Machine Vision and Applications Прикладные компьютерные науки
A view-normalized MediaPipe BlazePose framework for basketball movement stability quantification: a paired algorithmic validation study in Japanese collegiate players
Аннотация
Field-based basketball assessment requires affordable movement descriptors with explicit measurement limits. We analyzed 132 Japanese collegiate players and 792 retained shuffle-to-cut recordings, each processed with minimal-quality-control BlazePose (M0) and a complete filtering, body-scale, athlete-centered coordinate and landmark-plausibility pipeline (M4). Author-reported aggregate jitter values differed (3.42 vs. 2.15 mm/frame); their interpretation as physical error reduction requires common-unit confirmation across processing stages. The bounded internal alignment display score changed from 0.52 to 0.74. Agreement between calibrated task-specific PDMSI and coach ratings was ICC 0.51 for M0 and 0.74 for M4, with an M4 athlete-disjoint summary of 0.71 subject to calibration-only fitting. An auxiliary mixed-effects indirect association was 0.24 (95% cluster-bootstrap CI 0.04–0.45); shared pose-derived inputs preclude unique causal inference. An exploratory 10° projected-angle rule had F1 0.836 against expert frame labels, without anatomical or clinical validation. A Pixel 6 achieved reported throughput of 26.1 frames/s and compute-only time of 44 ms, which do not demonstrate sustained 60-frame/s feedback. These results characterize a controlled task-specific processing comparison, not athlete improvement or verified three-dimensional biomechanics.