Related Experiment Video
Updated: Nov 17, 2025
08:15
Capturing Dynamic Finger Gesturing with High-resolution Surface Electromyography and Computer Vision
Published on: March 28, 2025
954
A Transferable Adaptive Domain Adversarial Neural Network for Virtual Reality Augmented EMG-Based Gesture Recognition
Summary
A new dynamic dataset for electromyography (EMG) gesture recognition bridges the gap between offline accuracy and real-time use. This dataset, evaluated with a novel algorithm TADANN, shows superior long-term performance over fine-tuning.
Area of Science:
- Biomedical Engineering
- Human-Computer Interaction
- Machine Learning
Background:
- Electromyography (EMG) based gesture recognition faces challenges in real-time application due to discrepancies between offline accuracy and practical usability.
- Existing datasets often lack real-world control dynamics like gesture intensity, limb position, electrode shift, and signal transients, or are tied to specific controllers.
Purpose of the Study:
- To introduce a novel dynamic dataset for EMG gesture recognition recorded using a real-time experimental protocol.
- To create an intermediate dataset type that bridges the gap between static offline and specific online datasets.
- To benchmark recalibration techniques for long-term EMG gesture recognition.
Main Methods:
- A virtual reality experimental protocol was designed to capture EMG data incorporating four dynamic factors: gesture intensity, limb position, electrode shift, and transient signal changes.
- An EMG-independent controller guided movements, ensuring user-in-the-loop recording for a dynamic, benchmark-ready dataset.
- The dataset was collected from 20 participants over 14-21 days, involving multiple recording sessions.
Main Results:
- The dynamic dataset was utilized to evaluate various recalibration techniques for across-day gesture recognition.
- A novel algorithm, TADANN (Temporal Adaptive Domain Adaptation Network), was introduced and tested as a recalibration method.
- TADANN demonstrated consistent and statistically significant superior performance compared to standard fine-tuning for long-term gesture recognition.
Conclusions:
- The proposed dynamic dataset effectively addresses limitations of existing EMG datasets, enabling more realistic performance evaluation.
- TADANN presents a significant advancement in recalibration strategies, improving the robustness and usability of EMG-based gesture recognition systems over extended periods.
- This work facilitates the development of more practical and reliable EMG-controlled interfaces.
