Abstract
The aim of this article is to present Levenberg-Marquardt and line-search extensions of the classical iterated extended Kalman smoother (IEKS) which has previously been shown to be equivalent to the Gauss-Newton method. The algorithms are derived by rewriting the algorithm's steps in forms that can be efficiently implemented using modified EKS iterations. The resulting algorithms are experimentally shown to have superior convergence properties over the classical IEKS.
Introduction
In the fields of signal processing and navigation, the Extended Kalman Smoother (EKS) is a standard tool for estimating system states from noisy measurements. While the Iterated Extended Kalman Smoother (IEKS) improves upon the basic EKS, it is mathematically equivalent to the Gauss-Newton optimization method. This equivalence makes the IEKS susceptible to instability and divergence in highly non-linear systems or when initialized far from the true state. This paper introduces two robust extensions—Levenberg-Marquardt (LM-IEKS) and Line-Search (LS-IEKS)—designed to stabilize convergence while maintaining the recursive efficiency of the Kalman framework.
Robust Extensions to IEKS
The authors propose two primary strategies to "robustify" the IEKS:
- Levenberg-Marquardt IEKS (LM-IEKS): By incorporating a damping term into the iteration, the algorithm prevents excessively large steps that lead to divergence. A significant contribution of this work is demonstrating that this damping can be implemented as a "pseudo-measurement" update, allowing for seamless integration into existing EKS codebases.
- Line-Search IEKS (LS-IEKS): This method introduces a step-size parameter $\alpha$ to ensure that each iteration results in a decrease of the cost function. The paper explores both exact and inexact line-search techniques.
Crucially, both methods retain $O(N)$ computational complexity, ensuring they remain viable for real-time processing of long data sequences.
Experimental Evaluation
The proposed methods were evaluated using a simulated Coordinated Turn Bearings-Only Tracking scenario. In this benchmark, two stationary sensors track a target moving in a sharp turn. Because the sensors only measure the angle (bearing) to the target, the measurement model is highly non-linear, providing a rigorous test for the stability of smoothing algorithms.
Results and Discussion
The experimental results highlight the superior performance of the robust smoothers. While a single-pass EKS resulted in a Root Mean Square Error (RMSE) of 2.6, both the LM-IEKS and LS-IEKS achieved an RMSE of 0.6.
Beyond accuracy, the primary advantage is stability. As shown in the cost function analysis, the standard IEKS (Gauss-Newton) often exhibits extreme spikes in the cost function during early iterations, which can lead to divergence in more complex scenarios. The proposed LM and Line-Search variants, however, ensure stable, monotonic convergence.
Conclusion
By integrating classical optimization safeguards into the recursive Kalman smoothing framework, this research provides a more reliable approach to non-linear state estimation. The mathematical insight that regularization can be treated as a measurement update simplifies implementation and broadens the applicability of these robust methods to a wide range of tracking and navigation problems.
Summary updated