Markerless pose-derived movement quality assessment in physical education: a multi-task student performance analysis

Markerless motion capture offers a practical opportunity to support objective movement assessment in physical education, but its use for student performance evaluation remains limited by concerns about landmark reliability, task variability, camera-view effects, and insufficient validation. This study developed and internally validated a markerless pose-derived framework for evaluating student movement quality across common physical-education tasks. A publicly available synchronized marker-based and markerless motion-capture dataset was analyzed. The final analytical dataset included 21 male physical-education students, seven tasks, two camera views, and 294 eligible task-view trials. Markerless keypoints following the COCO-17 anatomical landmark convention were reconstructed into frame-level skeletons, cleaned using confidence-based filtering, interpolated for short gaps, smoothed, and normalized by body scale. Trial-level features were extracted to quantify movement amplitude, symmetry, smoothness, rhythm, postural control, and detection reliability. A Markerless Movement Quality Score was generated using equal-weighted, expert-weighted, and PCA-derived scoring strategies. Trials were stratified into low, moderate, and high pose-derived movement-quality groups within each task-view combination. Internal validation used grouped 5-fold, leave-one-subject-out, final subject holdout, leave-one-task-out, and leave-one-view-out strategies. The pipeline reconstructed 951,067 frame-level skeletons and more than 16 million keypoint observations. Trunk and lower-limb landmarks showed high reliability, with hips, knees, ankles, and shoulders demonstrating near-complete tracking availability, whereas facial landmarks were less stable. Task-level features captured distinct movement signatures across walking, jogging, high knees, jumping jacks, lunges, and squats. The scoring methods showed strong agreement, including Spearman correlations of 0.97 between equal-weighted and expert-weighted scores, 0.96 between equal-weighted and final scores, and 0.94 between PCA-derived and final scores. The PCA performance landscape showed partial separation of low, moderate, and high movement-quality groups. Machine-learning validation was modest, with the best grouped 5-fold Macro-F1 reaching approximately 0.39, and weaker cross-view generalization. The confusion matrix showed better recognition of low-quality trials than moderate or high-quality trials. Markerless pose-derived features can provide interpretable trial-level descriptors of movement quality across physical-education tasks, particularly when based on reliably detected trunk and lower-limb landmarks. However, predictive validation was modest, with limited discrimination between adjacent movement-quality groups and weak generalization across some task and camera-view conditions. The internally derived MMQS groups were not independently expert-rated, and the present results therefore support feasibility and internal methodological validation rather than reliable automated grading or generalizable performance prediction. External validation against expert assessments, marker-based trajectories, larger cohorts, and real classroom conditions is required before practical educational deployment.