Multi-Sensor Wearable Device With Transformer-Powered Two-Stream Fusion Model for Real-Time Leg Workout Monitoring.

Abstract

Leg workout-based monitoring provides valuable insights into physical and neurological health, supporting healthcare professionals and facilitating in-depth analysis. However, current single sensing modalities technologies are limited by size constraints, environmental sensitivity, and accuracy issues. Furthermore, despite the widespread use of deep learning (DL) methods for sensor-based gesture recognition methods, they still encounter challenges in feature extraction. To address the limitations, this study 1) presents the development of a multi-modal wearable device for leg workout monitoring with real-time gait analysis capabilities, 2) introduces a novel Transformer-powered Two-Stream Fusion, namely TTSF, for efficient and accurate extraction of temporal and spatial features. The experimental results on our leg workout dataset demonstrate the superior performance of the proposed TTSF model with Precision, Recall, and F1-Score values of 90.7%, 90.6%, and 89.1%, respectively. Overall, this research contributes to the advancement of using multi-sensor fusion with DL and Medical Internet of Things (MIoT) techniques for advanced gait monitoring and analysis. These techniques have potential applications in personalized training programs and enhanced rehabilitation assessment.