Markerless Motion Capture

Why markerless, and what it costs you

Marker-based optical motion capture remains the reference standard for joint kinematics, and nothing here claims to replace it for laboratory biomechanics. What it cannot do is measure a full squad on a Tuesday. Attaching a marker set takes twenty minutes per athlete, requires a controlled volume, and changes the task — athletes do not sprint the way they normally sprint while covered in reflective markers inside a capture volume.

Markerless capture trades some accuracy for the ability to measure real movement in real settings. Our system uses four to eight hardware-synchronised cameras at 120 to 240 frames per second, calibrated to the capture volume with a standard wand routine. A convolutional pose model detects 2D keypoints per camera; triangulation across views produces a 3D skeleton of 37 keypoints; an inverse-kinematics stage fits a constrained body model to those points so that joint angles respect anatomical limits rather than floating freely.

Where the error actually is

Being specific about accuracy is more useful than a single headline figure, because error is not uniform across joints, planes or tasks. Sagittal-plane angles at large joints — knee flexion during running, hip flexion during a squat — are the strongest case. Frontal and transverse plane angles at the same joints are noticeably weaker, and small-joint or distal-segment rotation is the weakest of all. Occlusion makes everything worse, which is why camera count and placement matter more than model sophistication.

We do not report values we cannot defend. Where a keypoint's triangulation confidence falls below threshold for a frame, the affected angle is marked as interpolated and excluded from summary statistics rather than quietly filled in.

Typical set-ups

A gym installation with four cameras around a lifting platform covers squat, deadlift and jump-landing analysis. A six-camera arrangement along a 40 m indoor straight covers acceleration and maximal-velocity sprint mechanics. Pitch-side portable rigs exist and work, but honestly: outdoor variable lighting, longer camera-to-subject distances and multiple overlapping athletes all degrade the result. We will tell you which of the three your question needs.

What the output is good for

The most reliable use is intra-athlete comparison over time. An athlete's own sagittal knee angle at a matched running velocity, measured in the same rig with the same calibration, is a genuinely sensitive signal of change — through a rehabilitation progression, across a pre-season, or before and after a technical intervention. Comparing two different athletes, or the same athlete across two different rigs, introduces error terms that often exceed the difference you are trying to see.

The second good use is screening as a prompt. A frontal-plane pattern flagged during a jump-landing session is not a diagnosis; it is a reason for a clinician to look. Used that way, moderate confidence is entirely sufficient.

Processing and integration

Capture is recorded and processed offline at roughly 1.4 times real time on our cloud pipeline, so a twelve-minute session is available in about seventeen minutes. Results land in the same athlete timeline as telemetry and load data, so a flagged mechanical change sits next to that week's exposure rather than in a separate report nobody opens.

Export is C3D for biomechanics tooling, CSV for joint-angle time series, and JSON through the API. Video is retained according to the retention period you set, with per-athlete consent recorded, because footage of an identifiable person in training kit is personal data under both GDPR and most club employment agreements.