Particle filter divergence monitoring with application to terrain navigation
Abstract
Particle filters are an efficient Monte-Carlo method for Bayesian estimation in non-linear models. However, under certain circumstances, they are subject to divergence. Increasing the number of particles is not always possible so it is essential for many applications to assess the reliability of the solution provided by the filter. In terrain navigation, trusting an erroneous position estimate can be problematic for obvious reasons. We introduce a framework for detecting filter divergence in the case of scalar measurements. The detector is based on a sequential change detection algorithm and we illustrate its performance on several terrain navigation scenarios.