Distributed Kalman filtering: Consensus, diffusion, and mixed

Abstract

A distributed Kalman filtering technique is developed for tracking state-space processes via sensor networks. Considering the optimal solution to multi-agent sequential filtering of linear Gaussian state-space processes, that is the centralized Kalman filter, this work focuses on decomposing and distributing the operation of the centralized Kalman filter among the agents of the sensor network. This decomposition is performed in a fashion that allows each agent to maintain a local Kalman filtering operation and an intermediate estimate of the state vector, providing for a robust distributed Kalman filtering technique that is scalable with the size of the network. In contrast to state-of-the-art distributed Kalman filtering approaches that focus on the use of consensus or diffusion as the basis of their information fusion, a mixed approach is proposed that exploits advantages of both methods. The performance of the proposed distributed Kalman filtering technique is verified in a simulation example, where the proposed technique is shown to outperform state-of-the-art distributed Kalman filtering algorithms.