Human Activity Recognition of Figure Skating Jumps Using Abdominal Electromyography and Accelerometry

Abstract Automated recognition of figure skating jumps remains challenging for computer vision due to rapid body rotations and occlusions. This study introduces a novel multimodal Human Activity Recognition (HAR) framework combining triaxial accelerometry with abdominal surface electromyography (sEMG) to capture core muscle pre-activation (m. obliquus externus) and upper-body kinematics across six multi-rotational jump types (Axel, Salchow, Loop) recorded from twelve competitive female adolescent skaters. Systematic investigation of temporal window sizes and sliding overlap configurations established an optimal setup of 600 samples with a 25-sample overlap. Evaluating five machine learning architectures showed that tree-based ensembles on handcrafted features significantly outperformed deep learning models. The optimized Gradient Boosting model achieved a peak test accuracy of 98.81% and a macro F1-score of 98.80%. A sensor ablation study confirmed that lateral and vertical accelerations combined with bilateral core sEMG are critical for jump differentiation. By achieving state-of-the-art accuracy using a lightweight, single-node sensor setup, this approach minimizes athlete payload while enabling real-time wearable sports analytics.