Interpretable Dance Assessment with Geometry-Preserving Pose Normalization and Constrained Temporal Alignment
Abstract
Low-cost monocular coaching can make community dance practice more accessible, but reference-based systems often confuse a learner’s slower tempo with incorrect posture and provide unstable, opaque scores. This paper presents an interpretable dance coaching prototype combining 2D keypoints, geometry-preserving isotropic pose normalization, confidence-weighted joint similarity, and band-constrained dynamic time warping (DTW). A collapsed path produces exactly one joint-level feedback item per learner frame, and an optional subsequence mode handles partial recordings. Reproducible stress tests use two real 480-frame pose trajectories as motion bases with known synthetic ground truth. Across five tempo conditions, DTW obtains 0–1.60-frame mean absolute alignment error, compared with 0–34.14 for fixed-offset matching. In a matched 128-case normalization sweep, isotropic localization reaches $\mathbf{F 1}=\mathbf{0 . 9 9 8}$ and 1.000 at deviations of 0.15 and 0.24 body scales, versus 0.785 and 0.811 for bounding-box normalization. The CPU-only post-keypoint stage processes a 19.2-s pair in about 0.52 s using one software thread. These results validate prototype mechanics, not learning efficacy or clinical safety in older adults.