Triangulation-Based Approaches to Three-Dimensional Scene Reconstruction - 3D Computer Vision: Efficient Methods and Applications - page 31

1.4.7.2 Existing Algorithms for Extracting the Calibration Rig
To extract the image coordinates of the control points from the images, the software
package by Bouguet ( 2007 ) provides a manual solution which prompts the user to
click near the outer corners of the rig. It is completely out of scope for an auto-
matic calibration process, and it makes even the occasional calibration process very
cumbersome and prone to errors. A good calibration requires about 20 images per
camera. At four clicks per image (if no mistake is made), this amounts to 240 clicks
with a precision of less than four pixels for a trinocular camera system.
The OpenCV library 1 provides a corner detection algorithm as well. It is based
on an initial scanning process with a subsequent subpixel accurate search of the cor-
ners. The initial scanning process operates on binary images, extracts the potential
corner candidates, and tries to sort them with a polygonal approximation in a loop.
If all expected corners are determined, the result is assigned to the subpixel accurate
location. It is a gradient minimum search with a relocation of a neighbourhood win-
dow until the centre stays within a given threshold. In practise, this contour analysis
turns out to be rather instable. The OpenCV algorithm is not able to properly detect
and sort the corners under outdoor lighting conditions or under those in an indus-
trial production environment. In order to avoid singularities due to parallel image
and calibration rig planes, the calibration rig must be imaged at a certain obliquity.
The corner analysis of the OpenCV implementation often fails if the angle or the
camera-to-rig distance is too large, mainly due to an imprecise polygon approxima-
tion.
Two proprietary implementations based on cross-correlation template matching,
line fitting, and subsequent precise position estimation at subpixel accuracy have
been conceived as the predecessors of the algorithm described in Sect. 1.4.7.3 .The
first one operates on the assumption of a fully visible calibration rig. Hence, it ex-
tracts a number of most prominent features, equal to the number of corners in the
calibration rig, by means of cross-correlation matching (Aschwanden, 1993 ). Sub-
sequently, an outlier detection is performed based on a Hough transform (Jähne,
2005 ) and geometric constraint evaluation. Upon this, the lines are approximated
by least-squares methods. Independent of the previously detected features, the line
intersections are used as coarse corner guesses, and a maximum search followed by
fitting a paraboloid to the correlation coefficients is performed. The second method
discards the assumption of a fully visible calibration rig. Instead of extracting a fixed
number, it selects features according to their reliability. The coarse feature sorting
and outlier detection is performed as described above; then, for each potential rig
corner, it is determined if a rig-corner template is displayed. The final sorting before
the determination of feature position at subpixel accuracy is performed by accumu-
lating features along line segments into approximated lines.
These two state-of-the-art algorithms are very robust with respect to occlusion, il-
lumination, and noise. Due to the efficient implementation of the correlation process
1 The OpenCV library is accessible at http://opencv.willowgarage.com .