What has been tested, and what has not
This page exists to be checked. Where something is still missing it says so, an evidence page that lists only the strong points is advertising with footnotes.
What is under the bonnet
- 1
Pose estimation
RTMPose with the Halpe26 point set: twenty-six points including heel and toe. Apache-2.0, so usable and distributable without licence restriction. It runs in the browser through onnxruntime-web, on WebGPU where available and WASM otherwise.
- 2
Person detection
YOLOX, also Apache-2.0. It does not run on every frame but periodically, with the box carried between detections. When someone walks through shot, the subject is tracked on continuity of centre rather than on size.
- 3
Filtering
Gaps up to five frames are bridged, then a zero-lag Butterworth filter at 10 Hz. Zero-lag because an ordinary filter shifts events in time, and time is exactly what is being measured.
- 4
Skeleton correction
Bone lengths should not change across a clip. They are estimated and then enforced, and the foot is fitted as a rigid body to the raw points. That removes fifteen to twenty-two per cent of the angular error around foot, ankle and shank.
How an event is established
Nearly every temporal measure hangs on initial contact and toe-off. Those are found by three independent methods: the height of the foot above an estimated floor line, the peak in vertical acceleration at landing, and the position of the foot relative to the pelvis.
An event only counts when at least two methods see it. The timing then comes from the acceleration method, which is the sharpest. That split, the majority decides whether, the sharpest decides when: turned out to be necessary: ordinary majority clustering threw away precisely the sharpest method whenever it sat a few frames away from the two blunt ones.
When the spread between the three methods grows too large, that appears on page one of the report. It is the signal that the recording itself is the problem.
What has been tested
Against a simulated runner with known ground truth. That tests the arithmetic end to end: an error in filtering, in finding cycles or in computing an angle surfaces here.
| What | Deviation from ground truth |
|---|---|
| Initial contact | 0.0 frames (sd 0.28) |
| Toe-off | +2.3 frames |
| Cadence | within 1 step per minute |
| Speed and step length | within 7% |
| Vertical oscillation | within 1% |
| Asymmetry on a symmetric runner | 0.0% |
What has not been tested yet
There is no study on real footage against a reference system. That is the largest outstanding item and it is not talked around here: a simulated runner has no clothing that moves, no shoe between the heel point and the heel, and no changing light.
The protocol for that validation is written and has three parts. One: concurrent measurement with force plates or an optoelectronic system, to establish systematic bias per measure. Two: repeatability: the same person twice in quick succession, and two raters on the same clip. Three: from those, the smallest change that reliably clears the noise.
Until that exists, the report makes no statement that would lean on it, and this page states what there is rather than what there ought to be.
Why this is not a medical device
The software measures and assesses the quality of that measurement. It does not link the measured values to a condition, predict anything, or propose treatment. That places it outside the MDR’s definition of medical device software, and Rule 11 does not arise.
The reasoning is written out against the five decision steps of MDCG 2019-11 and is public. It also states what would change with a module linking a patient’s values to literature and exercises: Class IIa. Such a module was built and deliberately removed again.
What is included is a reference library: what the literature reports about each measure, with no reference to any patient and no parameters. That is a textbook in digital form and does not change the classification.
Frequently asked
- Which reference values appear in the report?
- Ranges from the literature, as orientation beside the measurement. Not as cut-offs, and nothing is labelled abnormal: most of these measures depend heavily on speed and stature, and a range from a study on recreational runners is not a norm for an individual.
- Is a large image database involved?
- Not by us, and yes by the model we use: RTMPose is trained on public datasets with hundreds of thousands of annotated images. We do not retrain it; we use the published weights. No patient footage is used to train anything, and no footage is retained.
- Can the code be inspected?
- The underlying models are open (Apache-2.0) and the sources appear in the report and in the reference library. For a substantive review of the measurement method there is documentation describing step by step how each measure is arrived at; ask for it.
The recording protocol belongs with this
The largest source of error is not the software but the recording. One page, to tape up beside the treadmill.
Read the protocolRead further
- Running gait analysis for physiotherapists
Measure the running pattern from a slow-motion clip, put it in the notes, and at review see whether anything actually changed. From € 20 an analysis.
- Running gait analysis in sport and exercise medicine
What 2D video can and cannot measure, with the uncertainty stated. For anyone deciding whether a gait measurement will carry what gets built on it.
- For podiatrists: what video gait analysis does not measure
Honest answer: foot and ankle are the weakest part on 2D video. What this adds is what happens above the foot, and what changes when the orthosis goes in.
- Running gait analysis for coaches
Cadence, overstride and contact time as numbers, so that across a season you can show what changed instead of asserting it. From € 20 an analysis.
- For running stores: not the gait analysis you already do
This does not tell you which shoe fits. It does show what measurably changes in the running pattern between two shoes: which is a service you can sell on its own.
- Pricing for running gait analysis
Unlimited for € 129 a month, or from € 20 an analysis on a bundle. Bundles do not expire. First analysis free.
- Recording protocol for running gait analysis
Camera position, frame rate, clothing and duration. The recording determines the measurement more than the software does, and this is what has to be right.
- Compared: camera, IMU sensors, a motion lab or the naked eye
What camera-based analysis can and cannot do against IMU sensors, a gait lab and assessment by eye. Including the places where the others win.
- Try it: the first analysis is free
One complete analysis with the report, no watermark and no card details. You need a twenty-second slow-motion clip and nothing else.
- Bike fitting with the camera you already own
Measure your client on the trainer from an ordinary clip. Joint angles per crank revolution, adjustments in millimetres, a report for the file. From € 28 an analysis.
- Overpronation: what it is and what it does not predict
What pronation is, why sorting runners into neutral, over- and underpronating says less than long assumed, and what you can usefully look at instead.