Visible Light Positioning (VLP) technology has gained increasing attention due to its potential for providing low-cost, high-precision indoor location services. However, traditional VLP systems rely on Line-of-Sight (LOS) paths and cannot function properly when obstructed by obstacles.

To address this issue, we propose a novel Non-Line-of-Sight (NLOS) VLP system based on deep learning. This system utilizes reflected light for VLP, overcoming the challenge of LOS obstruction and enhancing the robustness of the VLP system. Considering the low signal-to-noise ratio of the reflected light, the accuracy and adaptability of conventional image detection methods for extracting Light Emitting Diode (LED) spots are limited, resulting in reduced positioning accuracy for NLOS VLP. Therefore, the proposed system employs the deep learning model U-shaped Network (U-Net) to detect LED spots, which demonstrates high accuracy and adaptability after being trained on datasets collected from various environments, thereby improving the system performance. In the simulation, the system estimates the Three-Dimensional (3D) position of the receiver using the Perspective-Three-Point (P3P) algorithm.

This paper constructed a 1.84 m×1.84 m ×1.96 m 3D space simulating an indoor environment for indoor positioning experiments. The experimental results show that under NLOS paths, the system's 3D mean error and Root Mean Square Error (RMSE) are 16.09 and 17.18 cm, respectively. The Two-Dimensional (2D) positioning error has a 90% confidence level at less than 21 cm, and the 3D positioning error has a 90% confidence level at less than 24 cm.

The proposed system has high positioning accuracy and robustness, which can meet the positioning requirements of most indoor applications.