What is markerless motion capture and how can it be used in healthcare?

What makes it ‘markerless’

Traditional motion capture attaches reflective markers to anatomical landmarks and tracks them with an array of calibrated cameras. Markerless motion capture removes the markers: a computer-vision model estimates where the body’s landmarks are directly from ordinary video.

In exchange for losing marker-level precision, it gains something valuable in healthcare — a measurement that needs no preparation of the person, no specialist facility, and in many cases no more than one camera.

Body keypoints and pose-estimation algorithms

Pose-estimation models are trained on large annotated image datasets to predict the location of body keypoints: shoulders, elbows, wrists, hips, knees, ankles and others. For each video frame the model returns estimated coordinates, usually with an associated confidence.

• Different models make different trade-offs between speed and accuracy.
• Some are designed for real-time use on modest hardware, others for maximum accuracy offline.
• Model choice affects results — the same movement analysed with two models can produce different agreement with a reference standard.
• Every downstream measure inherits the quality of the underlying keypoint estimate.

From keypoints to clinical measures

• Joint angles — calculated from the geometry of three keypoints, producing an angle for every frame and therefore a time series across the movement.
• Range of motion — derived from that time series, typically the span between the smallest and largest angle achieved.
• Repetitions — counted by detecting the repeating cycles in the movement, as in sit-to-stand or a prescribed exercise.
• Movement consistency — how similar each repetition is to the others, which can reveal fatigue, hesitation or compensation across a set.
• Timing — how long a movement or a phase of it takes.

These are measures of observable movement. They describe what the body did, not why it did it, and interpretation remains a clinical task.

A worked example: MoveLab’s 2024 validation

MoveLab’s 2024 peer-reviewed study compared MediaPipe and HRNet 2D pose estimation against Qualisys marker-based 3D motion capture in 22 healthy adults, deriving joint angles from body keypoint coordinates.

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

The pattern is instructive. Clearly visible, largely single-plane movements — seated knee flexion, elbow flexion — agreed closely. Sit-to-stand, where a chair and the person’s own body obscure the joint and the movement is less cleanly planar, produced markedly weaker agreement on one side for both models. Performance was movement-specific and model-specific rather than uniform.

Camera perspective

• A single camera records a two-dimensional projection, so movement towards or away from it is foreshortened.
• An oblique view introduces systematic error in the measured angle.
• Camera height, distance and angle should be replicated across repeat assessments.
• Stability matters — a hand-held camera adds movement that has nothing to do with the person.

Occlusion

• Furniture, walking aids and equipment hide keypoints; chairs are a frequent problem during sit-to-stand.
• The person’s own limbs and trunk can obscure a joint during rotation.
• Loose clothing blurs the outline the model relies on.
• When a keypoint is hidden, its position is inferred rather than observed, and any angle derived from it is less trustworthy.

Lighting

• Even, adequate light gives the model the clearest signal to work with.
• Strong backlighting, such as a bright window behind the person, is one of the most common causes of poor capture at home.
• Low contrast between clothing and background reduces reliability.
• Very low light introduces noise and reduces effective frame rate.

Movement plane

Single-camera markerless analysis is best matched to movement that happens predominantly in the plane facing the camera. Flexion and extension of a large joint, performed square to the camera, is a favourable case.

Defined, predominantly single-plane movements with a clear camera view provide the strongest basis for a repeatable measurement protocol.