Can joint range of motion be measured remotely using a camera?

What markerless pose estimation does

Markerless pose estimation is computer vision applied to the human body. A model trained on large quantities of annotated imagery examines each video frame and estimates where the body’s landmarks are, without any markers, suits or reflective dots.

The output is a set of coordinates per frame. Played across the video, those coordinates describe how the body moved.

Body keypoints and how joint angles are calculated

The landmarks the model estimates are usually called keypoints: shoulders, elbows, wrists, hips, knees and ankles among others.

1. 1The model estimates keypoint positions in each frame.
2. 2Three keypoints define a joint — for the knee, the hip, knee and ankle.
3. 3The angle between the two resulting segments is calculated geometrically.
4. 4Repeating this across every frame produces a joint-angle time series for the whole movement.
5. 5Range of motion is derived from that series, typically as the span between the smallest and largest angle reached.

Because the angle depends entirely on estimated keypoint positions, anything that degrades keypoint estimation degrades the angle.

2D pose estimation versus 3D motion capture

A single camera sees a flat projection of a three-dimensional movement. Marker-based 3D motion capture uses multiple calibrated cameras and physical markers to reconstruct positions in true three-dimensional space.

• Movement that stays within the plane facing the camera is captured comparatively well.
• Movement that travels towards or away from the camera is foreshortened, and out-of-plane rotation is difficult to resolve.
• 3D motion capture also delivers a level of precision, and access to full joint kinematics, that single-camera analysis does not.

This is why single-camera measurement is best matched to specific, clearly visible, largely single-plane movements rather than used as a general substitute for laboratory kinematics.

What the published MoveLab validation found

MoveLab’s 2024 peer-reviewed study compared two markerless 2D pose-estimation models, MediaPipe and HRNet, against Qualisys marker-based 3D motion capture in 22 healthy adults.

All tested joint-angle time-series comparisons had coefficient-of-variation values below 10%.

Two things stand out. Agreement was strongest for simple, clearly visible single-plane movements such as seated knee and elbow flexion/extension. And agreement depended on which model was used as well as which movement was performed — prone knee movement differed substantially between the two models.

Camera position

• The joint being measured must remain visible throughout the movement.
• The camera should face the plane the movement happens in; an oblique view systematically distorts the measured angle.
• Camera height and distance should be consistent, because changing them changes the projection.
• The whole relevant body segment needs to stay in frame for the entire repetition.
• For repeat assessment, replicating the setup matters more than achieving any particular ideal setup.

Lighting and occlusion

• Even, adequate lighting helps keypoint estimation; strong backlighting, such as a bright window behind the person, is a common cause of poor results.
• Clothing that obscures the outline of a limb makes the joint harder to locate.
• Furniture, walking aids or the person’s own limbs can hide a keypoint — a chair is a frequent culprit during sit-to-stand.
• Low contrast between the person and the background reduces reliability.
• When a keypoint is hidden, its position is inferred rather than observed, and the derived angle becomes less trustworthy.

Matching the movement plane

Some measurements are simply better suited to single-camera analysis than others. Flexion and extension of a large joint, performed facing the camera, is a favourable case. Rotation, small-amplitude movement, and multi-planar movement are much less favourable.

The practical consequence is that remote camera-based range of motion should be treated as a set of specific validated measurements, each with its own protocol, rather than as a general instrument that measures any joint in any direction.