TinyML Architecture for Indoor Localization and Posture Recognition in Older Adults: Validation in a Clinical Simulation Environment

Abstract

The elderly population requires continuous health monitoring solutions that are non-invasive and preserve independence in daily living environments. Current approaches fail to integrate real-time indoor localization with posture recognition on low-cost embedded platforms, limiting their deployment in assisted living environments. This article presents a distributed edge computing architecture based on ESP32-C3 microcontrollers and TinyML. The system integrates indoor localization and posture recognition on the same embedded platform using a state machine strategy that performs indoor localization via RSSI fingerprints and posture recognition via IMU sensor fusion. The system comprises four Room Beacons built with ESP32-WROOM-32E modules communicating via the ESP-NOW protocol, and a Smart Tag AI device equipped with a BNO055 inertial measurement unit running two quantified deep neural networks for on-device inference. Post-training quantification with INT8 precision achieved 91.53% accuracy for room-level localization using 5014 RSSI samples and 100% accuracy for posture classification using 2048 IMU samples, including 201 supine and 209 seated instances, with inference latencies of 0.25 milliseconds and 0.18 milliseconds, respectively, while reducing the model size from 40.9 kilobytes to 17.8 kilobytes. The system was validated in a Clinical Simulation Environment at Universidad del Rosario with older adult participants, demonstrating its real-world feasibility. This edge-based approach eliminates cloud dependency and supports scalable environmental assistance deployments for elder care monitoring.