Feature-dependent agreement between markerless and marker-based motion capture during continuous Taijiquan movement

Markerless motion capture is increasingly used outside specialist laboratories, but its validity may differ across kinematic features. We analysed paired Kinect V2 and Qualisys recordings from 12 participants performing 13 standardised Taijiquan gestures, comprising 1815 strictly paired segments and 156 participant-by-gesture units. Analyses used the official source files, timestamp-aware preprocessing, participant-cluster uncertainty estimates, absolute-agreement and consistency metrics, native-unit errors, and sensitivity analyses for anatomical mapping, filtering, bandwidth, synchronisation, scale normalisation, and data quality. Agreement was strongly feature-dependent. Pelvis features showed high agreement; knee and elbow angles preserved high relative consistency despite systematic offsets; hip agreement was mapping-sensitive; sagittal trunk orientation showed negligible cross-system association; and derivative-based velocity and smoothness measures remained distorted across sensitivity conditions. Participant-isolated feature-specific calibration corrected stable knee and elbow bias but did not restore features with weak underlying association. Exploratory participant-level quality ranking showed relatively high observed correlations for some models, but neither Ridge reference condition met the Holm-adjusted permutation threshold. These findings support feature-specific validation that distinguishes systematic bias, anatomical-mapping sensitivity, and residual measurement error.