A context-aware GraphSAGE–Transformer model for LED-based indoor positioning in narrow corridors

Indoor localization in confined spaces such as hallways presents significant challenges due to limited direct signal access and increased multipath reflections. Improving localization accuracy in narrow and irregularly shaped hallways is quite difficult. This paper proposes a novel context-aware GraphSAGE–Transformer model for indoor visible light localization. A context-based feature weighting mechanism is designed to learn global received signal strength reliability patterns across multiple LEDs. Simultaneously, the model attempts to eliminate unreliable measurements caused by noise and multipath effects. GraphSAGE is then employed to model local spatial correlations among neighboring LEDs, while a Transformer encoder captures long-range and global dependencies through self-attention. Additionally, we utilize the Huber loss to enhance robustness against severe received signal strength outliers in non-line-of-sight environments. Applied in narrow L-shaped and T-shaped hallways with minimal lighting infrastructure, the proposed method helps to achieve a root mean squared error of 0.2 meters, reducing the error by 55% compared to using Transformer alone and by 44% compared to using GraphSAGE alone. Furthermore, experimental results also demonstrate that the model achieves approximately 3 times better accuracy than traditional machine learning models, such as Random Forests and k-Nearest Neighbors. These results confirm the effectiveness of the proposed model in overcoming the inherent limitations of indoor hallway environments.