Data-Efficient Training of Gaussian Process Regression Models for Indoor Visible Light Positioning.

Area of Science:

• Computer Science
• Electrical Engineering
• Robotics

Background:

• Visible Light Positioning (VLP) systems offer precise indoor localization capabilities.
• Gaussian Process Regression (GPR) is a powerful tool for VLP, but often requires extensive training data.
• Data efficiency in training GPR models is crucial for practical VLP system deployment.

Purpose of the Study:

• To propose a novel, data-efficient training method for GPR-based VLP systems.
• To leverage active learning (AL) to optimize the training dataset for GPR-VLP.
• To evaluate the performance improvements of the proposed method against existing techniques.

Main Methods:

• Introduced Q-AL-GPR, a training methodology combining Gaussian Process Regression (GPR) with Active Learning (AL).
• The AL component progressively selects data points with low similarity to enrich the training set.
• Experimental validation was conducted in a 3D GPR-VLP system.

Main Results:

• Q-AL-GPR demonstrated superior performance compared to random draw and line-based AL training methods.
• The method achieved approximately 27.8% reduction in training data for a 3 cm mean positioning accuracy.
• A 36.4% improvement in the 97th percentile positioning error was observed with 300 training data points.

Conclusions:

• The proposed Q-AL-GPR method significantly enhances data efficiency and positioning accuracy in VLP systems.
• Optimal selection of active learning parameters (step size and initial data) is important for balancing performance and complexity.
• Q-AL-GPR offers a practical solution for deploying accurate and efficient GPR-based VLP systems.