78
CMRT09: Object Extraction for 3D City Models, Road Databases and Traffic Monitoring - Concepts, Algorithms, and Evaluation
Looking across track (camera tilt perpendicular to base)
Looking along track (camera tilt parallel to base)
►x
Figure 1: Geometry of across track (left) and along track (right) baseline in oblique stereo images
2 STEREO GEOMETRY AND ACCURACY IN
OBLIQUE IMAGES
The dense matching results in point correspondences in a stereo
pair 5 . The image rays as resulting from those matching image
pairs are forward intersected in order to compute 3D points in ob
ject space. In this section a brief theoretic approximatoin of the
expected accuracy from the given intersection geometry in obli
que airborne images is derived. Two cases can be distinguished:
A) the two cameras used for the stereo intersection are oriented
across track, i.e. they inclose a right angle with the base, and B)
the cameras are looking in flight direction. The intersection ge
ometry of the first case can be derived from the standard normal
case, but the varying scale and the tilt of the camera coordinate
system wrt the actual coordinate system need consideration. The
second case can be compared with the longitudinal tilt-setup (Al-
bertz and Kreiling, 1980), i.e. the normal case with different ca
mera heights. Only here the whole base is tilted. In Fig. 1 both
camera geometries are sketched.
The scale within an oblique image depends on the flying height
H, the focal length c the tilt angle t, and the angle between the
viewing ray to a target and the vertical ß (symbols according to
(Höhle, 2008)):
H ■ cos(ß — t)
TO = — -
c • cos ß
(1)
where to: scale at target point. At the principal point 0 equals f,
whereas at the fore- and the background /3 is determined from t
and the half field of view a: 0f O re = t — a and ftback = t + a.
A: tilt across track (side-looking) In the vertical image case,
the accuracy for triangulated points in height (s' H ), and in X-Y
plane (sx,y) can be estimated by:
Finally, the respective error components need to be projected
from the tilted system to the actual coordinate system:
sh ~ yj(s' H ■ cos t) 2 + (s[ x y ■ sinf) 2 , (5)
Sx,y ~ yj{s’ H ■ sini) 2 + (s' x Y ■ cost) 2 , (6)
thus for a tilt angle of 45° both components will be identical.
B: tilt along track (forward-looking) To derive the accuracy
in the tilted system H, X', Y 1 , first the necessary parameters for
the case of longitudinal tilt need to be computed: Base B' in the
tilted system and the heights of the cameras I and //:
H
= B ■ cos t,
(7)
AH'
= B-sint,
(8)
Hi = to • c,
and H'u = Hi- AH'.
(9)
Applying partial derivation wrt the image and parallax measure
ments to the formulas given in (Albertz and Kreiling, 1980), the
accuracies for the coordinate components in the tilted system can
be derived:
s'hj ~ s' HlI « y/(™) 2 ' s p x + '5, 00)
s'x,Y ~ ~ ■ S x - (11)
Note that the actual parallax needs to be computed for the estima
tion. In the approximations for the given data, see below, a mean
parallax according to a mean depth in fore- and background was
assumed. For the planar accuracy the more pessimistic estima
tion, assuming the smaller image scale, is given here. Finally, the
planar and height components in the actual coordinate system are
computed according to equations 5 and 6.
3 METHODS ADOPTED
sh ~ -g--m-s px , (2)
s 'x,y ~ s x ■ to « s y • to, (3)
where s x « s y « 0.5 • s px are the standard deviations for image
coordinate and parallax measurements; the errors in the orienta
tion components are neglected. In the case of the tilted camera
system, these formulas are applicable to the tilted system, so the
varying scale needs to be considered, according to equation 1,
also H' needs to be adopted accordingly:
H' — to • c. (4)
5 For the combination of multiple views see the experiment section
3.1 Block adjustment for multiple platforms
In (Gerke and Nyaruhuma, 2009) a method to incorporate scene
constraints into the bundle block adjustment is described and tested.
The bundle block adjustment algorithm uses horizontal and ver
tical line features, as well as right angles to support the stability
of block geometry. Those features can be identified at building
façades, as visible in oblique images. In addition, the approach
is able to perform self-calibration on all devices which are incor
porated in the block. This is an important issue in the case of
oblique images, as those are often acquired by non-metric came
ras. The extension to the setup used for this paper where images
from different platforms are involved is done without any change
to the core approach.