Item

Abstract

Estimating the pose of a rigid body means to determine the rigid body motion in
the 3D space from 2D images. For this purpose, it is reasonable to make use of
existing knowledge of the object. Our approach exploits the 3D shape and the
texture of the tracked object in form of a 3D textured model to establish 3D-2D
correspondences for pose estimation. While the surface of the 3D free-form
model is matched to the contour extracted by segmentation, additional reliable
correspondences are obtained by matching local descriptors of interest points
between the textured model and the images. The fusion of these complementary
features provides a robust pose estimation. Moreover, the initial pose is
automatically detected and the pose is predicted for each frame. Using the
predicted pose as shape prior makes the contour extraction less sensitive. The
performance of our method is demonstrated by stereo tracking experiments.