Distributed particle filter using Gaussian approximated likelihood function

Abstract

In this paper, we propose a distributed particle filtering algorithm for sensor networks in which multiple sensors collaborate to monitor and track an object in a nonlinear/non-Gaussian dynamic system. According to the algorithm, the sensors collaboratively compute the global likelihood function in order to make local estimates that takes into account measurements from all the sensors. To compute the global likelihood, each sensor first approximates its local likelihood function using Gaussian function, and exchange its approximated local likelihood with the other sensors. Such approximation saves communication overhead because it requires the sensors to exchange only the mean and the covariance of the approximated Gaussian local likelihood functions. The exchange of the parameters of the likelihood functions between sensors is accomplished using an average consensus filter or by implementing forward-backward propagation strategy.