Enhanced TransFormer-Based Algorithm for Key-Frame Action Recognition in Basketball Shooting

This paper proposes an enhanced TransFormer-based algorithm for key-frame action recognition in basketball shooting. The research addresses the challenges of accurate temporal localization and feature extraction in complex basketball environments through architectural innovations in deep learning models. The proposed approach integrates a multi-scale feature fusion mechanism with an improved spatio-temporal attention structure, enabling robust recognition of basketball shooting actions across varying conditions. A novel position encoding scheme is introduced to better capture temporal relationships in shooting sequences, while the enhanced attention mechanism facilitates more precise key-frame identification. Experimental evaluations on basketball shooting datasets demonstrate that the proposed model achieves 92.8% accuracy in action recognition tasks, outperforming existing approaches by 4.3% in mean average precision. The architecture maintains computational efficiency while improving recognition accuracy, processing video sequences in real-time at 30 frames per second. Ablation studies confirm the effectiveness of individual components, with the spatio-temporal attention mechanism contributing the most significant performance gains. The system demonstrates robust performance across different shooting styles and environmental conditions, making it suitable for practical applications in basketball training and analysis.