Two-Layer Nonlinear FIR Filter and Unscented Kalman Filter Fusion with Application to Mobile Robot Localization

Abstract

In this paper, we propose a new state estimator called the two-layer nonlinear finite impulse response (TLNF) filter and adopt this new filter and unscented Kalman filter (UKF) as subfilters to create the fusion TLNF/UK filter. The TLNF filter is constructed with measurements that are redefined by weighting the estimated states acquired through minimizing the cost function based on the Frobenius norm. The efficient iterative form of the TLNF filter is also developed in this paper. Using the fact that the UKF and the TLNF filter each takes a different type of memory structure, the fusion TLNF/UK filter is designed as a robust nonlinear state estimator taking both advantages of each filter. To obtain the best fusion estimates, probabilistic weights are computed based on Bayes' rule and the likelihood of each filter. Both simulation and experimental results for mobile robot indoor localization have shown that the fusion TLNF/UK filter achieves a higher level of accuracy and robustness under practical situations.