Semantic-aware dark-light visible light communication for indoor localization

Visible light communication (VLC) can support indoor positioning by exploiting existing lighting infrastructure, but conventional VLC positioning often remains coupled to active illumination. Dark-light VLC (DL-VLC) decouples localization from illumination by using short optical pulses that are detectable by photodetectors while remaining visually imperceptible. This study evaluates a semantic-aware DL-VLC localization framework in which received signal strength (RSS), angle of arrival (AOA), incidence angle, and signal-to-noise ratio (SNR) features are mapped to a compact task-oriented representation before coordinate regression. A simulated \(5\times 5\times 3~\textrm{m}^{3}\) room with four ceiling-mounted transmitters and a photodetector receiver at \(0.85~\textrm{m}\) height was used to generate line-of-sight Lambertian channel features. The model was compared with regressors based on decision trees, random forests, support vector regression, AdaBoost, k-nearest neighbors, stochastic gradient descent, and a direct multilayer perceptron under different training densities and channel degradation settings. With 5000 training points, the semantic-aware model achieved a mean localization error of \(1.740~\textrm{cm}\) and a root-mean-square error (RMSE) of \(2.360~\textrm{cm}\) , compared with \(2.233~\textrm{cm}\) and \(3.120~\textrm{cm}\) for the best-performing classical baseline. These results indicate that, in the adopted simulation model, a semantic-aware representation improves localization accuracy, task success rate, robustness to modeled channel degradation, and normalized energy per successful decision relative to the tested baselines.