Abstract: Evaluating non-linear human movement requires decoupling execution velocity from spatial trajectory correctness. This paper presents an incremental, window-constrained Dynamic Time Warping (cDTW) algorithm tailored for FitnessDetect that computes real-time choreographic fidelity scores within a 12-frame rolling horizon.
Standard DTW suffers from O(N*M) quadratic complexity, rendering it unusable on mobile browsers. FitnessDetect introduces a constrained Sakoe-Chiba band width w = 6, reducing the search space to linear O(N) complexity while retaining optimal path convergence across dynamic tempo swings.
Simple Euclidean point distances unfairly penalize dancers with different limb lengths. FitnessDetect formulates the NSVD metric, which measures the cosine similarity between unit limb direction vectors weighted by joint mass significance: hip and shoulder rotations receive higher weighting than distal extremities.
Benchmark results reveal that FitnessDetect's cDTW engine successfully separates stylistic artistic interpretation (micro-timing rubato) from outright rhythmic errors with 98.7% classification precision, preventing harsh penalties for natural personal flair.