ion. Sensor
ossibly) the
dvantage of
sions. This
D vectors x
nt the same
), from Eu-
CS).
ace X is de-
ined by pro-
r orientation
,» Un). scale
es 9.
we collected
d by an arbi-
ative orienta-
logous points
ith the 3 x 3
on for the ho-
e and elegant
1 images, i.e.,
a, there is no
je reduction is
1 S zz c 0 070
x= [0 17m 9 0° 0710 (5 a TC (1)
[iol aad, 0 gg } 9 1
e For all points x" in the second image which lie on the straight
line
ie FX
holds lx" — 0 and, therefore, the coplanarity condition. Le.,
1” is the epipolar line of x" in the second image. It can be used
to predict the geometrical location of x" in the second image
in the form of a straight line. The computation can be based
solely on F. There is no need to know the parameters of the
orientation of the two cameras.
e The epipoles e; 2 and e»,1 in the first and the second image,
respectively, are mathematically the left and right eigenvectors
corresponding to the eigenvalue O of the fundamental matrix:
el oF = Qand Fez; = 0.
e The bilinear form is linear in the coefficients of the fundamen-
tal matrix. This allows for a direct determination from homol-
ogous points.
e The 3 x 3 fundamental matrix has 9 elements and is singular
with rank 2. Because it is additionally homogenous, i.e., a mul-
tiplication with a scalar # 0 does not change the result, it has
only 7 DOF. The condition |F| = 0 has to be enforced, which
is cubic in the parameters of F.
If calibration data is available, the fundamental matrix reduces to the
essential matrix E and its bilinear form "x TE"x” = 0 with the
reduced image coordinates "x’ = K’~'x’ and "x" = K"~'x". The
essential matrix can be obtained from the fundamental matrix via
E — K"'FK' with the calibration matrices K' and K" of the first
and the second image, respectively. From the point of view of the
geometrical spaces, the fundamental matrix describes a projective
space and the essential matrix a "similar" space.
For three images we use the trifocal tensor. Its concept and its linear
computation were presented in (Hartley, 1997). In photogrammetry
(Fórstner, 2000, Ressel, 2000, Theiss et al., 2000) have described
the trifocal tensor and reported about experiments. The orientation
based on image triplets has some advantages over the orientation
based on image pairs:
e The orientation can be based upon homologous points in the
same way as on (infinitely long) homologous straight lines.
e The conditions for the homologous points as well as for the
homologous straight lines are linear in the observed homoge-
nous coordinates. They are additionally linear in the elements
of three 3 x 3-matrices or more precisely of the 3 x 3 x 3-tensor
constructed from these matrices.
e Practical experience shows that the local geometry of an image
strip or an image sequence can be much more precisely and,
what is more important, also much more robustly determined
from image triplets and their conditions than from only weakly
overdetermined pairs. This is true for the trifocal tensor but also
for bundle triangulation. Opposed to the latter, the trifocal ten-
sor has like the fundamental matrix its strength in its linearity.
Linearity equals speed and this makes the determination of ap-
proximate solutions, e.g., based on RANSAC (cf. Section 4.1)
possible.
The trifocal tensor can be introduced intuitively based on homol-
ogous straight lines (Hartley and Zisserman, 2000). Given are a
straight line 1’ in the first and a straight line 1” in the second im-
age. The planes ©’ = PI” and x” PTY" constructed from
these lines intersect in the 3D straight line L, the image of which in
the third camera is in general the straight line 1”.
rU = PTT Y"
T is a bilinear transformation and defines the trifocal tensor which
is usually written as qu It is a 3 x 3 x 3 cube made up of 27
elements. To has 18 DOF at maximum. Le., not all cubes are
trifocal tensors. The number 18 is, e.g., obtained by subtracting from
the 33 parameters of the three projection matrices the 15 parameters
for a projective transformation (homogenous 4 x 4 matrix) of space.
3 POINT TRANSFER WITH THE TRIFOCAL TENSOR
Based on the trifocal tensor a prediction of points and straight lines
in the third image is feasible without determining the point or the
straight line in space. Le., the trifocal tensor describes relations be-
tween measurements in the images without the need to reconstruct
3D geometry explicitly. In principle this corresponds to the epipolar
line for the image pair, but opposed to it the result is unique.
For the general case the prediction for points could be done by in-
tersecting the epipolar lines in the third image corresponding to the
homologous points in the first and the second image, respectively.
This is true only if the epipolar lines are not parallel, which is the
case if a point lies on the trifocal plane, or if the projection centers
are collinear. The latter is often valid or at least nearly valid, e.g., for
aerial images from one flight strip.
The restriction of the preceding paragraph does not hold if we em-
ploy the trifocal tensor: Given two homologous points x' and x"
in the first and the second image one chooses a line 1” through x”.
Then, the point x" can be computed by transferring x' from the
first to the third view via the homography defined by 1/7; 5 ie,
ak — a.
This transfer via the homography implies the intersection of the
plane defined by the projection center of the second camera O" and
the line 1” with the ray defined by the projection center of the first
camera O' and x' (cf. Fig. 1). This intersection is not defined if 1”
is taken to be the epipolar line corresponding to x'. The plane de-
fined by the epipolar line and O" comprises the point X which is
projected to x’, x", and x” as well as the projection center O" of the
first camera. In this plane lies also the ray defined by O' and x".
On the other hand, if one takes the line perpendicular to the epipolar
line through x", also the projection plane becomes in one direction
perpendicular to the ray defined by x' and O' and thus the intersec-
tion geometry becomes optimal. In (Hartley and Zisserman, 2000)
it is recommended to use optimal triangulation to compute a pair x
and x” which satisfies x""F,2x" = 0. We do not do this because
we are focusing on speed rather than on optimum quality.
Thus, our basic algorithm looks as follows:
—193—