A Marker-less Human Motion Analysis System
Motion-Based Biomarkers
in Knee Disorders.
Turning ordinary video into objective measurement — a full pipeline from a single RGB camera to a clinical report, and a study design clever enough to prove the numbers respond to treatment rather than to noise.
20
Participants
9
Significant Biomarkers
1
RGB Camera
The Problem
Too many low-risk patients,
too little objective data.
The UK performs over 90,000 total knee replacements a year, at an average cost above £7,000 per procedure, with the NHS spending more than £600 million annually on knee-related care. Waiting lists and staff shortages fall hardest on the low-risk patients whose movement could be monitored rather than escalated — if there were a cheap, repeatable way to measure it.
Marker-based motion capture answers the measurement question but needs trained experts, slow data collection, and a dedicated laboratory. The gap this study closes is not capturing movement — it is knowing which movement features are clinically meaningful.
Designed for the clinic, not the laboratory
Uncontrolled lighting
Clinic rooms have natural light whose direction shifts through the day.
Ordinary clothing
Patient attire varies and can occlude the body parts being tracked.
Mixed ability
Every patient has different functional capacity, so the test actions must be ones everyone can perform.
A standardised protocol
A clinician-designed protocol holds these variables steady so the data stays comparable.
The Pipeline
From a video file to a clinical report.
Participants performed at least three repeats each of a sit-to-stand and a squat — simple actions chosen by clinicians because everyone can attempt them, and because their diagnostic value is already established in the literature. Everything after that is automated.
Record
A single 1080p RGB video at 30 frames per second — camera at 1.2 m height, 3 m from the participant, who faces the lens to minimise occluded joints.
Reconstruct
Video is fed to a VIBE-based mesh reconstruction pipeline that recovers an SMPL-X body model — 10,475 mesh vertices around 54 joints — encoding both spatial and temporal cues.
Derive features
Joint angles are computed frame by frame in the sagittal and coronal planes, then reduced to two clinically meaningful measures: smoothness and cumulative rotational acceleration.
Rank with PCA
Principal component analysis ranks feature importance per participant and per action; features appearing most often at the top become candidate biomarkers.
Report
The surviving biomarkers are assembled into a clinical report written for both clinician and patient — explainable features, not a black-box score.
Explainable by design
Features are calculated manually from joint kinematics rather than learned end to end, so every number in the report traces back to a joint, a plane, and a formula that a clinician can interrogate.
What Gets Measured
Two measures that describe how a movement is made.
Range of motion says how far a joint travelled. These two say how it got there — and they turned out to be the measures that moved when pain was removed.
Smoothness
The standard deviation of the gradient of a joint-angle curve — how much the slope varies from point to point across the movement. A value closer to 1 identifies a smoother curve. Hesitation, guarding and compensation all show up here.
Computed for the mean, maximum and minimum curves across a participant’s repeats.
Cumulative acceleration
The cumulative absolute rotational acceleration at a joint across an action — a single number standing in for the explosiveness of the movement, or the overall abruptness of changes in angular velocity.
Derived from the change in joint angle over time, without time-normalising — a longer action may itself be the pathology.
How the Biomarkers Were Validated
Remove the pain.
See what changes.
Twenty participants — men and women over 55 with diagnosed knee pain — were each filmed before and after a local anaesthetic injection into the painful knee. The anaesthetic strips out the psychological alteration of movement caused by pain, producing a clean before-and-after within the same person, 15–30 minutes apart, in the same clothing, lighting and camera position.
Two-tailed paired t-tests at a 0.05 significance level compared pre- and post-injection values for the features PCA had ranked most important, with Bland-Altman plots assessing agreement and systematic bias. Nine biomarkers reached significance.
Sit-to-stand
7 significant
| Biomarker | t | p |
|---|---|---|
| Right elbow flexion (max) — smoothness | 3.592 | 0.002 |
| Left arm abduction (mean) — smoothness | 3.586 | 0.002 |
| Left knee flexion (max) — smoothness | 2.976 | 0.008 |
| Left elbow flexion (max) — smoothness | 2.604 | 0.017 |
| Right elbow flexion (max) — cumulative acceleration | 2.451 | 0.024 |
| Right knee flexion (max) — smoothness | 2.401 | 0.027 |
| Left elbow flexion (max) — cumulative acceleration | 2.364 | 0.029 |
Squat
2 significant
| Biomarker | t | p |
|---|---|---|
| Left knee flexion (max) — smoothness | 2.528 | 0.021 |
| Right knee flexion (max) — smoothness | 2.324 | 0.031 |
Values from Tables 1 and 2 of the paper; all listed biomarkers reached p < 0.05. Each showed a median increase from pre- to post-injection, with some individual outliers.
The unexpected result
The sit-to-stand produced more significant biomarkers than the squat — and several of them were in the arms. Elbow flexion and arm abduction changed measurably once knee pain was removed, because participants were permitted to push up with their arms. The system detected the compensation strategy, not just the painful joint.
Measurement Error, In Context
No method is error-free.
The paper places its own error alongside the established alternatives rather than claiming precision it does not have.
Marker-less joint centres
20–40 mm
Mean absolute error in joint centre location reported for marker-less capture.
Marker-based registration
≈7°
Peak joint angle variability produced by 50 mm of marker registration uncertainty.
Inertial measurement units
up to 11.4°
Reported measurement error for IMU-based joint angle estimation.
Why the design holds
Because every variable was held constant between the two measurements — same day, same clothing, same lighting, same camera, 15 to 30 minutes apart — measurement error is non-differential: it applies equally before and after. The comparison is one person against themselves, which is exactly where that assumption is safest.
What the Authors Flag
Stated limitations, not footnotes.
A small, knee-focused case study
Twenty participants in a single clinical context. The methods generalise in principle to other joints and to neurological movement disorders, but that breadth has not yet been demonstrated.
Biomarkers are not universally applicable
Bland-Altman plots show most participants clustered tightly, but outliers suggest the selected biomarkers, while effective for many, do not fit everyone. The squat showed greater variability than the sit-to-stand.
An uncontrolled environment brings occlusion and jitter
Without a controlled capture space, occlusion and frame-to-frame jitter remain open problems the authors name as needing future attention.
Right-side prevalence is unexplained
Whether the prevalence of right-sided biomechanics reflects a right-dominant sample or bilateral pain is unresolved — a limitation of establishing a truly representative clinical sample.
The study builds directly on the 2022 marker-less capture comparison and feeds forward into the sit-to-stand protocol validation and the biomechanics work now delivered through MAI Motion.
A standard camera. An objective answer.
The techniques offer a standardised route to musculoskeletal analysis achievable with any standard camera — even a mobile phone — opening the door to remote monitoring and earlier identification of pre-disease stages.
Study Identity
- Authors:
- Kai Armstrong · Lei Zhang · Yan Wen · Alexander P. Willmott · Paul Lee · Xujiong Ye
- Institutions:
- Laboratory of Vision Engineering and School of Sport & Exercise Science, University of Lincoln · MSK Doctors, Sleaford
- Journal:
- Frontiers in Digital Health 6:1324511, published 23 January 2024
- Ethics:
- University of Lincoln Ethics, Governance & Regulatory Compliance Committee
- Licence:
- Creative Commons Attribution (CC BY)
Citation
Armstrong K, Zhang L, Wen Y, Willmott AP, Lee P, Ye X. A marker-less human motion analysis system for motion-based biomarker identification and quantification in knee disorders. Front Digit Health. 2024;6:1324511. doi:10.3389/fdgth.2024.1324511