Robust 3D Human Motion Reconstruction from Multi-View Mobile Videos for Quantitative Ski Technique Analysis

Quantitative evaluation of skiing technique in real-world settings is inherently challenging due to complex whole-body coordination, rapid posture transitions, and frequent viewpoint changes during motion. Conventional assessment still relies heavily on expert observation, which is subjective and difficult to standardize. In addition, many existing motion-analysis pipelines assume fixed and calibrated camera setups, which limits practical deployment in realistic skiing environments. In this study, we present a practical multi-view framework for reconstructing stable 3D human poses from synchronized mobile videos without requiring fixed camera setups or prior calibration. The framework integrates skier tracking, single-view 3D pose estimation, canonical alignment, confidence-guided multi-view fusion, and temporal smoothing. Experiments on real skiing videos show that multi-view fusion improves the spatial consistency and temporal stability of reconstructed 3D poses and supports interpretable analysis of posture stability, joint kinematics, and inter-limb coordination, including observable differences between skiers with different experience levels. On simulated skiing data with ground truth, the proposed method achieves a fused 3D MPJPE of 0.5130, improving over single-view baselines of 0.7950 and 0.8498. The proposed framework provides a scalable and camera-agnostic solution for quantitative sports motion analysis and can be extended to other complex whole-body activities beyond skiing.