Markerless Pose Estimation and Explainable Machine Learning for Assessing Kicking Technique and Shooting Performance in Youth Football Players

Markerless pose estimation offers a practical route to field-based biomechanical assessment, but its ability to predict distinct dimensions of football kicking performance remains uncertain. This study evaluated 60 youth football players (48 male, 12 female; age 16.35 ± 1.19 years, range 14.31–18.28) from Romanian youth football programs, who contributed 1134 valid instep kicks across dominant- and non-dominant-leg and velocity- and accuracy-priority conditions. Ball velocity was measured by sports radar and shooting precision by continuous radial target error. Twelve prespecified markerless biomechanical variables were derived from synchronized multiview high-speed video and evaluated using participant-wise nested cross-validation; outer-fold predictions were generated only for players not used in model fitting or tuning. Velocity-priority trials increased ball velocity but also increased radial error, while non-dominant-leg kicks were slower and less accurate. For ball velocity, Elastic Net achieved an out-of-fold MAE of 1.19 m/s (95% CI 1.12–1.25), approximately 4.5% of the overall trial-level mean ball velocity (26.23 m/s), and R2 of 0.80 (95% CI 0.76–0.83); more complex nonlinear models provided no material improvement. Kicking-foot velocity, knee-extension angular velocity, and approach velocity were the most stable predictors. In contrast, radial-error prediction remained near baseline (Elastic Net R2 = 0.01, 95% CI −0.02 to 0.04). These findings support cautious use of the workflow for internally validated ball-velocity profiling in similar youth players, but not as a stand-alone tool for assessing shooting precision.