Reproduce “Decoding High-Dimensional Finger Motion from EMG Using Riemannian Features and RNNs” on OpenResearch

About this paper

Martin Colot, Cédric Simar, Guy Cheron, Ana Maria Cebolla Alvarez, Gianluca Bontempi

Researchers developed a cost-effective framework for EMG-based hand gesture recognition, introducing the EMG-Finger-Kinematics (EMG-FK) dataset of unconstrained movements. They propose the Temporal Riemannian Regressor (TRR), a lightweight recurrent neural network leveraging Riemannian features, which achieved an average absolute error of 9.79° on EMG-FK and demonstrated real-time, low-power inference on embedded systems, outperforming state-of-the-art baselines.

Key insights

• Riemannian covariance matrices, especially when computed across multiple frequency bands, provide a robust and effective feature representation for EMG signals, capturing complex muscle activation patterns more comprehensively than traditional time-domain or deep convolutional features for kinematics regression.
• The incorporation of temporal context through recurrent neural networks is crucial for accurately decoding the dynamic nature of high-dimensional finger movements from EMG signals, demonstrating superior performance over non-recurrent and purely convolutional architectures.
• A human-in-the-loop data acquisition protocol, where early model predictions guide participants to explore less-represented gestures, effectively enriches the dataset and improves the model's ability to generalize to diverse joint configurations.

Summary by alphaXiv.