Graphical Models for Nonlinear Distributed Estimation

Distributed estimation has advantages over centralized estimation in reducing communication bandwidth, distributing the processing load and improving system survivability. One important technical issue in designing distributed estimation architectures and algorithms is the proper treatment of dependent information. This paper presents graphical models to represent dependent information in general distributed estimation problems. It reviews the use of information graphs to represent dependence due to communication among processing agents so that common information can be identified to avoid double counting in fusion. It introduces Bayesian networks to represent conditional independence of measurements given the system states and recognize the minimal set of random variables that satisfy the conditional independence assumption. Distributed fusion algorithms that avoid double-counting and reduce communication can be designed by using both information graphs and Bayesian networks. Examples in tracking and classification illustrate the utility of this approach.