Robust Centimeter-Level Visible Light Positioning With Sparse RSS Fingerprints

Although fingerprint-based visible light positioning (VLP) systems can achieve centimeter-level accuracy, their performance strongly depends on dense received signal strength (RSS) fingerprints, limiting practical deployment. To address this challenge, we propose a centimeter-level VLP method on sparse RSS fingerprints, integrating three key components: RSS fingerprint regeneration, anomaly detection, and co-optimized positioning. In particular, the proposed method employs a network framework that leverages geometric distance vectors between location fingerprints and multiple light-emitting diodes (LEDs) as input features, enabling accurate and synchronous regeneration of RSS fingerprints across multiple LEDs. In addition, to mitigate potential outliers in the original RSS database, an anomaly detection framework based on leave-one-out cross-validation (LOOCV) and statistical thresholding is incorporated. Furthermore, we introduce a co-optimized positioning algorithm that integrates RSS fingerprint regeneration. Experimental results demonstrate that in a2× 2× 2-m room with four LED light sources, the proposed method achieves comparable localization accuracy using only 20% of the original dense RSS database, outperforming seven baseline methods. With this sparse RSS database, the system yields average positioning errors (PEs) of 0.28 cm (horizontal receiver orientation), 0.48 cm (tilted receiver orientation), and 0.96 cm in 3-D positioning. The proposed method exhibits robustness to sparse sampling, orientation changes, natural light, and geometric perturbations.