Dynamic time warping for similarity analysis of skeleton keypoint sequences in dance scenarios
Abstract
Skeleton keypoint sequences provide a compact and interpretable representation for human motion analysis. Dynamic Time Warping (DTW), as a classical non-linear temporal alignment method, can find an optimal matching path between similar motion phases when two motion sequences differ in execution speed or contain local temporal shifts, thereby alleviating the temporal misalignment problem caused by direct frame-by-frame comparison. DTW has been widely applied to temporal sequence alignment tasks in domains such as human motion analysis and rehabilitation assessment, demonstrating its capability in handling variable-speed motion recordings. However, in dance scenarios, the application of DTW to skeleton keypoint sequence similarity analysis remains insufficiently explored, particularly regarding its quantitative contribution to rhythm-sensitive similarity analysis. Based on this motivation, this paper proposes Coord-DTW, a method that applies DTW alignment to 24-dimensional normalized joint-coordinate descriptors extracted from 2D skeleton keypoint sequences for dance motion similarity analysis. We further introduce a decomposed scoring scheme that separately quantifies postural accuracy through a Pose Score and rhythmic consistency through a Rhythm Score, enabling interpretable multi-dimensional feedback. Experiments on a dataset of ten performers executing the same choreography demonstrate that DTW alignment plays a critical role in skeleton-based dance motion similarity analysis, with its removal decreasing the average Final Score from 66.2 to 59.2. The proposed method produces well-differentiated scores ranging from 50.7 to 83.3 across 36 cross-performer pairs, and effectively ranks nine student performers against a professional reference with scores ranging from 47.2 to 72.8.