Running form analysis from a side-on treadmill clip
Stand the phone beside the treadmill, film about 20 seconds, and get a scorecard of thirteen gait metrics plus a plan built around the worst bands — not a vibe check.
A headline score that is the mean of six categories, not the mean of every row, so adding a metric later does not rewrite history. Each row is a band (good / fair / poor) with a number that stays inside that band. Fair and poor rows come with a why and a drill.
How to film
Phone side-on at about hip height, square to the treadmill, whole body in frame from head to feet, near leg unobstructed. Easy conversational pace for about 20 seconds; trim the stepping-on. Live capture and a library upload produce the same landmark type, so either path is valid. Centimetre rows (overstride, bounce) need a known height and a fairly constant camera distance — uploads still score angles and timing.
Running-form ranges (the research behind the colours)
These are the ranges the scorecard colours against (green = in range, amber = borderline, red = out). All measured from a side-on clip at self-selected easy running pace. Population ranges and a starting point, not a lab gait analysis. Cadence and overstride are the two most consistently supported form levers in the literature: a higher step rate shortens the stride, moves footstrike closer under the centre of mass, and cuts braking and peak joint loads. Overstriding — the foot landing well ahead of the knee and hips — is the single most common amateur fault.
Cadence
Green. Reported, not scored. Literature: 170–190 spm at easy pace (higher is fine)
Red. <162 at easy pace → raise turnover (shorter, quicker steps). We still do not deduct points — cadence scales with speed
Overstride (ankle ahead of hips at contact)
Green. 0–5 cm
Red. >10 cm = reaching/braking, land under your hips
Trunk lean from vertical
Green. 4–12° forward
Red. <2° leaning back; >15° folding at the waist
Knee bend at landing
Green. 10–25°
Red. Very stiff (near-straight) or collapsing. Soft knee absorbs the landing
Bounce (vertical oscillation)
Green. Under 8 cm
Red. >11 cm = excessive bounce, wasted energy
What we grade, and what we deliberately don't
Cadence is reported, not scored. The literature clusters around 170–190 spm at easy pace and higher is fine, but cadence moves with speed and height — a trusted 193 must not lose points. Rhythm (step-interval spread) is what that category colours. Overstride is the money metric we do colour: landing position relative to the body, in centimetres from your height. Trunk lean is graded softly — a slight lean from the ankles is efficient; bolt-upright or folding at the waist is not. Foot-strike TYPE (heel vs midfoot vs forefoot) is not graded, on purpose. The evidence does not support “heel strike = bad.” What matters is WHERE the foot lands relative to your body, not which part touches first. Grading strike type would be pretending to a certainty the research does not have. Heel landmarks are also blank on Apple Vision, so the camera could not name it honestly even if we wanted to.
Measurement tolerance
Monocular pose on a phone clip resolves joint angles to roughly ±2–3°, so a reading within a couple of degrees of the green window is treated as fair rather than red. Honest precision, not goalpost-moving: it applies to every angle metric and is far too small to change a genuine fault. An 8 cm overstride is fine; a 24 cm one is not.
Cadence math
The near-side foot completes one stride per up/down cycle; cadence counts BOTH feet, so steps-per-minute = (strides/sec) × 2 × 60. The analyser finds gait-cycle peaks (the ankle’s lowest point each stride), takes the median stride interval, and converts. Sampling is 10 frames per second with sub-sample peak refinement, so a true 180 spm is not snapped to 150 or 200. If tracking is too noisy, cadence prints as “—” rather than a plausible-looking number.
These are easy-pace ranges
Cadence and lean both rise with speed. If the clip is a hard interval or a sprint, a higher cadence and a bit more forward lean are normal. The scorecard assumes an easy, steady effort — film at conversational pace for the cleanest read.
Sources
The gates come from mainstream running biomechanics, not from a lab session of yours. The papers below are why cadence and overstride are the levers, why strike type is left off the card, and why 180 is a useful target rather than a law.
• Cadence (steps per minute) and step-interval rhythm
• Overstride and knee bend at landing
• Trunk lean and head carriage
• Bounce, arm swing sweep, elbow angle
• Knee drive and hip extension
• Left/right timing asymmetry and landing symmetry
What this camera cannot score
• Hip drop, knee valgus, arm crossover, or any frontal-plane alignment — those need a front camera, which is a different capture
• Foot-strike type (heel / mid / forefoot) — the evidence does not support “heel strike = bad”, and heel landmarks are blank on Apple Vision. We score landing position (overstride) instead
• Ground-contact time, stride length, or belt speed — the treadmill belt is invisible
• Impact force or injury risk — not derivable from pose
Questions
Does the video upload to a server?
No. Pose is extracted on the device (or in the browser on the free check). Only landmarks, metrics, and scores are saved to your account.
Why bands instead of one-decimal cadence?
Cadence and trunk lean are reliable from one 2D camera. Overstride and asymmetry are noisier. Printing false precision is how athletes stop trusting the card.
Will you tell me to hit 180 spm?
No. Literature clusters around 170–190 at easy pace and 180 is a useful cue, not a law — it scales with speed and height. Cadence is reported; rhythm is what that category scores. A trusted 193 does not lose points.
Why don't you grade heel strike vs forefoot?
Because the evidence does not support “heel strike = bad.” Braking is about where the foot lands relative to your body (overstride), not which part touches first. Heel landmarks are also blank on Apple Vision, so we could not name strike type honestly even if the research asked us to.
Is this a lab gait analysis?
No. These are population ranges from a side-on phone clip at easy pace. They are a starting point you can re-film against, not a force-plate session.
Keep your results
The free check scores one clip and forgets it. An account saves every analysis, writes the 4-week plan around your worst band, and shows whether the number actually moved when you re-film. Five analyses a month, free.