Visible light positioning (VLP) has demonstrated promising outcomes in indoor environments. However, underground mines present significant challenges compared to ideal indoor conditions, substantially increasing positioning complexity. Traditional indoor positioning methodologies no longer adequately address underground mine positioning requirements. Consequently, developing a reliable and efficient positioning system for underground mines has become a critical challenge. In underground mining operations, the unpredictable nature of workers' head movements and variations in receiver height significantly impact visible light positioning performance. Thus, three-dimensional (3D) positioning presents a more suitable approach for personnel tracking in underground environments. During operations, head tilting introduces directional reception uncertainties, causing random fluctuations in received signal strength (RSS). This substantially reduces the accuracy of fingerprinting methods based on vertical received power and potentially leads to positioning failures. Moreover, reflections from irregular mine walls affect the optical power received by the photodetector (PD), resulting in unstable received power signals near walls and increased positioning errors. To improve positioning accuracy in complex underground environments, this study implements an inclination correction for vertical received optical power to accommodate tilted reception conditions (fingerprint correction). Additionally, we introduce a deep learning network that combines Transformer with MobileNetV3 (ViTs-MNV3) to process datasets from complex underground mine environments. Using point cloud technology, we simulate authentic mine environments to conduct comprehensive research on visible light positioning under these challenging conditions.

Currently, commonly used ranging-based positioning methods include RSS, time of arrival (TOA), time difference of arrival (TDOA), and angle of arrival (AOA). Among these approaches, fingerprinting-based positioning methods utilizing RSS are extensively researched due to their simple hardware requirements, operational simplicity, and capacity to achieve high positioning accuracy when integrated with deep learning networks for fingerprint database training. This study initially employs a conventional method to vertically collect optical power data for establishing a fingerprint database. However, receiver tilting causes the received optical power to deviate from the pre-established vertical fingerprint database, resulting in mismatches and decreased positioning accuracy. To address this challenge, we introduce a variable gain factor to convert the vertical fingerprint database into an inclined power database, thereby adapting to received power variations caused by receiver tilting and enhancing positioning accuracy. Furthermore, we propose a Transformer-enhanced MobileNetV3 network (ViTs-MNV3) for visible light positioning. This network utilizes the Transformer's multi-head attention mechanism to identify relationships between different input sequence components and learn global features, demonstrating enhanced performance. To increase simulation authenticity, we employ a depth camera to gather real-world point cloud data of mine walls and incorporate point cloud technology to construct a simulated underground mine environment that accurately reflects real-world conditions.

In real-world environments, the fingerprint database collected under vertical reception conditions is used to train the proposed network. A comparative study is conducted between the proposed ViTs-MNV3 network and four other models: MobileNetV3, Transformer neural network, backpropagation (BP) neural network, and long short-term memory (LSTM) network. The results demonstrate that the proposed network outperforms the other four networks in terms of positioning accuracy in real-world scenarios, proving its feasibility and superior robustness (Table 2 and Fig. 6). In an ideal simulated environment, the proposed network achieves a root mean square error (RMSE) of only 2.36 cm and an average positioning error of 3.2 cm in 3D positioning (Fig. 7). However, in the simulation space incorporating point cloud data of walls, the influence of wall reflections became significant, increasing the RMSE to 5.42 cm at a reception height of 1.2 m (Fig. 9). When the receiver tilted, the received power no longer matched the vertical fingerprint database, leading to substantial positioning errors (Fig. 10). To validate the practical applicability of the proposed algorithm, real-world positioning experiments are conducted. Optical power data are collected at four different heights under vertical reception conditions, and after deep network training, the final 3D RMSE reaches 10.95 cm, with an average error of 8.36 cm. The results achieved centimeter-level positioning accuracy, with 90% of errors below 25 cm and a low proportion of large errors (Fig. 12). Finally, a variable gain factor is applied to correct the vertical fingerprint database in real-world environments, transforming it into an inclined fingerprint database to match the received power under tilted reception conditions (Fig. 13). The transformed power closely resembles the actual inclined power measurements, and the RMSE is reduced to 21.43 cm (Fig. 14). However, compared to Fig. 12, this approach results in higher positioning errors due to the additional errors introduced by wall reflections during power transformation.

This research presents a three-dimensional visible light positioning system designed for complex underground mine environments through the integration of 3D point cloud technology. The investigation examines multiple influential factors, including wall reflections, receiver height variations, and receiver tilting, each affecting positioning accuracy distinctly. Receiver tilting emerges as the most critical factor, resulting in an average positioning error of 2.17 m within a 4.0 m×3.0 m×2.5 m simulated environment, significantly impacting personnel positioning capabilities in underground mines. To mitigate this challenge, the study implements a variable gain factor to modify the vertical fingerprint database, converting it into an inclined fingerprint database that corresponds to received power under tilted reception conditions. This methodology substantially enhances positioning accuracy and facilitates the expansion of fingerprint databases for various tilt angles without requiring extensive manual data collection, thus optimizing resource efficiency. The positioning algorithm incorporates the strengths of the Transformer neural network and MobileNetV3 network, forming an innovative hybrid model, ViTs-MNV3, which exhibits exceptional performance in positioning applications. The algorithm achieves a RMSE of 2.36 cm in ideal simulated conditions and 5.42 cm in simulated environments with wall reflections. Experimental validation in real-world conditions demonstrates a RMSE of 10.95 cm under vertical reception conditions. Following power correction for tilted reception, although the RMSE increased to 21.43 cm, the corrected power values demonstrate substantial improvement in positioning accuracy compared to uncorrected tilted reception scenarios.