Abstract
Most Kalman filter extensions assume Gaussian noise and when the noise is non-Gaussian, usually other types of filters are used. These filters, such as particle filter variants, are computationally more demanding than Kalman type filters. In this paper, we present an algorithm for building models and using them with a Kalman type filter when there is empirically measured data of the measurement errors. The paper evaluates the proposed algorithm in three examples. The first example uses simulated Student-t distributed measurement errors and the proposed algorithm is compared with algorithms designed specifically for Student-t distribution. Last two examples use real measured errors, one with real data from an Ultra Wideband (UWB) ranging system, and the other using low-Earth orbiting satellite magnetometer measurements. The results show that the proposed algorithm is more accurate than algorithms that use Gaussian assumptions and has similar accuracy to algorithms that are specifically designed for a certain probability distribution.
Introduction
The Kalman Filter (KF) is a cornerstone of state estimation, yet its reliance on the Gaussian noise assumption often limits its performance in real-world applications. Sensors frequently exhibit non-Gaussian characteristics, such as heavy tails in satellite magnetometers or outliers in radio-based ranging systems. While Particle Filters offer a solution for non-Gaussian noise, their high computational cost makes them impractical for many real-time systems. Robust Kalman Filters exist but typically require a predefined analytical distribution (e.g., Student-t). This work proposes a method to utilize raw empirical error data directly within a Kalman framework, bridging the gap between computational efficiency and distributional flexibility.
Methodology
The core of the proposed approach is a two-step process to integrate empirical noise into the filter.
Noise Transformation via PCHIP
Measurement noise $\epsilon_E$ is modeled as a transformation of a standard normal variable $\epsilon_N$ through a function $f(\cdot)$. To estimate this function from empirical samples, the authors employ Piecewise Cubic Hermite Interpolating Polynomials (PCHIP). PCHIP is selected for its strict monotonicity, ensuring a unique mapping between the Gaussian and empirical domains without the oscillatory "wiggles" common in other splines.
Iterative Filtering and Stability
The filter utilizes an Iterated Posterior Linearization Filter (IPLF), which refines the linearization of the non-linear noise model around the posterior estimate rather than the prior. To ensure numerical stability when working with empirical data, two modifications are introduced:
- Covariance Inflation ($\kappa$): A small diagonal term is added to the augmented state covariance to prevent singularities.
- Damping ($\alpha$): A line-search factor prevents the filter from making excessive updates into "data deserts"—regions where empirical samples are sparse.
Experimental Evaluation
The algorithm was tested across three scenarios:
- Simulated Student-t: Comparing against specialized robust filters.
- Satellite Magnetometer Data: Using residuals from the SWARM-A satellite to model complex, heavy-tailed noise.
- Ultra Wideband (UWB) Ranging: Real-world pedestrian tracking where a noise model was learned from a training set and verified on separate tracks.
Results and Discussion
The proposed method demonstrated significant accuracy gains. In UWB tracking, it achieved a mean error of 0.5m, compared to 1.5m for a standard Unscented Kalman Filter (UKF). In Student-t simulations, it matched the performance of filters specifically designed for that distribution.
As shown in the figure above, the filter's performance is highly dependent on the number of empirical samples used to build the model. Accuracy improves rapidly and plateaus after approximately 300 to 1000 samples, suggesting that a relatively small amount of empirical data is sufficient for robust modeling.
Conclusion
By mapping empirical noise to a Gaussian space using monotonic splines, this research provides a practical way to enhance Kalman filters without the overhead of particle filters or the rigidity of analytical noise models. Future work will focus on improving tail extrapolation and managing the computational complexity of high-dimensional augmented states.
Summary updated