Tiny-ML based activity recognition combined with indoor positioning using ultra-wideband sensors for elderly care

Elderly care systems play a crucial role in ensuring the well-being of older adults, as nurses cannot constantly monitor them manually. Elderly activity monitoring systems are therefore invaluable, especially in indoor environments where precise activity recognition at low power and cost is essential. Tiny Machine Learning (TinyML) technology, which enables intelligent low-power Microcontroller Units (MCUs), significantly enhances activity recognition, contributing to more efficient elderly care systems. Combining Indoor Positioning Systems (IPS) with TinyML-based activity recognition can create a highly accurate and cost-effective elderly care system. This study focuses on improving the precision of Ultra-Wideband (UWB) technology for indoor positioning and emphasizes the importance of TinyML in elderly care applications, exploring the integration of IPS with TinyML for comprehensive activity monitoring. The findings indicate substantial advancements and provide valuable insights for future developments. The indoor positioning system was evaluated in three scenarios: Line-of-Sight (LoS), Obstructed-Line-of-Sight (OLoS), and a real-world simulation where the tag was concealed in an individual's pocket, achieving average positioning errors of 17.21 cm, 48.27 cm, and 46.17 cm, respectively. The highest accuracy was observed under LoS conditions, while challenges arose in OLoS and pocket-carrying scenarios due to electromagnetic wave propagation through obstructions. Despite these challenges, the system demonstrated satisfactory accuracy for indoor tracking, indicating its potential for practical deployment. The TinyML model for activity recognition achieved an exceptional accuracy of 97.22\% on the testing dataset, supported by a robust confusion matrix and F1 Score, underscoring its reliability in real-world applications where accurate activity recognition is crucial. The integrated system, combining IPS and activity recognition, was tested across 47 activities over 11 rounds, achieving an overall accuracy of 91.48\%. The system excelled in identifying activities involving continuous and minimal movement, making it advantageous for elderly activity monitoring and healthcare settings. While challenges were noted in scenarios requiring differentiation within shorter observation windows, the system's overall performance confirmed its suitability for practical implementation in real-world environments.