Predicting continuous hand pose from wearable EMG sensor data using transformer-based deep-learning models
Saved in:
| Title: | Predicting continuous hand pose from wearable EMG sensor data using transformer-based deep-learning models |
|---|---|
| Authors: | McDowell, Jason |
| Committee Members: | Kaplanoglu, Erkan; Varol, Serkan; Abrha, Wolday D.; College of Engineering and Computer Science |
| Summary: | This thesis investigated Transformer-based deep-learning models for predicting continuous hand pose from electromyography (EMG) signals collected with a low-cost, eight-channel wearable armband. A modular software laboratory was developed to support data acquisition, synchronization, visualization, model training, and inference. A single-subject, three-hour dataset of synchronized EMG and hand-tracking data was collected, with hand pose represented both as 15-dimensional joint flexion angles and 84-dimensional finger-bone orientation quaternions in a hand-centered frame. Compared with a Long Short-Term Memory (LSTM)-based model, the Transformer-based model reduced whole-hand median prediction error from 3.6° to 3.3° for a joint angle model and from 17.4° to 15.1° for a bone orientation model. Experimental results demonstrated that Transformer models outperformed LSTM models in both median and 90th-percentile prediction error, particularly for angle-based outputs. These findings support the use of Transformer architectures for accurate, continuous hand-pose estimation with wearable EMG, relevant to applications in prosthetics and human-machine interfaces. |
| URL: | https://scholar.utc.edu/theses/1019 |
| Database: | OpenDissertations |
| Abstract: | This thesis investigated Transformer-based deep-learning models for predicting continuous hand pose from electromyography (EMG) signals collected with a low-cost, eight-channel wearable armband. A modular software laboratory was developed to support data acquisition, synchronization, visualization, model training, and inference. A single-subject, three-hour dataset of synchronized EMG and hand-tracking data was collected, with hand pose represented both as 15-dimensional joint flexion angles and 84-dimensional finger-bone orientation quaternions in a hand-centered frame. Compared with a Long Short-Term Memory (LSTM)-based model, the Transformer-based model reduced whole-hand median prediction error from 3.6° to 3.3° for a joint angle model and from 17.4° to 15.1° for a bone orientation model. Experimental results demonstrated that Transformer models outperformed LSTM models in both median and 90th-percentile prediction error, particularly for angle-based outputs. These findings support the use of Transformer architectures for accurate, continuous hand-pose estimation with wearable EMG, relevant to applications in prosthetics and human-machine interfaces. |
|---|
