Running form analysis from a phone video

A pose model places 33 body landmarks on every frame. Gait events (initial contact, midstance, toe-off) are found from the motion of the feet, each stride is segmented, and the metrics are averaged over strides so one noisy step doesn't set your number. Angles are reported as whole degrees, because single-camera pose is accurate to a few degrees and a decimal place would be theatre.

Most running apps compare everyone against one band: 180 steps per minute, so many millimetres of vertical oscillation. The published data says that is wrong. In the reference cohort of 860 recreational runners, step rate drops by about 44 steps per minute for every extra metre of height, so a 30 cm height difference is worth 13 steps per minute on its own, and ground contact time shortens by about 17 ms for every extra km/h of speed. Both sexes in that cohort averaged 164 steps per minute, not 180.

Mocta-X uses those regression equations to predict the six stride metrics for you, from your sex, age, height, weight and running speed. On a treadmill you enter the pace; overground, enter it too, or the affected target bands widen rather than silently falling back to a population average. The joint-angle metrics use fixed bands with a cited centre where one has been published, and are tagged as an estimate where we had to interpolate. Each metric carries its reliability label in the interface.

• Measured: the target is stated for this convention in the cited source.
• Modelled: predicted for you from the published equations.
• Estimate: the convention is cited but the band is ours. Guidance, not a standard.

Some running patterns show up more often in injured runners. A study of 72 injured and 36 healthy runners found injured runners had more contralateral pelvic drop and forward trunk lean at midstance, and a more extended knee and dorsiflexed ankle at contact; pelvic drop was the strongest single discriminator. Mocta-X builds an index from those patterns, weighted by the published effect sizes.

It is an index of load patterns, not a probability of injury, and it is built to say so. Weights are renormalised over the contributors actually measured, so a side-only clip is scored on what it could see rather than punished for the rear view it never had; because a side clip can't see the two strongest contributors, its confidence is capped and the card names what is missing. Pelvic drop is discounted separately for how poorly it can be measured from video (trained raters watching the same clip barely agree), so the least trustworthy number never dominates. The formula is shown in the interface.

The form score compares this run to your own personalised targets under the tolerance of the profile you chose. It is the coaching number. The canonical score grades the same measurements against one fixed rubric, using only the side-view metrics, so it is the comparable number: history, progress charts and any leaderboard read that one, and uploading a rear clip can never move it.

An AI coach then reads the graded session, with its joint scores worst first, phase analysis and recommendations, and writes a read you can question. It works from the measurements, never from the video. How the coach works describes it for rowing; running uses the same coach with the running session tools.

1. Side-on, at hip height, square to the line you run along. Joint angles are measured in the image plane, so past about 20–25° off side-on they distort.
2. Whole body in frame for several strides. On a treadmill, step back until head and feet stay in shot. Overground, run past a fixed phone, not a panning one.
3. Light behind the camera. A window behind you makes a silhouette, and silhouettes track badly.
4. Normal video mode at 30 fps or better. Slow motion isn't needed. A 20–30 second clip is enough.
5. For the rear view, film from directly behind on a treadmill or a straight path, level with the hips.

A capture-quality check runs before scoring. If the clip is too far away, too dark or too far off-axis, the affected metrics are withheld and the reason is shown, rather than a wrong number being reported.

Running analysis runs in Mocta-X Dashboard in any browser and in Mocta-X for iPhone. The free plan includes two video analyses a day with clips up to 30 seconds and covers running, rowing and golf. Pro at $9.99 a month raises the limits and adds the streaming coach and PDF reports; see pricing. The side and rear views also work at a walking pace, so the same tools serve as a phone-based gait check for rehabilitation and home assessment.