EMG Gesture Recognition: Generalization & Calibration

Posted on In , Word count in article: 392 Reading time ≈ 1 mins.
Discussing the severe cross-subject distribution shift in EMG signals and proposing a solution based on Universal Models and Few-Shot Rapid Calibration.

The Challenge of Cross-Subject Generalization

In the field of biometrics, Cross-Subject generalization is always a core difficulty. Significant differences in physiological structures (arm circumference, muscle position, subcutaneous fat thickness) between individuals lead to massive Domain Gaps in EMG signal data distribution.

Directly applying a model trained on Subject A to Subject B typically causes accuracy to plummet to around 50%, indicating that the model has not learnt universal gesture features independent of the individual. Previous attempts, even using Unsupervised Domain Adaptation (e.g., DANN), yielded suboptimal results.

Universal Model: From Individual to Population

To extract more robust common features, we changed our training strategy and adopted a Leave-One-Subject-Out (LOSO) approach to build a Universal Model.

We leveraged the diversity of the GRABMyo dataset (43 subjects) to train a Universal TCN model covering various arm anatomies.

  • The goal of this model is no longer to memorize specific individual patterns but to learn Invariant Representations of gesture actions across populations, such as the synergistic contraction laws of muscle groups for specific gestures.
  • In Zero-Shot testing, the average accuracy of this universal model improved to the 85%-92% range. Although there is still a gap compared to intra-subject testing, this proves the model possesses certain cross-subject generalization capabilities.

Rapid Calibration

To bridge the remaining performance gap and achieve product-level precision, we introduced a Registration/Calibration mechanism similar to biometric systems. We propose a Rapid Calibration Pipeline:

  1. Backbone Freezing: Keep the parameter weights of the Universal Model’s feature extraction part (Backbone) unchanged to preserve the learned general muscle patterns.
  2. Head Fine-tuning: Update only the parameters of the model’s final classifier layer (Linear Head).
  3. Few-Shot Data: Collect a very small amount of labeled data from the new user (e.g., 10 repetitions per gesture, about 1 minute of data).

Experimental Results

Testing on a completely unseen subject showed:

  • Before Calibration (Zero-Shot): Accuracy ~50.1% (in challenging cases).
  • After Calibration (Few-Shot): Average accuracy soared to 96.53%, with some users reaching 100%.

This result verifies the effectiveness of the “Universal Model + Few-Shot Calibration” paradigm. By combining universal features trained on big data with fine-tuning for specific users, we successfully solved the individual difference problem in EMG signals, realizing a system that has both generalization potential and the ability to adapt quickly to new users.