ALGORITHM

How the AI reads a turn

No black-box scoring. From pixels to physics, every step is explainable and reproducible.

🎥

① Input: any phone video

Portrait or landscape, 30–60 fps works best. Angles are most accurate with the camera directly up or down the fall line.

🦴

② Pose estimation: 17 keypoints

Frame-by-frame skeleton detection; in crowded frames the rider is locked via tracking plus area × confidence. 90%+ coverage in practice.

📐

③ Physical features, all explainable

Body-axis inclination, angulation, hip–shoulder separation, knee angle and CoM travel — computed directly from keypoint geometry as angles and ratios, not neural-net black-box outputs.

🌀

④ Turn segmentation at zero-crossings

Inclination crossing zero is an edge change; between two crossings is one turn. Amplitude and duration filters drop straight-running, producing a per-turn table.

👻

⑤ Phase-normalized comparison

Each turn is stretched to 0–100% phase and averaged, so your average turn overlays the master's — comparable across speed and turn size.

Seven metrics in plain words

Body-axis inclination
Angle of the hip-to-shoulder line from vertical. Bigger means more aggressive. Off-axis cameras shrink it, so compare like with like.
Angulation
Leg lean minus torso lean, sign-corrected by lean direction. Positive means the torso is more upright than the legs — you are flexing into the turn rather than falling in, the hallmark of quality carving.
Hip–shoulder separation
Angle between the shoulder line and the hip line, showing upper- and lower-body rotation timing. View-dependent, so same-camera comparisons only.
Knee angle
Hip–knee–ankle angle, where 180° is standing tall. The curve reveals how deep and how early you pressure and release.
CoM travel
Hip height over the ankles, normalized by median leg length — dimensionless, so riders of any height compare fairly.
Edge-change rate
Rate of inclination change through the transition (°/s). Higher means snappier edge changes; sluggish transitions show up instantly.
Rhythm & symmetry
Turn-duration variation (lower is steadier) plus the left/right peak-inclination ratio (100% is perfectly symmetric). Most riders have a weak side and the data names it.

Capability limits: video vs sensors

MeasurementVideoSensors (in development)
Turn shape, rhythm, symmetry✓✓
Edge-change timing & speed✓✓
Inclination & angulation△✓
Exact edge angle✗✓
Foot pressure & force split✗✓

✓ measured　△ image-plane projection (same-camera comparable)　✗ we say so when we can't

We stay honest about physics. Video delivers geometry and timing; the force layer belongs to the sensor kit (IMU plus pressure insoles, currently in collection with elite riders). Every report states how camera placement affects the values — coaches endorse it because we don't oversell.

Data demo Enter the library →