Embedded Machine Learning for 3D Indoor Visible Light Positioning via Optimized Fingerprinting

In this paper, a low power node featuring 3D indoor localization via Visible Light Positioning (VLP) and embedded Machine Learning (ML) is presented. The coordinates estimation is performed by a low-complexity shallow Neural Network (NN) running on board of a microcontroller and which approximates the regression model linking received light intensities and position in the workspace. The received signal strengths (RSSs) of the optical signals, coming from four LEDs chopped at unique frequencies and used as fixed anchors, are evaluated via fast Fourier transform (FFT). The generation of the datasets used to test and train the NN exploits a novel fingerprinting procedure which combines the simulation of the data through the canonical model for the light propagation and the acquisition of measurements in the workspace. Field tests performed in a $1 \mathrm{m}\times1 \mathrm{m}\times0.2\mathrm{m}$ volume prove an overall mean accuracy of about 3 cm with standard deviation of 1.5 cm and maximum error below 6 cm.