Robust Localization Method Based on Non-Parametric Probability Density Estimation

Chee-Hyun Park, Joon-Hyuk ChangIEEE Access 2023 · OpenReview

Abstract

This paper presents robust localization techniques that calculate location using distance observations. In enclosed and heavily populated urban environments, the positive measurement bias introduced by a non-line-of-sight (NLOS) signal can have a considerable adverse impact on estimation performance. To mitigate the detrimental effects of the multipath effect caused by the NLOS signal, robust localization techniques are considered. In particular, the k-nearest neighbor (KNN)-based and orthogonal series (OSERIES)-based localization approaches are proposed. Unlike conventional probability density estimation (PDF) methods, the proposed methods use the first-peak information of the estimated PDF to obtain actual distance information, rather than just the PDF shape. Specifically, the proposed methods use the mean calculated from observations selected by statistical testing, as the mean estimate generally outperforms the mode estimate. Additionally, the Rao test in the context of the two-mode Gaussian mixture model (GMM) is demonstrated to be the uniformly most powerful (UMP) test. The conditional variance of the range measurement is derived, and the proposed techniques are shown to outperform competing algorithms in terms of localization accuracy.

Introduction: The Challenge of NLOS Localization

In ideal conditions, localization—the process of determining a target's position—relies on direct Line-of-Sight (LOS) distance measurements between a sensor and a target. However, in "enclosed and heavily populated urban environments," signals rarely travel in a straight line. Instead, they often bounce off walls, buildings, or other obstacles. This phenomenon, known as the Non-Line-of-Sight (NLOS) problem, introduces a significant positive measurement bias. Because the reflected signal travels a longer path than the direct one, the distance appears longer than it actually is.

Traditional localization algorithms, such as Maximum Likelihood Estimation (MLE) or standard Kalman filters, typically assume that measurement noise is symmetric and follows a Gaussian distribution. In reality, NLOS noise is unpredictable, time-variant, and heavily skewed. The primary challenge addressed in this research is to accurately "see through" this noise to identify the true LOS distance, even when the majority of signals are corrupted.

Proposed Methodology: A Non-Parametric Approach

The authors propose a multi-stage approach that avoids rigid assumptions about noise distribution shapes. The core of the method involves:

  1. Non-Parametric PDF Estimation: Rather than assuming a Gaussian model, the system uses k-nearest neighbor (KNN) and Orthogonal Series (OSERIES) to estimate the Probability Density Function (PDF) of the distance observations. This flexibility allows the model to adapt to various complex noise environments.
  2. The "First-Peak" Strategy: In a multipath environment, the shortest distance measured usually corresponds to the true LOS path. The proposed algorithms search for the first peak in the estimated PDF, which represents the earliest arriving signal component, rather than simply selecting the highest peak (the mode).
  3. Statistical Outlier Rejection (Rao Test): To distinguish between "inliers" (accurate LOS data) and "outliers" (biased NLOS data), the researchers employ the Rao test. A significant theoretical contribution of this paper is the mathematical proof that the Rao test is the Uniformly Most Powerful (UMP) test for identifying NLOS outliers within a two-mode Gaussian Mixture Model.

Refinement and Localization

Once the PDF is estimated and outliers are identified, the method refines the distance estimate using the Nadaraya-Watson estimator, a type of conditional mean. The authors argue that once the data has been filtered for inliers, the mean of the selected observations provides a more stable and accurate estimate than the mode, which can be sensitive to small fluctuations in the PDF shape.

Finally, these refined distance estimates are converted into precise (x, y) coordinates using a two-step Weighted Least Squares (WLS) method. This ensures that more reliable distance measurements (those with lower estimated variance) have a greater influence on the final position calculation.

Experimental Evaluation and Results

The proposed methods were evaluated using both simulated environments and real-world benchmark data.

Simulation Results

In synthetic tests where NLOS noise followed complex distributions (such as Skew-t or Gaussian-exponential mixtures), the proposed methods demonstrated remarkable resilience. Even in scenarios with high contamination—where 60% of sensors were affected by NLOS bias—the RMSE (Root Mean Square Error) remained close to the Cramer-Rao Lower Bound (CRLB), the theoretical limit of accuracy for any unbiased estimator.

Real-World Performance

The algorithms were further validated using an ultra-wideband (UWB) dataset involving eight sensors in a three-dimensional physical space. The KNN-WLS method achieved an RMSE of 0.2273m, significantly outperforming traditional alternatives like the Adaptive MCC-EKF (1.1294m) and the standard SSM (1.3341m). This represents a nearly five-fold improvement in accuracy over some existing state-of-the-art methods.

Conclusion and Future Directions

This research provides a robust framework for localization in challenging urban and indoor environments. By combining non-parametric PDF estimation with rigorous statistical testing, the proposed method achieves high accuracy without requiring prior knowledge of the noise environment.

However, the authors note certain limitations. Estimating a PDF using a large number of grid points is computationally intensive, making the current implementation better suited for static or slow-moving targets rather than high-speed dynamic tracking. Future work will focus on reducing this computational overhead and exploring "single observation" robust localization, aiming to achieve similar accuracy levels when only one data point is available per sensor.

Summary updated

Abstract

This paper presents robust localization techniques that calculate location using distance observations. In enclosed and heavily populated urban environments, the positive measurement bias introduced by a non-line-of-sight signal can have a considerable adverse impact on estimation performance. Therefore, to mitigate the detrimental effects of the multipath effect caused by the non-line-of-sight signal, robust localization techniques are considered. In particular, the ${k}$ -nearest neighbor (KNN)-based and orthogonal series (OSERIES)-based localization approaches are proposed. The difference from conventional probability density estimation (PDF) estimation methods is that the proposed methods use the first-peak information of the estimated PDF to obtain the actual distance information, not just the PDF shape estimation. More specifically, the proposed methods use the mean calculated from observations selected by statistical testing because the mean estimate generally outperforms the mode estimate. In addition, the Rao test in the context of the two-mode Gaussian mixture model (GMM) is demonstrated to be uniformly most powerful (UMP) test. Furthermore, the conditional variance of the range measurement is derived. Also, the proposed techniques outperforms that of competing algorithms in terms of localization accuracy.
BibTeX
@inproceedings{park2023robust,
  author = {Chee-Hyun Park and Joon-Hyuk Chang},
  title = {{Robust Localization Method Based on Non-Parametric Probability Density Estimation}},
  booktitle = {IEEE Access},
  year = {2023},
  url = {https://openreview.net/forum?id=XOU8y0rWoF},
}
Figure