Item

Abstract

This work presents a
marker-less motion capture system that
incorporates an approach to smoothly adapt a generic model mesh
to the individual shape of a tracked person.
This is done relying on extracted silhouettes only.
Thus, during the capture process the 3D model of a tracked person is learned.

Depending on a sparse number of 2D-3D correspondences, that are computed
along normal directions
from image sequences of different cameras,
a Laplacian mesh editing tool generates
the final adapted model.
With the increasing number of frames
an approach for temporal coherence reduces the effects of insufficient
correspondence data to a minimum and guarantees smooth adaptation results.
Further, we present experiments on non-optimal data that
show the robustness of our algorithm.