A Machine Learning Model for Microcontrollers Enabling Low Power Indoor Positioning Systems via Visible Light Communication

Abstract

This paper presents a low-power Visible Light Localisation (VLL) Artificial Intelligence (AI)-enabled system for Indoor Positioning (IP) purposes. Compared to other IP techniques, VLL offers a similar positioning accuracy, but with the extremely desirable feature of low energy consumption, an aspect of primary relevance in the framework of Wireless Sensor Networks (WSN), self-sufficient sensing systems, Industry 4.0 and Internet of Things (IoT). The proposed system is composed of three modulated optical sources (i.e. LEDs) and a photodiode receiver mounted on the target to be localised. The localisation task is performed by processing the received light intensities through Machine Learning (ML) regression models trained with a set of data gathered during a calibration phase. The regressors are designed to be executed on a low-power microcontroller present in the target, hence establishing an embedded ML paradigm also preserving reduced power consumption features. The proposed models are trained exploiting datasets with different sizes, searching for a trade-off between the training set size, i.e. the duration and complexity of the calibration phase, and the maximum tolerable root mean square error (RMSE). In both cases, some localisation tests show that a satisfactory accuracy can be reached even with a limited complexity of the calibration procedure and that the obtained results fulfil the error constraint used for model design.