BandX : An Intelligent IoT-band for Human Activity Recognition based on TinyML
Bidyut Saha, Riya Samanta, Soumya K. Ghosh, Ram Babu Roy
Abstract
Human Activity Recognition (HAR) is used in several human-centric real-world situations. In many learning-based systems, sensors capture situational data and transfer it to adjacent computing devices for processing and analysis. This concerns user data privacy and network availability, as well as increasing response latency and unnecessary power consumption. To address these, we developed a device BandX that classifies human activities from wearable sensors using an on-device deep learning inference mechanism on a microcontroller unit (MCU) based on TinyML. BandX performs with an accuracy of around 94% and reduces significant network traffic in comparison to computational offloading methods. We have also designed a framework that manages BandX sensor/inference data and provides a logical layer for user activity-level inferences. This framework includes an interactive user interface that lets people track their activity patterns and other health statistics.