sEMG-Based Deep Metric Learning With Regulated Centroid-Nested Triplet Loss: From Hand Gestures to Elite Soccer Drills in the English Premier League

sEMG-based motion classification, traditionally applied for hand-gesture recognition (HGR) in prosthetics, presents transformative potential in sports science. Its broader application, however, is limited by the lack of extensive sports-specific data. This study adopts a unique few-shot metric learning (FSL) framework using a laboratory-collected hand-gesture dataset (captured with Delsys Trigno sensors). The method is then evaluated on a real-world sports dataset from the lower extremities, acquired via Neurocess sEMG sensors. This soccer dataset comprises data collected from 45 elite athletes representing two English Premier League (EPL) clubs. In addition, this article presents the novel regulated centroid-nested triplet loss (
$\mathcal {L}_{\text {RCTL}}$
) function, which addresses the limitations of the traditional triplet loss. Our proposed FSL technique achieves a remarkable 90% accuracy in classifying nine distinct soccer motions using just 50 shots, registering a 41% enhancement over traditional feed-forward learning. Furthermore, this article’s innovative
$\mathcal {L}_{\text {RCTL}}$
loss function significantly outperforms the conventional triplet loss.
$\mathcal {L}_{\text {RCTL}}$
achieves accuracies of 65%, 78%, 83%, and 90% for 5-, 10-, 20-, and 50-shot testing scenarios, marking improvements of 13%, 12%, 5%, and 4% over prevailing methods. The findings of this study demonstrate the effectiveness of using the Ninapro dataset through the FSL scheme in sEMG-based motion classification for sports science. Notably, this approach showcases reproducibility and adaptability, becoming a reliable solution, particularly in situations where dataset size is limited.