Applications to Industrial Quality Inspection - 3D Computer Vision: Efficient Methods and Applications - page 255

robot places the cable back to its original position, as shown in Fig. 6.11 f. A lo-
calisation accuracy of 1 mm is required for mechanical reasons to enable the robot
to grasp the cable, which is faithfully achieved by the multiocular ziplock snake
approach.
In the application scenario of three-dimensional reconstruction of tubes, a sim-
plified variant of the multiocular contracting curve density (MOCCD) algorithm
is used by Krüger and Ellenrieder ( 2005 ). An extensive evaluation based on the
full MOCCD approach according to Hahn et al. ( 2007 , 2010a ) as described in
Sect. 2.2.1.2 is provided by Krüger ( 2007 ), who compares the three-dimensional
reconstruction accuracy of the multiocular ziplock snake method to that of the
MOCCD and the gradient sign table approach (cf. Sect. 2.2.1.2 ). The addressed
scenario is termed the 'dangling rope problem' by Krüger ( 2007 ), i.e. the three-
dimensional position of the non-rigid object and its tangential direction is known
a priori for only one of its ends (in the application scenarios regarded in this sec-
tion to evaluate the multiocular ziplock snake approach, this information has always
been available for both ends of the objects). The Digiclops trinocular camera sys-
tem is used for image acquisition, and the distance to the object corresponds to
about 1 m. Krüger ( 2007 ) arrives at the conclusion that the accuracy of the mul-
tiocular ziplock snake method degrades less strongly when the edge of the object
in the image is incomplete. It is shown by Krüger ( 2007 ) that the multiocular zi-
plock snake and the MOCCD yield comparable average deviations between the
three-dimensional reconstruction and the ground truth of 1-3 mm as long as the
contrast between object and background is high, and that under less favourable con-
ditions (low contrast or shaded object edges), the MOCCD algorithm may become
more accurate than the multiocular ziplock snake by up to an order of magnitude,
while the multiocular ziplock snake permits smaller apparent object sizes than the
MOCCD algorithm. Furthermore, Krüger ( 2007 ) finds that the accuracy of the mul-
tiocular gradient sign tables approach is generally similar to that of the multiocular
ziplock snake and the MOCCD but tends to be more robust with respect to low
contrast.
6.3 Inspection of Metallic Surfaces
This section describes applications of the previously described three-dimensional
surface reconstruction methods to the inspection of metallic surfaces mainly of au-
tomotive parts. Where possible, the results are compared to independently derived
ground truth values obtained by a laser profilometer or by tactile measurement.
Traditionally, photometric three-dimensional surface reconstruction techniques
such as shape from shading have been regarded as being unsuitable for industrial
applications (Mühlmann, 2002 ). State-of-the-art commercial three-dimensional sur-
face reconstruction systems for industrial quality inspection purposes are based
on active scanning techniques such as projection of coded structured light. It is
shown in this section, however, that especially for metallic surfaces, taking into