Article Dans Une Revue The Visual Computer Année : 2025

MoCoSys: human motion correction based on deep learning coupled with 3D+t Laplacian motion representation

Résumé

In this paper, we present a 3D motion correction system whose objective is to produce motion for computer animation purposes from 3D human pose estimation. Currently, most 3D pose estimation approaches fail to reconstruct motion that meets the specific requirements for 3D avatar animation. In particular, they often do not ensure the stability of the underlying skeleton or the temporal fluidity and coherence of the motion. We propose an additional process to address these limitations. Our system uses deep learning techniques in combination with Laplacian motion modeling, along with algorithms designed to enhance the temporal characteristics of the motion while preserving the integrity of the skeletal structure throughout the sequence. The approach is based on two deep neural networks. The first network utilizes a 3D+t graph representation of motion, combined with a discrete Laplacian operator, to improve the spatio-temporal deformation of the skeleton over time. The second network estimates fixed bone lengths within the skeletal structure, enabling the correction process to maintain skeletal consistency throughout the motion, while minimizing bone lengths errors in the reconstructed motion. Experiments conducted on the outputs of state-of-the-art neural architectures demonstrate that our system significantly enhances both the spatial and temporal characteristics of the reconstructed motion. This improvement ensures that the corrected movements are suitable for use in data-driven animation applications.

Fichier principal
Vignette du fichier
The-Visual-Computer_Motion-Correction-System_Tchenegnon-Gibet-LeNaour.pdf (2.99 Mo) Télécharger le fichier
Origine Fichiers produits par l'(les) auteur(s)
Licence

Dates et versions

hal-05446774 , version 1 (07-01-2026)

Licence

Identifiants

Citer

Mansour Tchenegnon, Sylvie Gibet, Thibaut Le Naour. MoCoSys: human motion correction based on deep learning coupled with 3D+t Laplacian motion representation. The Visual Computer, 2025, 41 (12), pp.10347-10362. ⟨10.1007/S00371-025-04041-1⟩. ⟨hal-05446774⟩
44 Consultations
82 Téléchargements

Altmetric

Partager

  • More