Revisiting Gaussian Process Regression Modeling for Localization in Wireless Sensor Networks.

Area of Science:

• Wireless Sensor Networks
• Localization and Positioning
• Machine Learning for Signal Processing

Background:

• Signal strength-based positioning is crucial for ubiquitous localization, especially in GNSS-denied environments.
• Wireless Local Area Network (WLAN) fingerprinting requires efficient radio map creation for large-scale accuracy.
• Gaussian process regression (GPR) shows promise but lacks thorough model fit assessment in prior studies.

Purpose of the Study:

• To investigate and select optimal Gaussian process regression models for WLAN fingerprinting.
• To evaluate and compare GPR model performance in indoor, outdoor, and combined environments.
• To assess the direct fit of GPR models using quantitative measures and residual analysis.

Main Methods:

• Trained and evaluated multiple Gaussian process regression models for WLAN fingerprinting.
• Utilized indoor, outdoor, and combined area datasets for comprehensive model testing.
• Assessed model fit using quantitative measures and analyzed model residuals to understand performance.

Main Results:

• The standard GPR model (zero mean, squared exponential covariance) is not optimal for WLAN fingerprinting.
• Comparative experiments demonstrated significant differences in positioning performance based on model selection.
• Specific GPR model configurations were identified as superior for indoor, outdoor, and combined environments.

Conclusions:

• Thorough model selection and assessment are critical for accurate Gaussian process regression-based WLAN fingerprinting.
• The study presents a best-candidate GPR model, outperforming standard approaches.
• Findings contribute to more efficient and accurate radio map creation for large-scale localization systems.