Generating Light-based Fingerprints for Indoor Localization

TL;DR

This paper presents a novel indoor localization method using visible light spectral signatures captured by a low-cost sensor, employing GAN-augmented training data to improve accuracy and robustness over traditional radio-frequency solutions.

Contribution

Introduces a two-stage framework combining spectral fingerprinting with GAN-based data augmentation for improved indoor localization accuracy.

Findings

Reduced mean localization error from 62.9cm to 49.3cm

20% improvement in accuracy with minimal additional data collection

GAN augmentation enhances generalization in spectral fingerprinting

Abstract

Accurate indoor localization underpins applications ranging from wayfinding and emergency response to asset tracking and smart-building services. Radio-frequency solutions (e.g. Wi-Fi, RFID, UWB) are widely adopted but remain vulnerable to multipath fading, interference, and uncontrollable coverage variation. We explore an orthogonal modality -- visible light communication (VLC) -- and demonstrate that the spectral signatures captured by a low-cost AS7341 sensor can serve as robust location fingerprints. We introduce a two-stage framework that (i) trains a multi-layer perceptron (MLP) on real spectral measurements and (ii) enlarges the training corpus with synthetic samples produced by TabGAN. The augmented dataset reduces the mean localization error from 62.9cm to 49.3cm -- a 20% improvement -- while requiring only 5% additional data-collection effort. Experimental results obtained on 42 reference points in a U-shaped laboratory confirm that GAN-based augmentation mitigates data-scarcity issues and enhances generalization.