Triangulation-Based Approaches to Three-Dimensional Scene Reconstruction - 3D Computer Vision: Efficient Methods and Applications - page 58

smallest disparity is selected in order to avoid the formation of spurious obstacles at
small distances.
A similar assignment scheme is proposed by Stein ( 2004 ) in the context of real-
time optical flow analysis. This method regards local neighbourhoods of arbitrary
size. The relation between the intensities of the central pixel and the neighbour-
ing pixels is encoded as a chain of digits, called the 'signature vector'. Each digit
denotes if the corresponding pixel is significantly darker, of similar brightness, or
significantly brighter than the central pixel. This scheme can be regarded as an ex-
tension of the census transform introduced by Zabih and Woodfill ( 1994 ). Corre-
spondences are established by determining pixels in the two images with identical
signature vectors, where regions with many highly ambiguous assignments, such
as areas of uniform intensity, are excluded. Furthermore, point correspondences are
preferentially established based on signature vectors that do not occur frequently
in the images. Outliers are removed based on a temporal analysis of the resulting
optical flow vectors. Stein ( 2004 ) points out that the proposed assignment scheme
can also be used for establishing stereo correspondences.

A Contour-Based Stereo Vision Algorithm A feature-based stereo vision ap-
proach relying on the analysis of object contour segments is introduced by Wöhler
and Krüger ( 2003 ) in the context of surveillance of working areas in industrial pro-
duction. The presentation in this section is adopted from that work. This contour-
based stereo vision (CBS) algorithm is based on the comparison of the current im-
age pair with a pair of reference images. To detect changes between the current
image and the reference image, we compute the absolute difference image. There
are much more complex methods of change detection; cf. e.g. Durucan ( 2001 )for
an overview. Generally, however, these cannot guarantee that a zero image resulting
from change detection is equivalent to the fact that the current and the reference
image are identical. A zero difference image guarantees that the current and the
reference image are exactly identical, which is of significant importance for the ap-
plication scenario, as a person in the surveillance area must not be missed under any
circumstances.
The image pair is rectified to standard geometry. We transform the pair of differ-
ence images into binary images by thresholding with

√ 2 σ p , (1.104)
where the pixel noise σ p is the standard deviation of a came r a pixel signal over

θ 0 = with σ d =

qσ d

time, given a constant input intensity, and the noise σ d = √ 2 σ p is the resulting
pixel noise of the difference image. In our experiments, we set q = 3 in order to

detect only changes which are with 99 % certainty significant with respect to pixel
noise.
The image regions with pixel grey values above θ 0 are segmented using the
binary connected components (BCC) algorithm (Mandler and Oberländer, 1990 )
which yields, among others, the properties area, centre coordinates, and a con-
tour description for each blob. This computation yields n 1 blobs on the right and
n 2 blobs on the left image with centre coordinates (U (i)

,V (i)
1

), i = 1 ,...,n 1 and

1