Selection of relevant particles in nonparametric distributed message passing for cooperative localization

Distributed nonparametric belief propagation can be employed for cooperative localization. Its natural asset of providing a measure of uncertainty of position estimates makes distributed nonparametric belief propagation a powerful localization technique. However, distributed nonparametric belief propagation requires the exchange of arbitrarily complex messages over communication-constrained links. To cope with these communication constraints, parametric approximations (e.g. Gaussian mixtures) of these messages have evolved to become the de facto standard. These message approximations suffer from a limited representational power for highly nonstandard distributions, which results in limited localization accuracy. We propose two novel particle-based message approximations that select important particles according to mutual information shared with a relevant random variable. The first approach is based on the Information Bottleneck algorithm, while the second scheme is based on partial mutual information. We compare both particle-based methods through simulations, demonstrating an improvement over the parametrized approach in terms of localization error with only moderate increase in communications.

Options

Selection of relevant particles in nonparametric distributed message passing for cooperative localization