Video-Based Markerless Motion Capture for Clinical and Rehabilitation Biomechanics: A PRISMA-ScR Scoping Review of Validated Architectures, Clinical Readiness, and Emerging Methods
Florian Delaplace, Elodie Piche, Frédéric Chorin, Raphael Zory
cs.CV · physics.med-ph
Quick verdict
Scoping review of 117 studies: markerless capture at ~5-6° sagittal error is not yet clinically interchangeable; pooled metrics and small n limit precision.
Summary
This PRISMA-ScR scoping review, with a registered OSF protocol, searched PubMed, Scopus and IEEE Xplore (January 2015 to February 2026) and included 117 studies that validated video-based markerless motion capture against an external biomechanical reference. A deliberately non-exhaustive companion corpus of 572 computer-vision records (MICRO) was used qualitatively to map emerging building blocks. Validated studies cluster on healthy adults walking in laboratories (median n=17); pipelines fall into five architectural families, and 73 of 117 report raw geometric joint angles without biomechanical refinement. Median sagittal lower-limb errors cluster at 5-6°, with roughly half of estimates exceeding the 5° threshold McGinley et al. associate with misleading clinical interpretation; out-of-plane kinematics (knee: 65 sagittal vs 5 frontal vs 3 transverse studies), kinetics (9 studies), and pathological or older populations (~19% pathological) are rarely validated. Mesh recovery, differentiable inverse kinematics, and video kinetics are active in computer vision but nearly absent from validated work. The authors distil five hypothesis-generating design guidelines for clinically deployable pipelines. Limitations include pooling of RMSE and MAE, out-of-plane cells based on 1-5 studies, exclusion of commercial systems such as Theia3D, and no risk-of-bias appraisal.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
5 strengths · 5 weaknesses · 5 author questions
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- The 117 validated studies resolve into five architectural families across monocular and multi-camera modalities, and 73 of 117 report raw geometric joint angles with no biomechanical refinement.
- Sagittal lower-limb agreement clusters at roughly 5-6° median absolute error, with about half of estimates (57% pooling all metrics) exceeding McGinley's 5° 'likely to mislead' threshold.
- Out-of-plane kinematics (knee: 65 sagittal vs 5 frontal vs 3 transverse studies), kinetics (9 force-plate validations), and pathological or older populations (~19% pathological) are rarely validated.
- Foundation-model mesh recovery, differentiable inverse kinematics, and video kinetics are active in the computer-vision literature but essentially absent from validated biomedical studies.
- Markerless capture is not yet interchangeable with marker-based systems for clinical joint kinematics, while spatiotemporal parameters are ready now for monitoring, screening, and remote assessment.
Strengths
Genuinely novel dual-tier synthesis design
The MACRO/MICRO pairing (Sections 2.6, 3.5, Fig. 8) systematically maps 117 validated pipelines onto 572 curated CV records across 17 pipeline blocks, turning the usually anecdotal 'CV moves faster than validation' argument into a family-by-family bridge analysis showing triangulation is the only family importing emerging blocks.
Transparent, pre-registered methodology with explicit exclusion accounting
OSF-registered protocol (doi 10.17605/OSF.IO/TBYPC), PRISMA-ScR checklist, three databases with dates, dual Covidence screening with a senior adjudicator, and eight pre-specified A-H criteria with counts (e.g., 11 studies excluded for black-box systems under criterion A); the full 9,346-record flow is traceable (Section 3.1, Fig. 1).
Candid, two-sided clinical benchmarking
Unusually rigorous and honest LLM-assisted extraction QA
Clinically actionable outputs beyond description
Weaknesses
Heterogeneous error metrics pooled into single medians
Out-of-plane claims rest on tiny, internally inconsistent cells
Exclusion of the most clinically deployed systems narrows the clinical-readiness conclusions
No risk-of-bias or study-quality appraisal
Author-assigned architecture taxonomy without reported reliability
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
Is markerless motion capture accurate enough for clinical gait analysis?
Not yet for joint angles, according to this review of 117 validated studies: median sagittal-plane errors run about 5-6°, and roughly half of reported estimates exceed the 5° level above which McGinley et al. consider errors likely to mislead clinical interpretation. Walking speed and step-length measures agree much more closely and are considered ready for screening, monitoring, and remote assessment. Frontal/transverse angles and force-plate-free kinetics remain research-grade.
How many studies did this review include and from where?
It included 117 studies validated against an external biomechanical reference, screened from 9,346 records in PubMed, Scopus, and IEEE Xplore (January 2015 to February 2026). It also assembled a separate, deliberately non-exhaustive pool of 572 computer-vision papers (arXiv and CVPR/ICCV/ECCV/WACV/NeurIPS proceedings, 2024-2026) used only qualitatively to map emerging methods.
Which markerless motion-capture architecture is most accurate?
Per the review, two-dimensional monocular pipelines and multi-view triangulation performed similarly (median lower-limb sagittal errors of roughly 4-6°); monocular 2D-to-3D lifting was worse (about 6-7° with the widest spread, consistent with depth ambiguity); and mesh-recovery and end-to-end families had four or fewer studies each, too few to compare. No family reached the 2° 'acceptable' band at the median.
Why does the review exclude Theia3D and other commercial systems?
The pre-specified exclusion criterion A removes black-box commercial pipelines (Theia3D, Captury, KinaTrax) because their internal pose-estimation models cannot be identified, characterised, or reproduced. The trade-off, acknowledged by the authors, is that the review therefore describes the validated open-architecture evidence base rather than the systems most commonly deployed in clinics.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 1-9 could not be analysed (text extraction only); PRISMA counts and figure-based results were taken from the prose and tables.
- Additional files 1-7 (search strings, extraction table, MICRO corpus, per-metric breakdowns) are referenced but not contained in the embedded text, so several accuracy claims could not be independently verified.
- The abstract's final sentence and the reference list end appear truncated in the extraction; some tail content may be missing.
- Bibliographic entries were not checked against external records.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.