| Change-of-direction (COD) movements are critical components of performance and injury risk in field sports. However, the development of automated systems to classify COD movements typically requires large volumes of labelled data, which is costly and labour-intensive to obtain. This study investigates the application of self-supervised learning (SSL) to mitigate this challenge by leveraging unlabelled data collected from live match-play to support COD classification on a smaller, labelled dataset from a controlled environment. Three SSL frameworks, Autoencoder-based, SimCLR, and BYOL, were evaluated by utilising pretrained encoders as fixed feature extractors for downstream classification. Their performance was compared against fully supervised baselines using Random Forest (RF), Support Vector Machine (SVM), and Logistic Regression (LR). While supervised models (particularly RF) established the upper performance bound, the Autoencoder-based framework produced the most transferable representations among the SSL methods and consistently outperformed both SimCLR and BYOL. The results indicate that SSL can effectively support COD classification when labelled data is limited, though its efficacy is highly dependent on the choice of pretext objective and the alignment between the pretraining and downstream domains. |
*** Title, author list and abstract as submitted during Camera-Ready version delivery. Small changes that may have occurred during processing by Springer may not appear in this window.