Triangulation-Based Approaches to Three-Dimensional Scene Reconstruction - 3D Computer Vision: Efficient Methods and Applications - page 65

of the direction in the image. Correspondences are then established based on the
maximisation of flow in a graph.
A survey about dense stereo methods and an evaluation framework essentially
based on synthetic images to assess their performance is provided by Scharstein and
Szeliski ( 2001 ). Van der Mark and Gavrila ( 2006 ) examine dense stereo algorithms
with respect to their applicability in vehicle-based real-time vision systems. Based
on synthetic image data that mimic the properties of real vehicle-based images and
also on real-world images acquired during test drives in the urban traffic scenario,
they show that algorithms which rely on global optimisation of the disparity map
suffer more strongly from the variability of the encountered conditions under which
the scene is imaged than approaches relying on simpler selection criteria to estab-
lish point correspondences. Furthermore, it is demonstrated by Van der Mark and
Gavrila ( 2006 ) that the method proposed by Hirschmüller et al. ( 2002 ) in the context
of correlation-based blockmatching stereo vision (cf. Sect. 1.5.2.1 ) shows the best
performance of all examined algorithms when applied in a real-time setting for the
analysis of traffic scenes.
A general drawback of dense stereo algorithms is the fact that the established
depth values tend to be inaccurate for parts of the surface that do not show any sur-
face texture at all, or for corresponding parts of the stereo image pair which do not
display a similar structure. The latter behaviour may e.g. occur as a consequence
of specular reflectance properties leading to a different appearance of the respec-
tive surface part in the stereo images. In such cases of missing or contradictory
texture information, dense stereo algorithms usually interpolate the surface across
the ambiguous image parts, leading to an inaccurate three-dimensional reconstruc-
tion result for the corresponding region. This problem is addressed explicitly by
Hirschmüller ( 2006 ), who proposes the semi-global matching method which estab-
lishes correspondences based on similar pixel grey values to obtain plausible dispar-
ity values for surface parts free of small-scale structures. Furthermore, Hirschmüller
( 2006 ) suggests an interpolation approach which maintains sharp depth disconti-
nuities rather than blurring them and which is able to fill regions in the disparity
map where the disparity information is unavailable or incorrect. In the presence
of systematic differences between the pixel grey values of corresponding points,
Hirschmüller ( 2008 ) suggests the use of mutual information as a similarity measure
for disparity estimation. An important advantage of the semi-global matching ap-
proach by Hirschmüller ( 2006 , 2008 ) is that it has high computational efficiency,
while also providing disparity estimates of high subpixel accuracy.
1.5.2.4 Model-Based Stereo Vision Algorithms
A higher level stereo method which is complementary to the previously described
approaches is model-based stereo. The approach introduced by Tonko and Nagel
( 2000 ) relies on establishing correspondences between grey value gradients in the
image ('edge elements') and edges of object models in the scene ('model seg-
ments'). The detected objects are tracked using an extended Kalman filter. In the