Triangulation-Based Approaches to Three-Dimensional Scene Reconstruction - 3D Computer Vision: Efficient Methods and Applications - page 38

in the acquired calibration images. It has been used for the convergent stereo cam-
era setup employed in some of the three-dimensional surface inspection scenarios
described in Chap. 6 , which consists of two Baumer CCD cameras of 1032

×

776

pixel image size equipped with Cosmicar video lenses of f 25 mm, for the trinoc-

ular Point Grey Digiclops system used for the three-dimensional reconstruction of
flexible objects in Chap. 6 , and for the trinocular camera system used in the applica-
tion scenario of human-robot interaction described in Chap. 7 . In the first scenario,
the distance to the scene amounts to about 0 . 5 m, in the second scenario to approx-
imately 1 m, and in the third to 2-7 m. The uncertainty of the measured image
coordinates of the control points due to pixel noise is examined in Sect. 1.4.8 .
The quality of camera calibration strongly depends on the number of images
available for calibration and on the positions in space where the calibration rig is
located during image acquisition. In this context, Krüger ( 2007 ) states the following
four 'rules of thumb' to yield a suitable set of calibration images.

=

1. In order to be able to determine the lens distortion coefficients, images in which
the calibration rig completely fills the image must be acquired for each camera.
2. Images of the rig should be acquired when the rig is positioned at the near side,
in the middle, and at the far side of the space in which three-dimensional scene
reconstruction is performed.
3. For all three distances, the calibration rig should also be placed such that it ap-
pears close to the corners of the images, respectively.
4. At each selected position, the orientation of the calibration rig should be perpen-
dicular to the optical axis as well as rotated by an angle of about 30 ◦ ± around

the horizontal and the vertical axis.

As additional examples, Fig. 1.9 shows the result of automatic calibration rig ex-
traction from images acquired with somewhat 'exotic' lenses, an omnidirectional
catadioptric lens and a fisheye lens. The proposed graph-based rig extraction algo-
rithm performs well despite the strongly distorted shape of the calibration rig in the
images. Even when not all control points are found (cf. Fig. 1.9 a), the assignment
of the extracted control points is nevertheless correct. Hence, the proposed method
can be used for automatic calibration rig extraction for a variety of lens architec-
tures.

1.4.8 Accurate Localisation of Chequerboard Corners

This section provides a comparison between different calibration pattern types and
describes a model-based method for estimating the image position of chequerboard
patterns which provides highly accurate results even in the presence of strong dis-
tortions. The section is adopted from Krüger and Wöhler ( 2011 ).