Robustness of markerless biomechanical analysis using pose2sim: a sensitivity study on transmission and privacy constraints for cloud-based computation
Résumé
Biomechanical analysis for performance monitoring, fatigue analysis and injury risk prevention has traditionally been confined to laboratory settings using marker-based motion capture. While markerless approaches using multiview computer vision have emerged, their reliability under "in-the-wild" or broadcast constraints remains underexplored. In the context of cloud computation, where all content has to be sent over the network, we investigate the feasibility of near real-time privacy-preserving performance monitoring by assessing the sensitivity of the Pose2Sim pipeline to degradation in acquisition parameters. Specifically, we evaluate the impact of face deidentification, camera count, framerate, image resolution, and H.264 video compression (QP values) on the recovery of biomechanical quantities. Except for face deidentification, which is required to preserve participant anonymity, all other values influence the total bitrate, which is limited in real-time scenarios. Using the AthletePose3D dataset, we establish a high-fidelity ground truth and measure the joint angle deviations under constrained conditions. Our results provide a guideline for minimum technical requirements to maintain biomechanical fidelity in unconstrained environments. Specifically, we recommend using video compression, while framerate is the most critical factor in acquisition settings. More importantly, we report that face deidentification can damage the extraction of faithful biomechanical quantities.
| Origine | Fichiers produits par l'(les) auteur(s) |
|---|---|
| Licence |