Zum Inhalt

OpenPose Validation for Postural Assessment — Research Summary

Date: 2026-05-19 Status: [En desarrollo] — Validation protocol defined, implementation pending

Objective

Compare OpenPose BODY_25 automated pose estimation against manual Kinovea measurements and EPPA clinical evaluation for biomechanical analysis. The goal is to determine whether markerless 2D pose estimation can produce clinically equivalent angular measurements for posturographic variables currently assessed through the EPPA protocol.

Background

EPPA (Evaluacion Posturografica de Pie Automatizada) is a clinically validated postural assessment system that computes angular variables from standardized photographs with manual or semi-automatic marker placement. OpenPose is an open-source, real-time multi-person 2D pose estimation system developed at Carnegie Mellon University that uses Part Affinity Fields for bottom-up keypoint detection (Cao et al., 2021). Kinovea is a manual 2D video analysis tool widely used in sports science and clinical biomechanics as a reference standard for planar kinematic measurements.

This investigation line evaluates whether OpenPose BODY_25 keypoints can substitute or complement the manual landmark identification used in EPPA and Kinovea workflows.

Methodology

Pose estimation model

OpenPose BODY_25 detects 25 body keypoints per person from monocular RGB video or images. For this validation, the relevant keypoints include:

Keypoint BODY_25 ID Anatomical reference
Shoulder #2 (R) / #5 (L) Glenohumeral joint
Hip #9 (R) / #12 (L) Hip joint center
Knee #10 (R) / #13 (L) Knee joint center
Ankle #11 (R) / #14 (L) Ankle joint
Neck #1 C7-T1 estimated

A supplementary MediaPipe PoseLandmarker pipeline (33 keypoints) has been implemented and tested on synthetic images (see baseline-pipeline.md).

Analysis protocol

  1. Sagittal video analysis: Subjects perform load-lifting tasks (dorsal, squat, stoop) in two conditions — with and without exoskeleton. Camera positioned perpendicular to the sagittal plane at minimum 3 m distance, 1080p resolution, 30 FPS minimum.

  2. Three-way comparison:

  3. Automated (OpenPose/MediaPipe): Frame-by-frame keypoint extraction with confidence thresholds (visibility >= 0.5 per keypoint). Angular variables computed from keypoint triplets.
  4. Manual (Kinovea): Evaluator-driven marker placement on the same video frames. Serves as the 2D reference standard.

  5. Clinical (EPPA): Standardized posturographic evaluation from static photographs with anatomically defined markers.

  6. Variables measured: Trunk flexion, hip angle, knee angle, shoulder-hip vertical inclination. Full variable mapping is documented in variable-mapping.tsv and the sagittal lifting protocol in protocols/kinovea-openpose-sagittal-lifting.md.

  7. Statistical analysis: Mean Absolute Error (MAE), Root Mean Square Error (RMSE), Intraclass Correlation Coefficient (ICC 2,1 for concurrent validity; ICC 1,1 for test-retest), and Bland-Altman plots with limits of agreement. Analysis stratified by lifting type and exoskeleton condition.

  8. Calibration: Physical reference object of known length visible in frame. Kinovea calibration tool applied. OpenPose pixel-to-real conversion via FC = length_real / length_px.

Known limitations

  • OpenPose estimates joint centers, not bony anatomical landmarks used in clinical protocols (e.g., ASIS, C7 spinous process).
  • Equivalence between EPPA clinical markers and pose-estimation keypoints must be validated variable by variable; visual similarity does not imply measurement equivalence.
  • Exoskeleton wear may occlude shoulder and hip keypoints, reducing detection confidence.
  • Single-camera sagittal view cannot capture out-of-plane movements.
  • Frontal plane variables (pelvic tilt, shoulder height asymmetry) show poor accuracy in the literature (Ota et al., 2021).

Current status

Component Status
Literature review (17+ papers) Complete
Claim-evidence matrix In progress (claim-evidence-matrix.tsv)
Sagittal lifting protocol Defined (protocols/kinovea-openpose-sagittal-lifting.md)
Variable mapping (EPPA vs BODY_25 vs MediaPipe) Defined (variable-mapping.tsv)
MediaPipe baseline pipeline Functional on synthetic data
OpenPose Docker environment Built (v1.7.0, docker/)
Clinical video dataset Pending — requires authorized recordings with informed consent
Kinovea reference measurements Pending — depends on clinical dataset availability
Validation execution Not started

Dependencies

  • Clinical video dataset: Sagittal recordings of load-lifting tasks with and without exoskeleton. Must have informed consent and institutional authorization before processing.
  • Kinovea reference data: Manual measurements by trained evaluator on the same video set.
  • EPPA baseline: Static posturographic measurements from the same subject cohort for cross-method comparison.

Evidence from the literature

The following summary is drawn from the local bibliography collection (17 papers reviewed; see sources/local/investigacion-labis/openpose_validation_research.md for the full annotated review).

Sagittal plane accuracy (lower limb)

Joint MAE (deg) Context Source
Hip 3.7 +/- 1.3 Gait cycle Washabaugh et al., 2022
Hip 4.0 Walking Stenum et al., 2021
Knee 5.1 +/- 2.5 Gait cycle Washabaugh et al., 2022
Knee 5.6 Walking Stenum et al., 2021
Ankle 7.4 Walking Stenum et al., 2021

Squat validation (relevant to load-lifting)

Ota et al. (2020) found almost perfect test-retest reliability (ICC 1,3) for OpenPose during bilateral squat. Concurrent validity (ICC 2,1 vs. VICON) was almost perfect/substantial for trunk and knee, but fair for hip (~0.37) with proportional bias.

Upper body and limitations

Shoulder angles show the poorest performance across studies: RMSE 23.6--36.5 deg with ICC 0.28--0.72 depending on camera angle (Baldinger et al., 2025). Pelvic tilt in the frontal plane is unreliable with current pose-estimation methods (Ota et al., 2021).

Clinical populations

OpenPose has been validated for stroke (n=44) and Parkinson's disease (n=19) populations with comparable or better MAE than healthy cohorts, likely due to slower movement velocities (Stenum et al., 2024).

References

  1. Cao, Z., Hidalgo, G., Simon, T., Wei, S.E., & Sheikh, Y. (2021). OpenPose: Realtime Multi-Person 2D Pose Estimation Using Part Affinity Fields. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(1), 172--186. DOI: 10.1109/TPAMI.2019.2929257

  2. Stenum, J., Rossi, C., & Roemmich, R.T. (2021). Two-dimensional video-based analysis of human gait using pose estimation. PLoS Computational Biology, 17(4), e1008935. DOI: 10.1371/journal.pcbi.1008935

  3. Ota, M., Tateuchi, H., Hashiguchi, T., Kato, T., Ogino, Y., Yamagata, M., & Ichihashi, N. (2020). Verification of reliability and validity of motion analysis systems during bilateral squat using human pose tracking algorithm. Gait & Posture, 80, 62--67. DOI: 10.1016/j.gaitpost.2020.05.027

  4. Ota, M., Tateuchi, H., Hashiguchi, T., & Ichihashi, N. (2021). Verification of validity of gait analysis systems during treadmill walking and running using human pose tracking algorithm. Gait & Posture, 85, 290--297. DOI: 10.1016/j.gaitpost.2021.02.006

  5. Washabaugh, E.P., Shanmugam, T.A., Ranganathan, R., & Krishnan, C. (2022). Comparing the accuracy of open-source pose estimation methods for measuring gait kinematics. Gait & Posture, 97, 188--195. DOI: 10.1016/j.gaitpost.2022.08.008

  6. Baldinger, Reimer, & Senner. (2025). Influence of the Camera Viewing Angle on OpenPose Validity in Motion Analysis. Sensors, 25(3), 799.

  7. Stenum, J., Hsu, M.M., Pantelyat, A.Y., & Roemmich, R.T. (2024). Clinical gait analysis using video-based pose estimation: Multiple perspectives, clinical populations, and measuring change. PLoS Digital Health, 3(3), e0000467. DOI: 10.1371/journal.pdig.0000467

  8. Puig-Divi, A., Escalona-Marfil, C., Padulles-Riu, J.M., Busquets, A., Padulles-Chando, X., & Marcos-Ruiz, D. (2019). Validity and reliability of the Kinovea program in obtaining angles and distances using coordinates in 4 perspectives. PLoS ONE, 14(6), e0216448. DOI: 10.1371/journal.pone.0216448

  9. Kim, W., Sung, J., Saakes, D., Huang, C., & Xiong, S. (2021). Ergonomic postural assessment using a new open-source human pose estimation technology (OpenPose). International Journal of Industrial Ergonomics, 84, 103164. DOI: 10.1016/j.ergon.2021.103164

  10. Wade, L., Needham, L., McGuigan, P., & Bilzon, J. (2022). Applications and limitations of current markerless motion capture methods for clinical gait biomechanics. PeerJ, 10, e12995.