EMG Gesture Recognition: Intra-Baseline

Posted on In , Word count in article: 314 Reading time ≈ 1 mins.
Establishing a TCN-based EMG gesture recognition baseline to verify signal separability and model capacity under ideal conditions.

EMG Gesture Recognition Baseline

Before conducting any generalization research, we must first establish a baseline: What is the upper limit of model performance under ideal conditions?

“Ideal conditions” refer to Intra-Subject Intra-Day testing:

  1. Consistent Subject: Training and testing sets come from the same person.
  2. Consistent Time: Electrode positions remain unchanged, impedance is stable.
  3. Consistent Distribution: Data distribution is not significantly affected by environmental or physiological changes.

Under such conditions, traditional machine learning methods (such as SVM or Random Forest) can typically achieve 85%-90% accuracy, indicating good signal separability.

Deep Learning Model Design

To further exploit signal features, we designed a baseline model based on 1D Temporal Convolutional Network (1D-TCN). Compared to manual feature extraction (e.g., zero-crossing rate, waveform length), an end-to-end deep learning model can learn complex non-linear patterns directly from raw voltage signals.

We used Session 1 data from the GRABMyo Dataset for 5-Fold Cross Validation.

The results are as follows:

  • Training Set Accuracy: 99.8%
  • Validation Set Accuracy: 99.0%

This demonstrates the model’s extremely high fitting capability in distinguishing 10 fine gestures (such as index finger flexion, wrist movements) within the same session.

Conclusion and Limitations

This baseline test reveals two key facts:

  1. Reliable Signal Quality: The preprocessing pipeline preserves sufficient effective information, and muscle electrical signals contain features distinguishable enough for complex actions.
  2. Sufficient Model Capacity: For data distribution of a single subject, a small-parameter CNN/TCN model is sufficient to achieve overfitting-level performance.

However, it is important to note that this 99% accuracy represents performance only under Stationary Distribution. In real-world applications, slight electrode shifts, skin condition changes, and cross-subject physiological differences will cause drastic Distribution Shifts.

Therefore, starting from the next post, we will explore the more challenging Cross-Day and Cross-Subject testing, which are the true touchstones for model robustness.