1. Landmark detection
MediaPipe Pose Landmarker (lite) runs locally via WebAssembly/WebGL and returns 33 normalized landmarks with visibility scores per frame. It is Google's open baseline model, not a OrthoTraq invention.
Technology
Deterministic computer vision, on your device.
No frames to the cloud. No AI-invented numbers. Every measurement shown in the scanner is derived from detected landmarks with documented math.
2. Quality validation
Each frame is checked for landmark visibility (≥ 0.6), in-frame position, and whether the subject's orientation matches the required side or front view. Failing frames produce no angle.
3. Joint-angle geometry
For landmarks A, B, C the angle at B is atan2(|BA × BC|, BA · BC), clamped to [0°, 180°]. Coordinates are aspect-corrected so the image plane uses square pixels.
4. Temporal filtering
A median-of-3 pre-filter rejects single-frame spikes, followed by an exponential moving average to reduce jitter.
5. Repetition detection
A finite-state machine with two thresholds (hysteresis), minimum dwell frames and minimum/maximum rep durations prevents double counting and rejects noise.
6. Session metrics
Minimum, maximum and excursion angles, repetition timing, a jerk-based smoothness estimate and an excursion-consistency score, all computed from observed landmarks.
Limitations. 2D image-plane angles are not equivalent to true anatomical 3D joint angles. Results depend on camera placement, lighting, clothing and occlusion. Estimates are experimental and not clinically validated.