Applications to Industrial Quality Inspection - 3D Computer Vision: Efficient Methods and Applications - page 252

than 1 . 5 ◦ and 2 . 5 mm for the three rotation angles and the three translational pose
parameters, respectively.
The system of Stößel ( 2007 ) (cf. Sect. 2.2 ) has been developed for estimation
of the pose parameters of articulated objects in monocular images. When applied
to the real-world oil cap data by von Bank et al. ( 2003 ), the full extended kernel
particle filter approach shows the highest performance of the system configurations
examined by Stößel ( 2007 ), yielding a mean error between 0 . 1 ◦ and 0 . 3 ◦ for the
three rotation angles. Due to the stochastic nature of the employed optimisation
algorithm, however, the result of an optimisation run is non-deterministic. Hence,
the uncertainty of an individual rotation angle measurement amounts to 0 . 9-1 . 5 ◦ .
Stößel ( 2007 ) does not report accuracy values for the translational degrees of free-
dom.
This comparison illustrates that the pose refinement approach of Barrois and
Wöhler ( 2007 ) based on multiple monocular cues is able to estimate the rotational
pose parameters at an accuracy which is of the same order of magnitude as the
mean error of the method proposed by Stößel ( 2007 ), even under viewing directions
where geometric cues alone do not allow one to recover the object pose reliably.
It should be noted, however, that the uncertainty of an individual measurement by
Stößel ( 2007 ) is nearly an order of magnitude higher than the mean error. Presum-
ably, the high accuracy of the estimated rotational pose parameters achieved by our
method is due to the integration of photopolarimetric information. The translational
accuracy of our method in directions parallel to the image plane, which is largely
determined by the available edge information, is comparable to or slightly higher
than that observed for other pose refinement systems, which all rely to a consider-
able extent on edge information. The accuracy of the values obtained for the object
depth with the monocular depth from defocus technique comes close to that of the
system of Yoon et al. ( 2003 ), which is based on the evaluation of stereo image
pairs.
6.2 Inspection of Non-rigid Parts
In this section the three-dimensional active contour algorithm of d'Angelo et al.
( 2004 ) described in Sect. 2.2.2 is applied to the reconstruction of a cable, using im-
ages acquired with a Digiclops trinocular camera system at a distance to the object
of about 1 m, as shown in Fig. 6.9 . A three-dimensional ziplock ribbon snake has
been used in these examples. The approximate ribbon radius was used as model in-
formation, and constraints for the upper and lower bound of the ribbon width were
applied. The segmented contour part is displayed from a slightly different viewing
angle to show the spatial segmentation. The accuracy of reconstruction compared
to ground truth obtained with a calliper gauge is about 1 mm, corresponding to
1 . 5 pixels, in this example.
A different inspection scenario is the three-dimensional reconstruction of a glue
line on the non-planar free-form surface of a car body part, as shown in Fig. 6.10 .