Triangulation-Based Approaches to Three-Dimensional Scene Reconstruction - 3D Computer Vision: Efficient Methods and Applications - page 86

In all four example scenes, the number of three-dimensional points forming spu-
rious objects is significantly reduced; in particular, clusters of incorrectly assigned
points disappear. Accordingly, in mobile robot navigation or human-robot interac-
tion scenarios the proposed method is useful for distinguishing between real objects
between the camera and the repetitive structures and spurious objects, which would
otherwise induce a halt of the robot motion in such systems.
The applicability of the proposed technique is limited by high spatial frequencies
of the repetitive structures, corresponding to wavelengths (in pixels) comparable
to or shorter than the typical disparity differences between the model plane and
the corresponding three-dimensional points. The disparity difference between the
spurious objects and the true object corresponds to the wavelength of the repetitive
structures. Hence, given the typical inaccuracies of the model parameters derived in
Sect. 1.6.3 , the wavelength of the repetitive structures should be larger than about
2 pixels for the outdoor scenario and larger than about 20 pixels for the indoor
scenario to ensure that the adapted model does not extend into the spurious objects.
These minimum values decrease linearly with increasing accuracy of the estimated
model parameters. They do not imply strong restrictions for the proposed method:
In the indoor scenario, the minimum required wavelength of the repetitive structures
is one order of magnitude larger than the resolution limit of the image, while in the
outdoor scenario it approximately corresponds to the image resolution.
1.6.5 Discussion
The described model-based method by Barrois et al. ( 2010 ) for resolving stereo
matching ambiguities is independent of the specific stereo algorithm used. Scene
models have been applied which are represented by a single plane or several con-
nected planes. In three example scenes with extended areas of very strongly pro-
nounced repetitive structures, the proposed model-based refinement procedure de-
creases the fraction of false correspondences in the scene part displaying repetitive
structures by factors of up to 30, while the fraction of three-dimensional points
correctly assigned to the object decreases only moderately. In the example scene
showing an object in front of an outstretched arm, the improvement is somewhat
less pronounced but still significant. Furthermore, it has been shown that a state-
of-the-art model-based stereo approach is also able to provide a three-dimensional
reconstruction of scene parts displaying repetitive structures without generating spu-
rious objects, at least as long as initial pose parameters and information about the
metric extension of that scene part are provided. However, that method fails to cap-
ture scene parts or objects which are not part of the model. In contrast, the method
described in this section has proven capable of suppressing spurious objects while
at the same time achieving a correct three-dimensional reconstruction of objects in
front of the scene part displaying repetitive structures.