Forearm Orientation Invariant Hand Gesture Recognition using sEMG and Feature Driven Machine Learning

Achieving reliable hand gesture recognition using surface electromyography (sEMG) across varying forearm orientations remains a major challenge in the development of robust human–machine interfaces. This study explores the impact of feature engineering, classifier selection, and training strategies on orientation-invariant gesture recognition. We evaluated five feature extraction methods and three classical machine learning classifiers using a custom sEMG dataset recorded across pronation, rest, and supination orientations. Results show that feature type and training orientation have a significant influence on performance, while classifier choice has a minimal effect. Signal Normalized Time Domain Features (SNTDF) consistently outperformed other features, and Linear Discriminant Analysis (LDA) proved to be an effective and stable classifier. Training with multiple orientations significantly improved generalization, with the highest F1 score (97.78%) achieved when using all orientations. Deep learning models were also tested, but classical approaches with engineered features delivered superior accuracy. These findings offer practical insights for designing orientation-invariant, sEMG-based control systems for wearable and assistive technologies.