Distributed Particle Filtering in Networks of Agents

This thesis is focused on distributed particle filter (DPF) algorithms performing decentralized sequential state estimation in networks of agents. The individual agents are equipped with sensing, computation, and communication capabilities. Applications of networks of agents are manifold and include environmental and agricultural monitoring, target tracking, pollution source localization, healthcare monitoring, chemical plume tracking, and surveillance. Performing distributed (collaborative) estimation of certain states of the environment from the measurements obtained by the agents is an essential task in many applications. Particle filters (PFs) are a modern approach to sequential state estimation that offers superior performance in nonlinear and non-Gaussian systems. A distributed (decentralized) implementation of PFs in networks of agents is complicated by the fact that the measurements are dispersed among the agents. Diffusing the locally available information throughout the network is thus an essential component of DPF algorithms. In this thesis, we group the existing DPF algorithms into three main classes, which we call fusion center-based, statistics dissemination-based, and measurement dissemination-based DPFs. Subclasses of the statistics dissemination-based class include leader agent-based and consensus-based DPFs. We discuss the main features and properties of the DPF algorithms belonging to the various (sub)classes. Then, we propose a novel leader agent-based DPF with a “look-ahead” proposal density adaptation that significantly reduces the estimation error. In this DPF, Gaussian mixture representations of partial posteriors are propagated along an aggregation chain of agents. We also propose several consensus-based DPFs that employ the novel “likelihood consensus” method for a distributed approximate calculation of the global (all-agents) likelihood function. In these DPFs, each agent executes a local PF that calculates a particle representation of the global posterior. To reduce the amount of communications and the latency, we propose a modification of the likelihood consensus that requires only a single consensus iteration per PF recursion. Finally, we develop a distributed, consensus-based proposal adaptation scheme that yields a performance improvement or allows a reduction of the number of particles. Simulation results for a target tracking problem demonstrate the excellent performance of the proposed DPFs, which is often very close to the performance of a centralized PF.