A Data-Based Parametric Biomechanical Running Model driven by Pose Estimation - A proof of concept study
Authors
Hansen, Jesper Bang ; Kristiansen, Nikolaj Hoelgaard
Term
4. semester
Education
Publication year
2022
Abstract
This proof-of-concept study tests whether a statistical, parametric running model can reconstruct full-body kinematics from hip and knee flexion estimated from a single smartphone video (pose estimation), thereby approaching a laboratory marker-based reference. Fifty-one treadmill trials from 18 male participants were recorded simultaneously with smartphone and marker-based motion capture at 240 Hz. Hip and knee angles from pose estimation were transformed into Fourier coefficients; the first harmonic (a1 and b1), together with height, weight, sex, age, angular stride frequency, and running speed (7–15.5 km/h), were used to drive a data-based parametric running model. Model outputs were compared with the marker-based reference for selected kinematic variables. Results showed excellent correlations (r > 0.90) for hip, knee, and plantar flexion and vertical center-of-mass position, and strong correlations (0.67 < r ≤ 0.90) for hip abduction and elbow flexion. Root mean square differences (RMSD) were 4.0–8.5 degrees for knee and plantar flexion and hip abduction, 10.0–14.9 degrees for hip and elbow flexion, and 0.021 m for vertical center of mass. Overall, the approach can estimate realistic running kinematics from simple video input, but reducing magnitude errors will require more and more accurately measured input parameters.
Denne proof-of-concept-undersøgelse afprøver, om en statistisk, parametrisk løbemodel kan genskabe fuldkrops-kinematik ud fra hofte- og knæfleksion estimeret fra en enkelt smartphone-video (poseestimering) og dermed nærme sig en laboratoriereference med markører. I alt blev 51 løbebåndsforsøg fra 18 mandlige deltagere indsamlet med samtidig smartphone- og markerbaseret bevægelsesregistrering (240 Hz). Hofte- og knævinkler fra poseestimering blev omsat til Fourier-koefficienter, hvor første harmoniske (a1 og b1) sammen med højde, vægt, køn, alder, vinkelstegfrekvens og løbehastighed (7–15,5 km/t) blev brugt til at styre en data-baseret parametrisk løbemodel. Modellens forudsagte bevægelser blev sammenlignet med markerbaseret reference for udvalgte størrelser. Resultaterne viste fremragende korrelationer (r > 0,90) for hofte-, knæ- og plantar-fleksion samt vertikal massecenterposition og stærke korrelationer (0,67 < r ≤ 0,90) for hofteabduktion og albuefleksion. Rodmiddelkvadratfejl (RMSD) lå på 4,0–8,5 grader for knæ- og plantar-fleksion samt hofteabduktion, 10,0–14,9 grader for hofte- og albuefleksion, og 0,021 m for vertikalt massecenter. Samlet peger fundene på, at metoden kan estimere realistisk løbekinematik fra simple videodata, men at flere og mere præcist målte inputparametre er nødvendige for at reducere størrelsesfejl i modeludgangene.
[This apstract has been generated with the help of AI directly from the project full text]