Video-Based Markerless Motion Capture for Clinical and Rehabilitation Biomechanics: A PRISMA-ScR Scoping Review of Validated Architectures, Clinical Readiness, and Emerging Methods

Technical Summary: Video-Based Markerless Motion Capture for Clinical and Rehabilitation Biomechanics

Problem Statement
Three-dimensional motion capture (3D MoCap) using optoelectronic, marker-based systems (MBMC) remains the gold standard for clinical gait analysis, offering sub-millimeter spatial accuracy and sub-degree joint angle precision. However, its clinical adoption is hindered by high capital costs, dedicated space requirements, and operational overheads such as marker placement and soft-tissue artifact management. While video-based markerless motion capture (MMC) using RGB cameras and deep learning promises to democratize biomechanical analysis, a critical gap exists between the rapid advancements in computer vision (CV) and the validated evidence required for clinical deployment. It remains unclear whether current MMC pipelines deliver biomechanically accurate data that meets clinical acceptability thresholds, particularly for pathological populations and out-of-plane kinematics. Furthermore, the relationship between validated biomedical studies and the cutting-edge CV literature is poorly mapped, creating a lag in the translation of new algorithmic building blocks into clinically usable tools.

Methodology
The authors conducted a scoping review following the PRISMA-ScR (Preferred Reporting Items for Systematic Reviews and Meta-Analyses extension for Scoping Reviews) guidelines and the Joanna Briggs Institute (JBI) framework. The study employed a dual-tier design to address two distinct research questions:

1. MACRO Level (Systematic): A formal screening of biomedical databases (PubMed, Scopus, IEEE Xplore) from January 2015 to February 2026. This corpus included 117 studies that empirically validated an MMC pipeline against an external biomechanical gold standard (e.g., Vicon, Qualisys). Studies were excluded if they used black-box commercial systems, relied solely on manual digitization, used depth sensors (RGB-D) as the primary modality, or lacked an external gold standard.
2. MICRO Level (Curated Horizon Scan): A deliberately non-exhaustive, qualitative scan of emerging computer vision literature (arXiv, CVPR, ICCV, ECCV, NeurIPS) from 2024 to July 2026. This corpus identified 572 records representing "technological building blocks" (e.g., foundation models, differentiable inverse kinematics) that had not yet transitioned into validated biomedical publications.

Data extraction utilized a structured template covering 90+ fields, assisted by large language models (LLMs) for initial population and cross-verified by human adjudication to ensure accuracy. The analysis focused on architectural families, validation outcomes (joint angle accuracy, spatiotemporal parameters), and clinical readiness.

Key Contributions and Results

• Architectural Taxonomy: The review classified validated pipelines into five architectural families:

  1. 2D Monocular: Computes joint angles directly from 2D keypoints (40% of studies).
  2. Multi-view Triangulation: Reconstructs 3D positions via ray intersection from calibrated cameras (38%).
  3. Monocular Lifting: Regresses 3D pose from 2D detections (16%).
  4. Body Mesh Recovery: Fits parametric meshes (e.g., SMPL) to video (3%).
  5. End-to-End Direct: Regresses joint angles directly from images (3%).
     Most pipelines (73/117) reported raw geometric angles without biomechanical refinement (e.g., inverse kinematics or musculoskeletal modeling).

• Metrological Performance:

  • Accuracy: Sagittal-plane lower-limb joint angle agreement clustered around a median absolute error of 5° to 6°. According to McGinley et al.'s error bands, errors above 5° are likely to mislead clinical interpretation. Only ~6% of sagittal estimates fell within the "acceptable" (<2°) band, and ~37% were in the "reasonable" (2–5°) band.
  • Plane Coverage: There is a severe reporting bias toward the sagittal plane. Out-of-plane kinematics (frontal and transverse) are rarely validated; for the knee, 65 studies reported sagittal error versus only 5 for frontal and 3 for transverse. Where reported, out-of-plane errors were larger (e.g., ankle frontal error ~9°).
  • Kinetics: Kinetic validation (ground reaction forces, joint moments) was reported in only nine studies, leaving this critical clinical output largely unexamined.

• Population and Setting Limitations: The evidence base is narrow. Participants were overwhelmingly healthy adults (81%), with clinical/pathological populations (stroke, Parkinson's, cerebral palsy) comprising only ~19%. The vast majority of studies (100/117) were conducted in laboratory settings; home or field deployment is almost absent.

• The MACRO-MICRO Gap: The review identified a significant disconnect between validated biomedical literature and emerging CV state-of-the-art. Key building blocks present in the MICRO corpus but absent from validated MACRO studies include:

  • Foundation-model mesh recovery (e.g., SAM 3D Body).
  • Differentiable and neural inverse kinematics.
  • Video-based kinetics estimation.
  • Human-centric foundation models.

Significance and Claims
The paper concludes that video-based markerless motion capture is not yet interchangeable with marker-based systems for clinical joint kinematics. While spatiotemporal parameters (e.g., walking speed) show close agreement with reference systems and are suitable for screening, joint angle accuracy—particularly for out-of-plane movements and pathological populations—does not yet meet clinical standards.

The authors emphasize that the "clinical-acceptability gap" is a property of the field as a whole, not a single weak architecture. They argue that the most deployable tools (single-camera monocular pipelines) are currently the least refined, often stopping at raw 2D angles.

Design Guidelines and Future Directions
Rather than proposing a single validated method, the paper offers hypothesis-generating design guidelines to steer the next generation of pipelines. These guidelines advocate for:

1. Dense Parametric Mesh Front-Ends: Moving from sparse keypoints to dense meshes to capture segment orientation and out-of-plane rotation.
2. Subject-Specific, Model-Based Inverse Kinematics: Using differentiable or classical model-based solvers (e.g., OpenSim) rather than learned neural mappings to ensure generalizability to pathological gait.
3. Two-Level Mesh-to-Anatomy Correspondence: Calibrating a generic mesh-to-landmark map with a subject-specific residual to handle atypical morphologies.
4. Uncertainty Quantification: Integrating per-degree-of-freedom uncertainty estimates to inform clinical decision-making.
5. Closing the Validation Lag: Prioritizing validation in the populations and settings where it is most needed (older adults, pathological gait, real-world environments) over further benchmarking of healthy adult sagittal gait.

The paper positions itself as a bridge between the fast-moving computer vision frontier and the rigorous requirements of clinical biomechanics, urging that future progress depends on standardized, biomechanically grounded evaluation that reports joint angles rather than just keypoint position errors.