Comparative analysis of the performance of regression machine learning models for indoor visible light positioning systems

Abstract

This paper presents an Indoor Visible Light Positioning (VLP) system designed for deployment in enclosed environments, using four ceiling-mounted Light Emitting Diodes (LEDs) to serve both illumination and positioning functions. Each LED is fixed at predefined coordinates and transmits unique signals to a floor-level receiver via Visible Light Communication (VLC) technology. Upon receiving these signals, the system performs photoelectric conversion, translating light signals into electrical signals, and generating the dataset used for training machine learning models. Several models, including LSTM, GRU, Random Forest, KNN, Decision Tree, and XGBoost, were trained and evaluated for positioning accuracy. The experimental results indicate that XGBoost achieved the best performance, with remarkably low error rates, producing a MAPE of 0.0022%, an RMSE of 0.0011, and a perfect R2 score of 1, thus being the most effective model for this application. LSTM and GRU are neural network-based models that performed very close to XGBoost. XGBoost's exceptional error correction capabilities and consistent performance across all evaluation metrics make it particularly suitable for high-precision indoor VLP systems, demonstrating superior ability in handling complex spatial relationships and the inherent variability of indoor environments. The system's effectiveness in utilizing light signals for precise node positioning marks a significant advancement in the field of indoor positioning, offering a reliable solution for real-world applications where high accuracy is essential.