged in time, are optimal, equal to a centralised solution. The method is demonstrated in the context of dynamic node tracking and localisation, where a team of mobile robots track each other's position to obtain a joint estimate of the position of every team member.

Download

Full paper [pdf] (201 kb)