Enhancing the Accuracy of Visible Light Positioning Systems Using a Kolmogorov-Arnold Network and Feature Selection

Visible light positioning (VLP) technology has emerged as a cost-effective solution for indoor localization, offering reasonable accuracy. However, its performance is limited by real-world optical constraints, noise interference, and challenges in system modeling. This study aims to enhance VLP localization accuracy by processing high-dimensional input features using a Kolmogorov–Arnold Network (KAN) and comparing its performance with that of other modeling approaches. A VLP system was developed using multiple LEDs and receivers, with prime numbers assigned as LED transmission frequencies to enable distinct positional fingerprinting across the spectrum. To achieve accurate localization, four machine learning models were employed: KAN, multilayer perceptron (MLP), random forest regression (RFR), and Transformer. To enhance model accuracy, feature selection based on coverage criteria was employed. This approach helped identify essential features while minimizing the impact of noisy inputs. Experimental results show that the KAN model outperformed the others. Specifically, KAN achieved an overall localization error of 10.04 mm within a 300 mm × 300 mm × 200 mm workspace, and an XY-plane error of 4.48 mm, along with a more concentrated error distribution.