Efficient and robust 3D indoor visible light positioning via uniform and power-of-two quantization on multi-head...

Area of Science:

• Computer Vision
• Wireless Communication
• Machine Learning

Background:

• Visible light communication (VLC) systems offer potential for indoor positioning but face challenges with accuracy, computational cost, and reliance on extra sensors.
• Existing methods often require large databases and powerful hardware, limiting their use in resource-constrained environments.

Purpose of the Study:

• To develop an efficient and accurate indoor positioning algorithm for VLC systems.
• To achieve an optimal balance between positioning accuracy, inference time, and model size without additional sensors.

Main Methods:

• Proposed a uniform and power-of-two quantization on a multi-head ResNet50 (UPU-MH-ResNet50) algorithm.
• Implemented model quantization to reduce computational costs and enhance energy efficiency.
• Validated the algorithm on a self-built experimental testbed measuring 2.6m x 2.6m x 2.2m.

Main Results:

• The UPU-MH-ResNet50 algorithm accurately estimates 3D coordinates (X, Y, and rotation angle) without extra sensors.
• 90% of 3D positioning errors were controlled within 2 cm for receiver rotation angles up to ±30°.
• The 4-bit UPU-MH-ResNet50 achieved a 7.7x reduction in model size and a 2.7x speedup in inference.

Conclusions:

• The proposed UPU-MH-ResNet50 algorithm offers a robust and efficient solution for indoor positioning in VLC systems.
• Model quantization effectively reduces computational load, leading to significant improvements in speed and energy efficiency.
• The algorithm demonstrates high accuracy and robustness against random angle changes, making it suitable for practical applications.