Markerless Gymnastics Score Prediction Using Biomechanical Feature-Group Attention SVR

Objective assessment of technical execution in artistic gymnastics remains challenging because scoring relies on expert visual judgement of posture, movement control, alignment, symmetry, and action-specific execution quality. This study developed a markerless, video-based, and biomechanically interpretable machine learning framework for estimating gymnastics execution scores from pose-derived movement features. A total of 500 men’s artistic gymnastics floor-element video sequences were collected under controlled conditions and independently scored by three FIG-certified judges, with the mean score used as the ground-truth label. Pose trajectories were extracted using MediaPipe Pose, and judging-relevant landmarks were retained to represent key body segments. After hip-centred coordinate normalisation, biomechanical features were constructed from four groups: position magnitude, joint angle, temporal velocity, and left–right symmetry. As the core prediction method, this study proposes a Biomechanical Feature-Group Attention Support Vector Regression model (BFG-AttSVR). BFG-AttSVR introduces an action-specific feature-group attention mechanism to learn different group-level importance weights for different gymnastics actions. Within BFG-AttSVR, the attention weights are learned through a two-stage optimisation strategy, including data-driven initialisation based on cross-validated Ridge regression and gradient-free fine-tuning using an RBF-kernel SVR objective. Experimental results showed that BFG-AttSVR achieved an MAE of 0.0465 ± 0.0041, an RMSE of 0.0621 ± 0.0074, and an R2 of 0.9710 ± 0.0131, outperforming the strongest SVR (RBF) baseline with an RMSE of 0.0743 ± 0.0071. The proposed framework provides a low-cost, objective, and interpretable approach for video-based gymnastics score estimation.