Introduction
Accurate localization in urban and indoor environments is a significant challenge for modern positioning systems. In these settings, signals rarely travel in a direct Line-of-Sight (LOS) path. Instead, they bounce off buildings, walls, and other obstacles—a phenomenon known as Non-Line-of-Sight (NLOS) propagation. For standard localization algorithms like simple Least Squares, these NLOS signals appear as massive measurement errors or outliers. Even a single outlier can significantly degrade position estimates, pulling them hundreds of meters away from the true location.
While robust statistical methods exist to mitigate these effects, many current approaches rely on manual statistical testing and fixed thresholds to identify outliers. These thresholds are difficult to tune and often fail in dynamic, real-world environments. This paper, published in IEEE Access, introduces three novel algorithms designed to provide high-precision localization in outlier-heavy environments without the need for manual parameter tuning.
Proposed Methodology
The researchers developed three distinct approaches to enhance localization robustness:
1. Modified MM WLS-M
This method refines the Multi-stage Maximum Likelihood-type (MM) estimator. The authors identified a critical limitation in existing MM algorithms: the assumption that transformed range measurements and distance estimates are independent when calculating weights. By deriving a novel weight that accounts for the statistical correlation between these variables, the Modified MM WLS-M algorithm achieves significantly higher precision than its predecessors.
2. MAP-EM WLS
Utilizing a Bayesian Expectation-Maximization (EM) framework, this algorithm treats the "inlier vs. outlier" status of a signal as a hidden (latent) variable. It employs Maximum A Posteriori (MAP) estimation to iteratively refine the position estimate. This approach is particularly valuable because it eliminates the need for manual outlier thresholds, as the model learns the noise characteristics directly from the data.
3. VB-EM WLS
The Variational Bayes (VB) version of the EM algorithm provides an even more robust alternative. Unlike MAP, which finds a single point estimate, VB approximates the entire probability distribution of the parameters. This makes the algorithm less prone to overfitting, especially in environments with low signal-to-noise ratios.
Experimental Evaluation
The proposed methods were rigorously tested using simulated datasets to control the "contamination ratio" (the percentage of NLOS outliers). The simulation environment consisted of a 20m x 20m area with seven fixed sensors and ten randomly placed emitters. A two-mode Gaussian mixture model was used to simulate the mix of clean LOS and contaminated NLOS signals.
The performance of these algorithms was compared against industry standards, including the standard MM WLS, Robust Extended Kalman Filter (EKF), ESPUKF, and Semi-Definite Programming (SDP) methods.
Key Insights and Results
- The Importance of Correlation: The derivation of correlation-aware weights proved to be the most significant factor in improving accuracy. The Modified MM WLS-M algorithm consistently provided the highest overall accuracy (lowest Root Mean Square Error) across most tested outlier conditions.
- Threshold-Free Robustness: The EM-based methods (MAP and VB) demonstrated that "self-tuning" algorithms are viable for localization. By estimating outlier parameters directly from the data, these methods are far more practical for dynamic applications, such as mobile robotics or indoor navigation.
- Performance Limits: A critical finding was the "three-sensor rule": EM-based methods require at least three clean LOS sensors to maintain high performance. Below this threshold, the algorithms struggle to distinguish true signals from NLOS noise.
- Mathematical Consistency: The authors provided mathematical proof for the "asymptotic unbiasedness" of their algorithms, ensuring that the estimates converge toward the true location as more data is collected.
Conclusion and Future Work
While the proposed iterative algorithms offer superior accuracy and robustness, they come with a higher computational cost compared to closed-form algebraic solutions. This may limit their immediate application on low-power IoT devices. Additionally, like many iterative solvers, they can be sensitive to initial position guesses in extremely noisy environments.
Future research will focus on developing "algebraic robust localization algorithms" that maintain this high level of accuracy while reducing computational complexity, making them suitable for real-time deployment on resource-constrained hardware.
Summary updated