Registering Depth Maps from Multiple Views Recorded
by Color Image Sequences

Rüdiger Bess

Lehrstuhl für Mustererkennung (Informatik 5), Friedrich-Alexander-Universität Erlangen-Nürnberg,
Martensstr. 3, D-91058 Erlangen, Germany

tel.: +49 9131 85-7891; fax: +49 9131 303811

e-mail: bess@informatik.uni-erlangen.de

Abstract—In this paper we outline a system to search and classify objects in an unknown environment. As a
first step towards this goal a dense depth map of the complete object surface is computed by use of a stereo
approach adapted to monocular color image sequences. The resulting depth data of the stereo algorithm are
input of the judgement, registration and fusion step described in this paper.

After judging the depth values and removing the poor ones we use common lines of sight to identify depth val
ues in consecutive depth images which refer to the same 3D point. This allows to compute the registering trans
formation directly, no matching step of the 3D data is necessary. Advantages of this new approach are the inde
pendence of the registration step from accuracy in calibration and the possibility to register depth images con
taining only depth values of a single plane.

In the fusion step a 3D accumulator is used to combine depth values belonging to the same 3D point. In this
step we integrate filtering of neighboring depth values to restrict surface resolution and data size to the maximal
resolution of the resulting complete 3D map of the object surface. This enables an extremely efficient evaluation
of the depth images.

The algorithms described in this paper allow one to overcome the relatively inaccurate calibration and to com
pute 3D data from monocular image sequences by passive stereo approaches without placing a calibration pat
tern in the scene.

Pleiades Publishing home page | journal home page | top

If you have any problems with this server, contact webmaster.