A new approach for Kalman filtering on mobile robots in the presence of uncertainties

Abstract

In many practical Kalman filter applications, the quantity of most significance for the estimation error is the process noise matrix. When filters are stabilized or performance is sought to be improved, tuning of this matrix is the most common method. This tuning process cannot be done before the filter is implemented, as it is primarily made necessary by modelling errors. In this paper, two different methods for modelling the process noise are described and evaluated; a traditional one based on Gaussian noise models and a new one based on propagating modelling uncertainties. We discuss which method to use and how to tune the filter to achieve the lowest estimation error.