Abstract
With advances in computing power and deep learning, image-based pose estimation has become a viable tool for quantitative motion analysis. Compared to sensor-based systems, vision-based approaches are cost-effective, portable, and easy to deploy. However, when applied to golf swings, conventional similarity measures often fail to match expert perception, as they rely on static, frame-wise posture comparisons and require strict temporal alignment. We propose a Dynamic Motion Similarity Measurement (DMSM) framework that segments a swing into seven canonical phases-address, takeaway, half, top, impact, release, and finish-and evaluates the dynamic trajectories of joint keypoints within each phase. Unlike traditional DTW- or frame-based methods, our approach integrates continuous motion trajectories and normalizes joint coordinates to account for player body scale differences. Motion data are interpolated to improve temporal resolution, and numerical integration quantifies path differences, capturing motion flow rather than isolated postures. Quantitative experiments on side-view swing datasets show that DMSM yields stronger discrimination between same- and different-player pairs (phase-averaged separation: 0.092 vs. 0.090 for the DTW + cosine baseline) and achieves a clear biomechanical distinction in spine-angle trajectories (Δ = 38.68). Statistical analysis (paired t-test) confirmed that the improvement was significant (p < 0.05), and coach evaluations supported perceptual alignment. Although DMSM introduces a minor computational overhead (≈169 ms), it delivers more reliable phase-wise feedback and biomechanically interpretable motion analysis. This framework offers a practical foundation for AI-based golf swing analysis and real-time feedback systems in sports training, demonstrating improved perceptual consistency, biomechanical interpretability, and computational feasibility.