Robust 3D Visible Light Positioning Against Receiver Tilt via Attention-Based Deep Neural Networks

Visible Light Positioning (VLP) has emerged as a promising solution for accurate indoor positioning by leveraging the ubiquitous LED infrastructure. While camera-based VLP systems offer the advantage of high-resolution imaging and cost efficiency, their performance is significantly degraded when the receiver is tilted in random directions—a common occurrence in mobile or robotic applications. Such orientation-induced distortions violate the assumptions of conventional geometric models, leading to substantial positioning errors. To solve this limitation, we propose a novel VLP framework based on ResNet50 enhanced with a multi-head attention mechanism, designed to improve robustness against receiver tilt. The model directly learns the mapping between distorted image features and 3D coordinates, eliminating the need for additional sensors or calibration procedures. Evaluations conducted in a real-world indoor environment (2.6 m × 2.6 m × 2.2 m) with tilt angles up to ±30° demonstrate that our method achieves a positioning accuracy within 2.5 cm and superior performance compared to existing algorithms.