Graphics Reference
In-Depth Information
1.5.2.2 Feature-Based Stereo Vision Algorithms
General Overview
In the framework of feature-based stereo, correspondences
between pairs of stereo images are established based on the similarity between fea-
tures extracted for certain pixels of interest, such as conspicuous points or lines.
A comparative discussion of stereo algorithms based on corner feature points is
provided by Vincent and Laganière (
2001
). They describe the Plessey operator in-
troduced by Harris and Stephens (
1988
) as an important detector for corner points.
Here, however, we follow the original presentation of that detector as given by Har-
ris and Stephens (
1988
), who formulate the underlying structure tensor as
∂I
∂x
2
w
∂I
∂x
∂y
∗
∂I
∗
w
M
=
(1.103)
∂I
∂x
∂y
∗
w
∂I
∂y
2
∂I
∗
w
with
I
as the pixel grey value and '
...
w
' as a convolution with an image window
function
w
of predefined size, where a radially symmetric Gaussian shape of the
function
w
is given as an example. The eigenvalues of
M
are denoted by
λ
1
and
λ
2
. Harris and Stephens (
1988
) point out that edges are characterised by the occur-
rence of one large and one small eigenvalue, while large values of both eigenvalues
indicate a corner. Furthermore, Harris and Stephens (
1988
) suggest to use the value
R
=
∗
2
as a threshold value for corner extraction, where the
parameter
k
governs the sensitivity of the corner detector, instead of determining
and analysing the eigenvalues of
M
.
Vincent and Laganière (
2001
) point out that the eigenvalues
λ
1
and
λ
2
of
M
denote the curvatures of the intensity profile at an extracted corner point. In this
context, they present the method by Kung and Lacroix (
2001
), who define the 'cor-
nerness' as
c
det
(M)
−
k
·[
trace
(M)
]
λ
1
+
λ
2
|
=|
and establish correspondences between pixels in the stereo
images based on the smaller divided by the larger associated cornerness value. In
a different approach, Vincent and Laganière (
2001
) note that the directions of the
eigenvectors of the matrix
M
correspond to the directions of the edges which form
the corner. The average vector of the two eigenvectors then defines the 'corner orien-
tation', which can be used as a criterion for establishing a correspondence between
two corner points.
As a further feature point detector, Vincent and Laganière (
2001
) mention the
'univalue segment assimilating nuclei' (USAN) approach introduced by Smith and
Brady (
1997
). According to the presentation by Smith and Brady (
1997
), the USAN
area is computed by regarding a circular mask around a central pixel, called the
'nucleus', and determining the number of pixels inside the mask which have a sim-
ilar grey value. For uniform image areas, this value is similar to the area of the
mask, while it decreases at edges and further decreases at corners. These consid-
erations give rise to the concept of 'smallest univalue segment assimilating nuclei'
(SUSAN), which allows one to identify minima of the USAN area with edges and
corners in the image.
Zhang and Gimel'farb (
1999
) use the SUSAN-based corner detector according
to Smith and Brady (
1997
) to determine feature points in uncalibrated stereo image
Search WWH ::
Custom Search