AI-Enabled IoT for Hybrid Ultrasonic–RFID Indoor Positioning

Indoor positioning is a key enabler for many Internet of Things (IoT) applications, where high accuracy, low power consumption, and limited infrastructure cost must coexist. This paper presents a fully embedded implementation and experimental validation of the ultrasonic–RFID hybrid indoor localization architecture previously introduced in Merenda et al., 2022, in which standard UHF RFID infrastructure simultaneously provides synchronization and a backhaul channel for position data, while an ultrasonic front-end performs time-of-flight ranging. A microcontroller-based Beacon unit sequentially emits linear up-chirp signals from four ceiling-mounted transducers, and a mobile device equipped with an ultrasonic MEMS microphone acquires the waveform and extracts four peak-time features. These features feed a lightweight multi-output regression neural network deployed on the same microcontroller via an embedded machine learning workflow. The model is trained on a dataset of more than 32k frames acquired with a 6-DoF robotic arm in a Vicon-instrumented laboratory, using sub-millimeter optical ground truth as supervision. Experimental results show that the proposed edge implementation achieves a mean Euclidean positioning error of 7.83 cm, with 95% of estimates within 15.13 cm, while providing accuracy comparable to a traditional cross-correlation-based multilateration pipeline but with significantly reduced computational load, enabling update rates close to 10 Hz with millisecond-level latency. The resulting architecture is fully battery-powered, scalable, and suitable for AI-enabled IoT indoor positioning scenarios.