Efficient Tiny Machine Learning for Human Activity Recognition on Low-Power Edge Devices

Human Activity Recognition (HAR) continues to capture the attention of academic and industrial researchers because of its practical applications in healthcare and everyday living environments. To equate this technology for widespread use, it is crucial to ensure that HAR models are not only high-performing but also optimized for minimal resource usage. This paper employs Tiny Machine Learning (TinyML) techniques, including quantization, pruning, and knowledge distillation for reducing the model size and make a parsimonious use of resources and energy while maintaining high accuracy. The results demonstrate that the proposed approach can reduce the model footprint (4 × on average) with minimized accuracy deviation (4% on average). Additionally, this work extends the deployment of HAR models and the UCI-HAR dataset across three different hardware edge platforms: SensorTile.box PRO embodying multiple sensors, Raspberry Pi, and Arduino. The Arduino exhibited the most efficient use of the energy consumption while the Raspberry Pi the best accuracy.