Validation of Pose Estimation Models for Automated Kickboxing Technique Analysis Using 2TRAX3

Abstract

Biomechanical feedback can improve sports technique, but the 3D motion capturing gold standards typically used to generate such feedback are rarely available in everyday training. This study evaluates how the choice of pose estimation model affects automated movement analysis of the double side kick in Pointfighting, Kickboxing using ordinary 2D side-view video. In doing so, the study aims to determine whether advances in pose-estimation models translate into meaningful improvements in complete, sport-specific movement-analysis systems under realistic training conditions. The analysis compared MoveNet, YOLO11x-Pose, ViTPose-L, and RTMPose-L on 27 videos recorded under varying lighting conditions. Seventeen movement characteristics extracted from the videos were compared with 3D motion-capture measurements of the same kicks. Friedman tests with Wilcoxon signed-rank post-hoc tests showed that model choice significantly affected phase-duration errors (p<0.01, Kendall’s W=0.150) and femur-angle errors (p<0.01, Kendall’s W=0.174), but not leg-spreading angles or distance- and height-based characteristics. ViTPose-L achieved the lowest median errors for phase durations (0.043 s, IQR 0.035 to 0.062 s), femur angles (6.95°, IQR 4.35 to 11.65°), and leg-spreading angles (6.22°, IQR 5.08 to 10.04°), while RTMPose-L had the lowest median error for distance- and height-based characteristics (6.09 percentage points, IQR 3.84 to 9.10). Bland–Altman analysis showed generally small systematic biases, but acceptable agreement with the 3D reference system was achieved only for a small subset of the evaluated characteristics. The runtime benchmark, performed on the tested local hardware, further demonstrated substantial differences in computational efficiency, with mean processing speeds ranging from 6.50 ± 0.10 FPS for ViTPose-L to 66.87 ± 0.36 FPS for MoveNet. These findings indicate that, for 2D technique analysis of rapid sports movements from side-view recordings that meet minimum quality requirements, ViTPose-L and RTMPose-L are preferable when accuracy and keypoint completeness are prioritized, whereas MoveNet is preferable when runtime performance on comparable hardware is the main requirement. Therefore, the choice of pose-estimation model is strongly use-case dependent, and automated 2D analysis should be interpreted as approximate feedback for selected characteristics rather than precise individual-level biomechanical measurement.