An Explainable Dance Pose Evaluation System Using Velocity-Adaptive Filtering and Dynamic Alignment

Practicing dance from pre-recorded online tutorials allows students to learn motor skills independently, yet learners frequently struggle without an instructor to identify technical errors. This paper addresses the research problem of objective, monocular video dance evaluation by developing an end-to-end computer vision framework that compares a student’s practice video with an instructor’s reference recording. Our methodology extracts 33 body landmarks, normalizes coordinates by torso length to account for body differences and camera distances, and applies a velocity-adaptive Kalman filter that eliminates lag during fast steps while preserving stability during held poses. A Dynamic Time Warping module with a ±0.25 s tolerance window evaluates body posture without penalizing expressive musical timing. Continuous routines are segmented into discrete poses using feature-divergence tracking, with a rule-based engine outputting plain-language postural corrections. The system was evaluated across 15 full-body benchmark routines comprising 10,550 video frames from the AIST Dance Video Database and instructional exercises in both frontal and 90◦ side views. The results demonstrate that the system tracks fast direction changes accurately, detects technique breakdowns reliably, and matches professional instructor ratings closely (r = 0.88, Mean Absolute Error = 4.8 points). In conclusion, this automated framework offers accessible, actionable feedback for home physical education using standard consumer cameras.