Upper Limb Joint Force Estimation via Inverse Dynamics and Monocular Human Pose Estimation: Assessment with In-Vivo Shoulder Data

November 28, 2024

Marker-based motion capture systems are the gold standard for tracking human motion; however, their high cost and time-consuming setups limit their practicality in clinical and sports settings. Therefore, markerless systems utilizing Human Pose Estimation algorithms have gained attention. This work evaluates the accuracy of the monocular Human Pose Estimation algorithm VideoPose3D in estimating key biomechanical parameters during four movements using inverse dynamics and an upper limb musculoskeletal model. The evaluation is validated against a marker-based system, as well as in vivo force data measured from a shoulder implant and electromyography signals. To address VideoPose3D’s limitation in wrist tracking, a Hybrid Model was developed that integrates hand kinematic data from the MediaPipe Pose algorithm into the VideoPose3D output. The Hybrid Model displayed large errors in estimating certain body keypoint coordinates (>100 mm) and the glenohumeral joint Axial Rotation angle (>28º). Nevertheless, its predicted reaction forces at this joint exhibited strong to excellent correlations (0.89-0.95) with the in vivo forces, with Root Mean Squared Errors between 55.01 to 122.13 N. The Hybrid Model’s ability to detect shoulder muscle activations was comparable to the marker-based system. These findings suggest that monocular Human Pose Estimation algorithms show promise as alternatives to marker-based systems, though further improvements are needed.

Authors in the community:

• 

  M

  Maria João Candeias Pio

  ist195950

Supervisors of this institution:

• 

  C

  Carlos Quental

  ist154212

• 

  I

  Ivo Fialho Roupa

  ist427060

industrial-biotechnology - Industrial Biotechnology

eng - English

Instituto Superior Técnico