This paper presents COZDAL (Convolutional Octave-band Zooming-in with Depth-kernel Attention Learning), a versatile deep-learning model designed for surface Electromyography (sEMG) motion classification. Specifically focusing on sports movements involving the hamstring muscle, the model employs attention mechanisms across various frequency bands, kernel sizes, and hidden layer depths. The proposed method has been extensively evaluated on the benchmark Ninapro dataset and a custom soccer dataset. The results demonstrate substantial improvements over the existing state-of-the-art models, with an accuracy of 95.30% on Ninapro DB2, outperforming the previous best by 3.29%, and an accuracy of 98.80% on Ninapro DB2-B, a 8.6% enhancement. Remarkably, COZDAL exhibits a performance accuracy of 96.30% on a soccer dataset gathered from 45 elite-level athletes representing two clubs in the English Premier League (EPL). This result, achieved without parameter tuning, highlights the model’s adaptability and exceptional efficacy across diverse motion scenarios, sensors, subjects, and muscle types.

Motion classification with surface electromyog- raphy (sEMG) has been studied for practical applications in prosthesis limb control and human-machine interaction. Recent studies have shown that feature learning with deep neural networks (DNN) reaches considerable accuracy in motion classification tasks. However, DNNs require large datasets for acceptable performance and fail for tasks with few data samples available for training. Professional athlete training includes hundreds of exercises, and coupled with privacy and confidentiality issues acquiring a large dataset for all the exercises is not feasible. As a result, state-of-the-art DNN architectures are unsuitable for real-life sports applications. We utilise few-shot learning (FSL) techniques to overcome the small dataset problem of sports-related motion classification tasks. The employed methodology uses the knowledge gathered from a large set of tasks to classify unseen tasks with a few data samples. The FSL approach with a siamese network and triplet loss reached the best performance with a median F1-score of 72.01%, 76%, and 79% for 1, 5 and 10 shot datasets that include an unseen set of tasks, respectively. In contrast, DNN with transfer learning (TF) reached 49.27%, 51.58%, and 67.66% for the same set of tasks, respectively.