Indoor visible light positioning with machine learning

Visible Light Positioning (VLP) systems are increasingly being featured in the literature as an alternative method to radio frequency (RF)-based positioning systems. Thanks to their low-cost infrastructure requirements and high positioning accuracy, VLP systems offer significant advantages compared to RF-based systems. In addition, in recent years, there has been a growing trend in the use of machine learning (ML) algorithms to improve system performance, and various studies have been conducted in this direction. This study analyzes the effect of ML algorithms on positioning error in VLP systems. Error performances were obtained using the received signal strength (RSS) parameter in an indoor scenario that considers both line-of-sight (LOS) and the first reflection (non-line-of-sight, NLOS). Six different scenarios with 4, 5, 8, 9, 12, and 16 light-emitting diodes (LEDs) were considered. A dataset was created using RSS information obtained from LOS and NLOS signals in these scenarios. Various machine learning algorithms were run on this dataset and the results were analyzed. Among the algorithms used, XGBoost and Random Forest (RF) demonstrated the highest success rates. The k-nearest neighbors (kNN) algorithm showed lower accuracy compared to XGBoost and RF. The trilateration algorithm performed worse than kNN, XGBoost, and RF. Linear Regression (LR) was the algorithm with the lowest performance. As the number of LEDs increased, the performances of the algorithms converged. However, the LR model consistently showed very poor performance across all LED counts. While the RF algorithm performed better at lower LED counts, the performance of the XGBoost algorithm surpassed that of RF as the number of LEDs increased. Additionally, since reflection effects are more pronounced in indoor areas, especially near walls, the dataset was divided into two regions: center and edge. Based on this separation, ML algorithms were trained separately for each region. Using this method, the impact of NLOS signals was reduced, leading to lower average positioning errors compared to both the traditional trilateration algorithm and single-region ML-based methods.