TinyML-Driven On-Device Personalized Human Activity Recognition and Auto-Deployment to Smart Bands
Bidyut Saha, Riya Samanta, Soumya K. Ghosh, Ram Babu Roy
Abstract
Human activity recognition has revolutionized health and fitness monitoring, although significant hindrances still exist. One such difficulty is making a system that considers each person’s unique physical activity features, routines, and preferences to provide personalized outcomes. Another challenge is implementing on-device computing models that balance latency, energy use, performance, and accuracy. To bridge this research gap, in this paper, BandX 2.0, a wrist-worn, customized human activity recognition smart band, is presented. BandX 2.0 utilizes IMU sensors, along with a TinyML-driven on-device computing paradigm. It has preloaded activity classes that can be customized based on an individual’s unique movement style. Users can add any number of activity classes by performing and recording the activities for a short duration, thus incorporating minimal calibration and intervention. For personalised human activity recognition, a lightweight 1D Convolutional Neural Network is constructed by employing a transfer learning strategy, along with fine-tuning, using a small dataset gathered from the user’s environment. This approach led to a 37% increase in the accuracy of human activity recognition compared to generalized models. The evaluation of BandX 2.0’s performance was conducted by leveraging three benchmark datasets: WISDM, PAMAP2, and the BandX dataset. Furthermore, this work proposed a cloud-supported framework for the automatic wireless deployment of the TinyML model to remote wearables. The framework facilitates model customization in a cloud environment and enables on-device inference, even when limited target data is available.