IMU-Based Exercise Recognition and Angle Estimation via an Edge-Friendly Architecture
Alfonso Esposito, Ivan Zyrianoff, Yasamin Moghbelan, Marco Di Felice · 2025
Effective motor rehabilitation relies on correct exercise execution and timely feedback. Automatically assessing exercise quality remains a key challenge for remote rehabilitation systems. Although numerous approaches have been explored, each comes with its own set of limitations: camera-based systems can be costly and raise privacy concerns, while purely Inertial Measurement Unit (IMU)-based approaches often struggle with sensor drift and precise quality assessment. This paper investigates the feasibility of real-time exercise quality assessment using only simple wearable IMU sensors coupled with on-device edge intelligence. We propose a distributed, data-driven framework where Transformer-based Deep Learning (DL) models are deployed directly onto wearable IoT device placed on the patient joints when performing exercises. Each device processes local IMU data for initial exercise recognition and angle estimation. The system then consolidates these distributed inferences achieving a consensus on the performed exercise and to evaluate movement quality against physiotherapist-defined thresholds. Experimental results demonstrate high accuracy in exercise identification (97% F1-Score) and the capability for precise joint angle estimation (the MAE varied between 0.16 to 0.31 depending of the exercised performed).