Football Teaching Action Recognition via Temporal Consistency Based Data Augmentation

Instructional action recognition under small-scale and class-imbalanced datasets presents significant challenges in sports education applications. This paper proposes a temporal consistency-based data augmentation strategy to enhance video action recognition models in football teaching scenarios. The method applies localized frame perturbation combined with a motion flow consistency constraint to generate temporally coherent and semantically valid augmented sequences. An I3Dbased recognition framework is employed, with joint optimization of classification and consistency losses. Experiments conducted on a custom dataset constructed from UCF101 and real classroom footage demonstrate that the proposed method significantly improves recognition accuracy and robustness. Ablation studies and t-SNE visualizations validate the contribution of each component and the importance of maintaining rhythmic integrity during augmentation. The approach shows strong potential for deployment in intelligent teaching systems with limited training resources.