We benchmarked pose estimation accuracy across six models — MediaPipe, MeTRAbs, YOLO and MoveNet — against the VICON gold standard. Two clear tiers emerged; here’s which to use, and when.
Can markerless pose estimation replace marker-based motion capture in sports? Our validation study found a ~10-degree mean joint-angle difference — invisible to the eye, but too much for clinical precision. Here’s what that means.
A foundational overview: marker-based vs markerless pose estimation, the OpenPose/MediaPipe/MeTRAbs coordinate systems, ~10 degree accuracy and why errors happen, kinematics vs kinetics, capture frameworks, and a single-camera to OpenSim example.
A practical getting-started guide to two free markerless motion-capture tools: OpenPose (install the demo, fix the models, run it) and OpenCap (2 iOS cameras, calibration, and 3D capture).
The math behind atan2d(det, dot): the determinant is the cross-product magnitude (|a||b|sin), the dot product is |a||b|cos, so their ratio is tan – and why the two-argument atan2 preserves sign and handles angles beyond 90 degrees.
A step-by-step MATLAB walkthrough: import OpenPose JSON, compute joint angles with atan2d, filter with Butterworth (cutoff from the power spectrum), find phases and landing, and export a results table.
Breaking a football kick down phase by phase with a slow-motion phone video and OpenPose – backswing, leg cocking, acceleration, and the supporting leg’s shock absorption and energy transfer, read from the joint-angle and angular-velocity curves.
Part 5 (finale): kinetic/dynamic simulation with OpenSim Moco and automatic differentiation – muscle activations and ground reaction forces from the reconstructed motion.
Part 4: driving an OpenSim model with the reconstructed motion – convert to .trc, LSTM marker augmentation (OpenCap), scale the model, and run inverse kinematics.
Part 3: triangulating 2D keypoints into 3D with the DLT, weighting by OpenPose confidence, person tracking, residual-analysis filtering, and reprojection-error optimization.