Related Experiment Video
Updated: Aug 8, 2025
08:15
Capturing Dynamic Finger Gesturing with High-resolution Surface Electromyography and Computer Vision
Published on: March 28, 2025
684
Reducing the Energy Consumption of sEMG-Based Gesture Recognition at the Edge Using Transformers and Dynamic
Chen Xie1, Alessio Burrello2,3, Francesco Daghero1
1Department of Control and Computer Engineering, Politecnico di Torino, 10129 Turin, Italy.
Sensors (Basel, Switzerland)
|February 28, 2023
Summary
This study introduces bioformers, tiny transformer models for surface electromyographic (sEMG) hand gesture recognition, achieving higher accuracy and significant energy savings on edge devices. A dynamic inference system further optimizes energy consumption without compromising performance.
Area of Science:
- Biomedical Engineering
- Machine Learning
- Edge Computing
Background:
- Surface electromyographic (sEMG) signals are crucial for hand gesture recognition.
- On-device execution of deep learning (DL) models is desired for wearables but faces challenges with energy and memory constraints.
- Optimizing DL model architecture and runtime execution is essential for efficient edge deployment.
Purpose of the Study:
- To develop a novel, efficient, and accurate hand gesture recognition system for edge devices using sEMG signals.
- To introduce and evaluate tiny transformer models, termed 'bioformers', for sEMG-based gesture recognition.
- To implement a dynamic inference strategy for further energy optimization at runtime.
Main Methods:
- Applied tiny transformer models (bioformers) to sEMG gesture recognition, exploring various architectures.
- Evaluated model performance on the Ninapro DB6 dataset, comparing against state-of-the-art Convolutional Neural Network (CNN) models.
- Deployed models on a RISC-V low-power System-on-Chip (SoC), GAP8, and implemented a three-level dynamic inference approach combining a random forest (RF) classifier with bioformers.
Main Results:
- The most accurate bioformer achieved +3.1% higher classification accuracy than the state-of-the-art TEMPONet on the Ninapro DB6 dataset.
- Bioformers demonstrated 7.8×-44.5× lower energy consumption per inference on GAP8 compared to TEMPONet at similar accuracy levels.
- The dynamic inference approach provided an additional 1.03×-1.35× energy reduction on GAP8 at iso-accuracy, offering flexible operating points.
Conclusions:
- Bioformers represent a highly accurate and energy-efficient alternative to CNNs for sEMG hand gesture recognition on edge devices.
- The proposed dynamic inference system enhances energy efficiency by adaptively selecting model complexity based on input data.
- The integrated system offers a significant advancement in wearable gesture recognition technology, balancing accuracy and power consumption.
Related Concept Videos
01:21Energy Losses in Transformers
925
In an ideal transformer, it is assumed that there are no energy losses, and, hence, all the power at the primary winding is transferred to the secondary winding. However, in reality, the transformers always have some energy losses, and, hence, the output power obtained at the secondary winding is less than the input power at the primary winding due to energy losses.
There are four main reasons for energy losses in transformers.
The first cause can be the high resistance of the...
There are four main reasons for energy losses in transformers.
The first cause can be the high resistance of the...
925
01:25Transformers with Off-Nominal Turns Ratios
187
In scenarios involving parallel transformers with disparate ratings, developing per-unit models requires accommodating off-nominal turns ratios. This situation arises when the selected base voltages are not proportional to the transformer’s voltage ratings. Consider a transformer where the rated voltages are related by the term a. If the chosen voltage bases satisfy a relationship involving term b, term c is defined as the ratio of these bases. This ratio is then substituted into the...
187