# TraHGR: Transformer for Hand Gesture Recognition via ElectroMyography

**Authors:** Soheil Zabihi, Elahe Rahimian, Amir Asif, Arash Mohammadi  
**arXiv:** [2203.16336](https://arxiv.org/abs/2203.16336)  
**Date:** 2022-03-28  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/2203.16336v2/

## Key points

- Deep learning-based Hand Gesture Recognition (HGR) via surface Electromyogram (sEMG) signals is promising for myoelectric-controlled prostheses.
- Existing deep learning approaches struggle with generalization in changing scenarios.
- The TraHGR hybrid framework leverages hybrid models and transformers for improved performance.
- TraHGR architecture includes two parallel paths and a fusion center for robustness across scenarios.
- Evaluation on Ninapro dataset (DB2) shows recognition accuracies of 86.18% on 49 gestures, outperforming state-of-the-art results.
- Analysis of different window sizes (200ms, 150ms, 100ms) impacts model performance and complexity.
- Statistical analysis demonstrates significant accuracy improvements with TraHGR for 200ms window size.
- Position-wise cosine similarity analysis highlights the effectiveness of TraHGR in capturing similarities between embeddings.
- Comparison table showcases TraHGR's superiority over previous DNN approaches in terms of performance across window sizes and gesture subsets within the DB2 dataset.

## Summary

, , , , 


Deep learning-based Hand Gesture Recognition (HGR) via surface Electromyogram (sEMG) signals has shown significant potential for the development of advanced myoelectric-controlled prostheses. However, existing deep learning approaches often struggle to maintain acceptable generalization performance in changing scenarios. To address this challenge, this paper proposes a hybrid framework called the Transformer for Hand Gesture Recognition (TraHGR), which leverages recent advances in hybrid models and transformers. The TraHGR architecture consists of two parallel paths followed by a fusion center to integrate the advantages of each module and provide robustness across different scenarios. The proposed model was evaluated on the second Ninapro dataset (DB2), which includes sEMG signals from 40 healthy users performing 49 gestures in real-life conditions. The paper analyzes the effect of different window sizes (200ms, 150ms, and 100ms) on overall performance and model complexity. TraHGR achieves recognition accuracies of 86.18% on DB2 (49 gestures), 88.91% on DB2-B (17 gestures), 81.44% on DB2-C (23 gestures), and 93.84% on DB2-D (9 gestures). These results outperform state-of-the-art performance for each subset of gestures in the dataset. Statistical analysis was conducted to compare TraHGR with other proposed architectures, showing statistically significant improvements in accuracy for window sizes of 200ms. Position-wise cosine similarity analysis further demonstrates the effectiveness of TraHGR in capturing similarities between different embeddings. A comparison table showcases the superiority of TraHGR over previous DNN approaches, highlighting its superior performance across different window sizes and gesture subsets within the DB2 dataset. In conclusion, the proposed TraHGR architecture offers a novel approach to hand gesture recognition using sEMG signals, achieving high accuracy rates across various scenarios and outperforming existing methodologies in the field.