Introducing a new constraint reduces the number of tests to
investigate. This constraint on the coplanarity of two object
points involves two instead of three correspondences,
reducing the order of the number of tests to n*. Further
significant reduction of the number of tests is achieved by
computing a number of attributes for each constructed image
point. A correspondence hypothesis is only set up when the
attributes of the two points match to a sufficient degree.
The procedure for deriving relative position, epipole
detection in short, consists of the following five steps:
1 Compute points in each image through line
intersection.
2 Establish correspondence hypotheses between the
computed points.
3 Compute statistical tests for coplanarity of all
combinations of two correspondences.
4 Grouping of correspondences based on the test results
and epipolar plane intersection constraints.
5 Refinement of grouped correspondences by overall
adjustment and testing.
Each step is repeated for each of the four possible
permutations of the rotation matrix of the second image. The
exterior orientation of the first image determines the object
co-ordinate system.
3.3.1 From lines to points
Points are created in each image as intersection of two image
lines being projections of edges with a different
(perpendicular) orientation in object space. This orientation
is determined by the vanishing point detection procedure.
Two lines are intersected when an endpoint of a line is within
a preset distance (commonly set to 5 pixels) from the other
line. Furthermore, the following attributes are computed and
registered for each point:
e The orientation of the plane passing through the object
point (perpendicular to the orientations of the two
intersected lines)
e The type of junction (T-junction or L-junction)
e The orientation of the junction (4 options: Figure 4)
e The ratio of the lengths of the two edges
These attributes are all evaluated in object space, i.e. after
projecting the lines onto a plane of which the orientation is
known, again through the vanishing point detection.
n:
LT 291]
Figure 4: Four T-junctions (top) and four L-junctions
3.3.2 The correspondence hypotheses
In the next step of the procedure a list of possible
correponding image points is set up. A correspondence is
added to the list only when the attributes of points have been
compared. Both points have to have the same (plane)
orientation, the same junction type, and the same junction
orientation. The ratio of the line lengths should be similar for
both points with a typical maximum difference of 50%. The
correspondence search turned out not to be sensitive to this
criterion.
The creation of image points, their attributes, and
—230—
consequently the correspondence hypotheses depend on the
current permutation of the rotation matrix of the second
image. With each permutation the vanishing point (i.e. object
orientation) line labels for X and Y are exchanged.
3.3.3 Statistical testing of coplanarity
With the image orientations known, the epipole is defined by
a minimum of only two correspondences (the intersection of
the two related epipolar planes; Figure 3). Furthermore,
because each image point is constructed from the intersection
of two lines of which the spatial orientation is known from
the vanishing point detection, the orientation of the object
plane in that point is known. With the epipole derived from
the two correspondences, the location of the two object
points in model space can be computed by forward
intersection, and thus the position of the two parallel planes
through those points and the distance between the planes is
known. The coplanarity constraint requires this distance to be
Zero.
For the formulation of the coplanarity constraint the ray of an
image point is rotated to the object co-ordinate system. After
replacement of equation (1) by
x" zR'(x y-f) (6)
in which i refers to one of the two images, equations (2) and
(4) are used to compute the orientation vector of the epipole
(v) from two correspondences. Knowing the epipole, forward
intersection results in the distance (d) from the projection
center to an object point. This distance is projected onto the
normal n, of the object plane. For image 7 and
correspondence J:
=} i x! :n Dp
djzd,-— t (7)
i
Normal n, is a unit vector in the object X or Y direction,
dependent on the orientation of the façade. For the
formulation of the coplanarity constraints the first image is
used (in principle, using the second image would yield
identical results). The constraint for coplanarity of the two
object points can now be written as:
— 1 —
Qr did (8)
The statistical testing of the hypothesis is explained in (van
den Heuvel, 1998). The coefficients of the linearised form of
equation (8) are derived numerically.
3.3.4 Clustering of correspondences
For all combinations of two possible correspondences the
statistical test based on constraint (8) is evaluated. Each test
links two correspondences that define an epipole. The
clustering procedure aims at grouping of correspondences of
which the inter-correspondence tests are accepted, and at the
same time support the same epipole. In order to verify the
latter, a statistical test is used that is based on the intersection
of epipolar planes. This constraint is identical to the
intersection of interpretation planes constraint (3).
The clustering procedure includes the following steps:
e For each accepted test (8) it is checked whether one of
the two correspondences is present in an existing
cluster. If this is not the case, a new cluster is
established.
e [fa correspondence of an accepted test is present in an
existing cluster, the other correspondence becomes a
= 0 rs
tl
Les
c9 ds CS CN NT Rady
a IA WE MA m AY A fn S — - ("79 CH PF CH ed