Visible light indoor positioning provides novel theoretical and practical support for high-speed, environmentally-friendly, safe, and economical indoor localization. Traditional non-imaging-based positioning methods face operational challenges, and current algorithms using image sensors primarily capture local appearance information, often overlooking geometric structure, thus affecting positioning accuracy. In this study, we propose a model integrating attention mechanisms and graph neural networks (GNNs), enhancing the accuracy and robustness of indoor visible light positioning.

In this study, a novel approach combines attention mechanism with GNNs. Visible light images are represented as graphs, where GNNs aggregate both intra- and inter-graph information, embedding the spatial position of feature points into descriptors, which enriches them with geometric data. The attention module further enhances descriptor quality, improves matching accuracy and realizes precise indoor positioning.

To validate the model, a 4 m×4 m×3 m visible light experimental platform is constructed. Four 10 W LED light sources are positioned at the top of the model, and the visible light area is divided into a grid of equidistant 5 cm×5 cm cells. Visible light images are captured at each grid vertex, creating a fingerprint database with 3510 images. For testing, 80% of the database images are used for feature extraction and matching, while the remaining 20% are reserved for model testing. Simulation and practical experiments are conducted with the platform at heights of 0, 0.75, and 1.50 m. The results show centimeter-level accuracy, with average errors of 5.93, 7.21, and 9.15 cm, at each height. The robustness tests are also conducted, including rotation and tilt transformations of the mobile terminal. Compared to the SuperPoint algorithm, which extracts image features for deep learning, the experimental results show notable improvements in matching rates. When the visible light image is rotated by 5°, the matching rate increases by 9%; at 30°, it increases by 17% with a maximum improvement of 20% observed in the rotation experiments. For tilt angles, a 5° tilt results in a 12% increase in matching rate, while a 30° tilt yields a 13% increase. These results indicate that the algorithm proposed in this study surpasses the SuperPoint algorithm in matching accuracy, demonstrating its superior performance.

Indoor visible light environments are often complex, with frequent light and background interferences that challenge traditional algorithms. This model, combining attention mechanisms and GNNs, optimizes indoor visible light positioning by enhancing robustness and stability. Using deformable convolutional networks (DCNs) during sampling improves key information in visible light images. GNNs effectively aggregate both intra- and inter-image information, while the attention mechanism dynamically adjusts feature weights to emphasize features with greater discrimination and reliability. This reduces the influence of illumination and occlusion, enhancing matching accuracy. In this study, a 4 m×4 m×3 m visible light indoor positioning model is constructed for simulation testing. The experimental results show an average positioning error of 7.43 cm, highlighting this approach as a viable new algorithm for indoor visible light positioning.