Kalman filtering with empirical noise models

Matti Raitoharju, Henri Nurminen, Demet Cilden-Guler, Simo Särkkä

TL;DR

This paper introduces a method to incorporate raw, empirically measured noise data directly into a Kalman filter framework using monotonic splines (PCHIP). By mapping non-Gaussian measurement errors to a standard normal distribution, the authors achieve the accuracy of specialized robust filters and the efficiency of traditional Kalman extensions, significantly outperforming standard Gaussian-based methods in real-world tracking scenarios.

What Problem Does This Paper Address?

The Kalman Filter (KF) is the "gold standard" for state estimation, but it has a famous Achilles' heel: it assumes noise is Gaussian (bell-curved). In the real world, sensors often produce "non-Gaussian" noise—think of outliers in GPS data, skewed errors in radio ranging, or heavy-tailed residuals in satellite magnetometers.

When the Gaussian assumption fails, engineers usually turn to Particle Filters, which are accurate but computationally expensive, or Robust Kalman Filters, which require you to guess a specific distribution (like Student-$t$). This paper addresses the gap for a filter that is both computationally efficient and flexible enough to use actual measured error data without needing a predefined mathematical formula for the noise.

Methods

The authors propose a two-step approach to bridge the gap between empirical data and Kalman filtering:

  1. The Noise Transformation: They model the measurement noise $\epsilon_E$ as a transformation of a standard normal variable $\epsilon_N$. To find this transformation function $f(\cdot)$, they use Piecewise Cubic Hermite Interpolating Polynomials (PCHIP) fitted to empirical error samples. PCHIP is chosen because it is "strictly increasing," ensuring a one-to-one mapping that prevents the filter from getting stuck in local optima.
  2. The Iterative Filter: They use an Iterated Posterior Linearization Filter (IPLF). Unlike a standard Extended Kalman Filter (EKF) that linearizes once around a prior, the IPLF iteratively refines its linearization around the posterior estimate.
  3. Stability Tweaks: To make this work with empirical data, they introduce two key modifications:
  4. Covariance Inflation ($\kappa$): A small diagonal term added to the augmented state covariance to prevent mathematical singularities.
  5. Damping ($\alpha$): A line-search-inspired factor that prevents the filter from jumping into "data deserts"—regions where there are too few empirical samples to provide a reliable noise model.

Datasets

The algorithm was evaluated on three distinct datasets:

  • Simulated Student-$t$: A synthetic test with heavy-tailed noise to compare against filters specifically designed for Student-$t$ distributions.
  • Satellite Magnetometer Data: Real residuals from the SWARM-A satellite, used to model the complex, heavy-tailed noise found in low-Earth orbit magnetic field measurements.
  • Ultra Wideband (UWB) Ranging: Real-world pedestrian tracking data. The authors used a "learning set" of 6 tracks to build the PCHIP noise model and a "verification set" of 3 tracks to test the positioning accuracy.

Key Insights and Ideas

  • Empirical is Practical: You don't need a named distribution (like Cauchy or Skew-t) to be accurate. If you have 300+ samples of your sensor's error, you can build a better model than a "best-guess" analytical one.
  • Monotonicity Matters: The choice of PCHIP over other splines is critical. Because PCHIP preserves the shape of the data without "wiggling," it maintains the mathematical integrity of the transformation from Gaussian to non-Gaussian space.
  • The "Data Desert" Problem: The authors realized that filters can fail if they update the state to a value where no empirical noise data exists. Their damping factor $\alpha$ acts as a safety rail, keeping the estimate within the "statistically significant" regions of the empirical data.

Key Results

  • Superior Accuracy: In the UWB pedestrian tracking test, the proposed method achieved an average error of 0.5 meters. In contrast, a standard Unscented Kalman Filter (UKF) using Gaussian assumptions had an error of 1.5 meters.
  • Competitive with Specialists: In the Student-$t$ simulation, the proposed method performed significantly better than standard KFs and was nearly as accurate as the Recursive Outlier-Robust Filter (RORF)—despite the RORF being specifically "cheating" by knowing the exact distribution type beforehand.
  • Efficiency: The algorithm remains computationally efficient, operating in the same order of magnitude as other iterative Kalman extensions, making it much faster than Particle Filters for high-dimensional states.

Limitations and Future Work

  • Sample Dependency: The model requires a representative set of error samples. The authors noted that accuracy plateaus after about 300 samples, but very rare outliers might still be missed if they aren't in the training set.
  • Linear Extrapolation: Outside the range of the empirical data, the model reverts to linear extrapolation. Future work could explore more sophisticated ways to handle these "extreme" tails.
  • Dimensionality: While more efficient than Particle Filters, the augmented state approach still grows in complexity as the number of sensors (and thus noise variables) increases.

Summary updated

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.
BibTeX
@inproceedings{raitoharju2021kalman,
  author = {Matti Raitoharju and Henri Nurminen and Demet Cilden-Guler and Simo Särkkä},
  title = {{Kalman filtering with empirical noise models}},
  booktitle = {ICL-GNSS},
  year = {2021},
  url = {https://openreview.net/forum?id=TMPiDnaiKE},
}

Loading related content...

Figure