MediaPipe-based human topology tree and damped least-squares inverse kinematics framework for interpretable yoga pose assessment

Abstract Static yoga posture assessment requires more than pose recognition or two-dimensional joint-angle matching because small landmark errors can disrupt estimates of support-chain continuity and holding stability.We present VectorBody2.0,a monocular visual proxy framework that organizes MediaPipe landmarks in a pelvis-rooted topology and uses topology-constrained damped least-squares inverse kinematics to stabilize static sequences. Visual proxies are image-derived indicators of observable alignment, support continuity, temporal stability, and compensation; they are not estimates of internal biomechanical loads. The analysis-ready dataset comprised 1,500 clips from 30 participants performing eight postures, with the mean score of three expert teachers as the reference. VectorBody 2.0 achieved a mean absolute error of 4.982 (participant-clustered 95% confidence interval, 4.827–5.141), a root-mean-square error of 6.225, Pearson r of 0.918, and an intraclass correlation coefficient of 0.916. The basic two-dimensional stage yielded corresponding values of 8.536, 10.735, 0.787, and 0.774. Sequential component-addition analysis showed lower error after topology modelling, kinematic stabilization, and visual proxy fusion. These results support interpretable posture feedback under controlled single-person recording, while external populations, devices, and environments remain untested.