Spatiotemporal key point detection in golf swing sequences via hybrid CNN-TCN architecture and regression-based refinement

Finding structural keypoints is crucial for precise sports motion analysis, especially in golf, where subtle biomechanical variations influence performance. This study introduces a hybrid deep learning framework that integrates convolutional neural networks (CNNs) for spatial feature extraction and Temporal Convolutional Networks (TCNs) for dynamic modeling, enhanced by a regression-based refinement step for sub-pixel accuracy. We also extend the public GolfDB dataset by manually annotating 17 anatomical joints per frame, creating the first fine-grained keypoint-level benchmark for golf swing analysis. The proposed model achieves superior performance with a Mean Absolute Error (MAE) of 2.8 px, PCK@0.1 = 94.5% (Percentage of Correct Keypoints), and real-time inference speed (14.9 ms per frame). These results demonstrate improved spatial precision, temporal smoothness, and practical applicability for real-time coaching evaluation. This work establishes a reproducible foundation for fine-grained sports pose estimation and dataset development in the golf domain.