WLS Localization Using Skipped Filter, Hampel Filter, Bootstrapping and Gaussian Mixture EM in LOS/NLOS Conditions

Chee-Hyun Park, Joon-Hyuk ChangIEEE Access 2019 · DOI

Abstract

Accurate source localization in wireless sensor networks is often compromised by Non-Line-of-Sight (NLOS) propagation, which introduces significant outliers in range measurements. This paper presents four robust localization algorithms that integrate statistical filtering techniques—Hampel filter, Skipped filter, Robust Bootstrapping, and Gaussian Mixture Expectation-Maximization (EM)—into a Weighted Least Squares (WLS) framework. Unlike conventional methods that require prior knowledge of noise variance, the proposed approaches dynamically estimate reliability weights and identify outliers. Numerical simulations demonstrate that these methods achieve superior accuracy and stability in mixed LOS/NLOS environments, even under high contamination levels.

Introduction

Range-based localization is a fundamental component of wireless sensing applications, from indoor navigation to drone tracking. However, the prevalence of NLOS propagation in dense urban or indoor environments remains a major challenge. NLOS signals, caused by reflections and obstructions, result in measurement biases that can lead to substantial localization errors. While robust statistical methods exist to mitigate outliers, they often depend on pre-configured noise profiles (prior variance), which are rarely available in dynamic real-world scenarios. This research addresses the need for "blind" robust localization that can adapt to unknown noise characteristics on the fly.

Proposed Methodology

The core of the proposed solution is the integration of robust weighting mechanisms into a Weighted Least Squares (WLS) framework. The WLS approach assigns higher weights to reliable sensors and lower weights to those affected by NLOS errors. Four specific techniques were evaluated:

  1. Hampel Filter-based WLS: Utilizes Median Absolute Deviation (MAD) to detect outliers, replacing extreme values with the median to dampen their impact.
  2. Skipped Filter-based WLS: A more decisive approach that entirely discards measurements identified as outliers based on robust statistical thresholds.
  3. Robust Bootstrap-based WLS: Employs a resampling strategy to identify "clean" data subsets near the median, providing a reliable variance estimate for weighting.
  4. Gaussian Mixture EM-based WLS: An unsupervised learning approach that clusters measurements into LOS (inliers) and NLOS (outliers) groups using the Expectation-Maximization algorithm.

The authors derived closed-form algebraic expressions for the variance of the filtered data, enabling the WLS algorithm to automatically assign optimal weights without manual tuning.

Experimental Setup

The performance of the algorithms was validated through extensive numerical simulations. The test environment consisted of a 20x20 meter region equipped with seven receivers.

Configuration of the 7 receivers in the 20x20m simulation environment.

The simulation involved 10 different source locations with 200 Monte-Carlo trials each. Noise was modeled using a two-mode Gaussian mixture to represent both standard LOS noise and biased NLOS errors. The study specifically tested the algorithms' resilience to varying levels of NLOS contamination and bias intensity.

Results and Discussion

The proposed methods demonstrated significant improvements over conventional localization techniques:

  • High Precision: The algorithms achieved a Root Mean Square Error (RMSE) of approximately 2.8 cm, outperforming standard bisection and WLS methods (3.0 cm).
  • Robustness to Contamination: A key finding was the stability of these filters even when more than 50% of the sensors provided NLOS data. The Hampel and Skipped filters were particularly effective at "blocking" large biases.
  • Computational Efficiency: While the EM and Bootstrap methods are computationally intensive, the Hampel and Skipped filters provide closed-form solutions. This avoids the risk of local minima and makes them suitable for real-time implementation on low-power hardware.

Conclusion

This research provides a robust framework for localization in challenging LOS/NLOS environments. By combining statistical filtering with WLS, the proposed algorithms eliminate the need for prior noise knowledge while maintaining high accuracy. Future work will focus on validating these methods in physical multi-story environments and extending the framework to handle mobile receivers with rapidly changing LOS/NLOS status.

Summary updated

BibTeX
@article{park2019wls,
  author = {Chee-Hyun Park and Joon-Hyuk Chang},
  title = {{WLS Localization Using Skipped Filter, Hampel Filter, Bootstrapping and Gaussian Mixture EM in LOS/NLOS Conditions}},
  journal = {IEEE Access},
  year = {2019},
  doi = {10.1109/access.2019.2905367},
}
Figure