Cross-modal motion analysis: transferring motion-capture and force-plate features to video-based standing long jump prediction

The standing long jump is a widely used indicator of lower-limb explosive power and overall motor ability. Laboratory systems — optical motion capture and force plates — provide high-fidelity kinematic and kinetic measurements but are costly and constrained to controlled environments, limiting their routine or large-scale use. In contrast, markerless video-based pose estimation is inexpensive and easy to deploy but remains limited in prediction accuracy and physical interpretability. To bridge this gap, we propose a Teacher–Student cross-modal knowledge transfer framework. We first construct a feature set of 35 kinematic and kinetic variables from synchronized Qualisys motion-capture and AMTI force-plate recordings and train/compare four traditional regressors (XGBoost, Random Forest, Gradient Boosting, and MLP) under a leave-one-subject-out (LOSO) evaluation. We then train a Student model — a TCN-like temporal convolutional network — that takes 2D keypoint time series from video as input and is trained with a multi-task mean squared error loss to jointly regress the final jump distance and to reconstruct the Teacher's intermediate physical features (feature-level distillation). With a limited dataset (N = 10, one trial per subject), results indicate that the Teacher model provides reasonable score estimates from laboratory data, and that incorporating Teacher-level feature distillation improves video-based score estimation and helps preserve physical interpretability of the Student model. The proposed pipeline is applicable both for laboratory-grade modelling and as an interpretable, low-cost approach for video-based field assessment.