AUTOMATIC EXTERIOR ORIENTATION OF AERIAL IMAGES IN URBAN ENVIRONMENTS
C. Drewniok, K. Rohr
Fachbereich Informatik, Universität Hamburg
Vogt-Kölln-Straße 30, D-22527 Hamburg
|. drewniokQinformatik.uni-hamburg.de
http://kogs-www.informatik.uni-hamburg.de/staff/Drewniok.Christian.html
Commission lll, Working Group 2
KEY WORDS: Vision, Automation, Extraction, Orientation, Registration, Urban Scene, Model-Based Detection, Invariants
ABSTRACT
We present an approach for automatic exterior orientation of aerial images which is based on the use of manhole covers as
landmarks. The approach includes two main procedures: First, a landmark extraction scheme which enables us to automatically
detect and precisely localize many of the manhole covers visible in an image. Second, an automatic matching procedure which
robustly and efficiently matches constellations of detected landmarks with the correct landmarks from a cadastral database.
By combining both methods we are able to automatically generate a large number of landmark correspondences per image,
which allows for a reliable estimation of the exterior orientation parameters.
1. INTRODUCTION
The development of operational procedures for the automatic
orientation of aerial images is a matter of topical interest
in photogrammetric research (cf. Gülch (1995)). It requires
the automation of each step in the orientation process. For
absolute orientation this process includes:
1. Estimation of the interior orientation by detecting, pre-
cisely localizing, and identifying the fiducial camera
marks, knowing the type of the camera used.
2. Detection and precise localization of landmark features
within the image, assuming that these features corre-
spond to known geodetic coordinates.
3. Identification of the features by determining correspon-
dences between the landmarks detected in the image
and existing landmarks from the observed scene.
4. Estimation of the exterior orientation by spatial resec-
tion based on the landmark correspondences.
For automatic interior orientation operational techniques
have recently been worked out (Schickler, 1995), which make
use of the existence of well-defined geometric models of the
fiducial marks and profit from the well-structured appearance
of these marks within the image (completely isolated features
with high contrast). Opposed to this, in the automation of
exterior orientation one has to deal with real scene objects
and their complex appearance in aerial images. As a conse-
quence, only very few approaches to automatic exterior ori-
entation have been developed so far (see (Schickler, 1992)
and (Vosselman & Haala, 1992)).
In this contribution we present an approach to automatic ex-
terior orientation which is based on a specific type of circular
landmarks. We suggest that manhole covers are well suited
features to serve as landmarks for orientation of urban scenes.
The advantages of using this kind of landmark are manifold:
A great number of manhole covers can be found in urban
environments, most of them being placed in the middle of
a road; they are well distributed and located at the ground;
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996
geodetic coordinates (including altitudes) are available from
the cadaster of the city's sewerage system; and, as we will
show, they can be detected, localized with high precision,
and associated with geodetic coordinates from the cadastral
database, all in an automatic manner.
This contribution elaborates on the two major aspects in this
context: First, how to detect landmarks of the considered
type automatically and, second, how to match constellations
of detected landmarks with the manhole positions from the
cadastral database. Note that in the following the interior
orientation of the images is assumed to be known.
2. MODEL-BASED DETECTION AND
LOCALIZATION OF LANDMARKS
Our landmark extraction approach is based on a parametric
model which explicitly describes the location, size, shape, and
systematic intensity variations of depicted manhole covers.
Minimizing the squared intensity error between the model
and the image results in the best-fit model parameters and,
most important, determines the landmark's position in the
image with high sub-pixel precision. This can be shown for
simulated as well as for real image data. A short description of
the landmark extraction approach will be given below. More
details can be found in (Drewniok & Rohr, 1995).
2.1. Analytic Description of Circular Landmarks
While the appearance of manhole covers varies from country
to country we frequently find a specific type which consists
of a bright disk surrounded by a dark concentric ring (see
Figure 1, left). Since aerial images normally are recorded
approximately parallel to the ground plane, images of these
objects are circular. The idealized image intensities of a cross-
section through a manhole cover of the considered type form
a symmetric step function. When we also take into account
that the intensities are blurred because of the band-limiting
effect of the camera, we get a rounded shape as sketched in
Figure 1 (right). This profile can approximately be described
by three characteristics: haz, Amin, and rmin, where imas
146
and Amin €
and minimı
denotes th:
suggested -
profile usin
sponds tot
the shape :
eters (amp
needed for
landmark.
function, w
ond deriva:
three parar
MC
with r? — |
coordinate:
Given our 1
instance of
the image
the sum of
ties I and |
zo, and yo
centered ai
Since we a
iterative Le
ror functio
the first p:
parameters
result is gi
2.2. ‚The
In our lan:
of potenti:
cross-corre
using a la
marks det:
timization
lytic mode
The result
mally ada;
Both are c
decide wh:
instance. |
of false de
the actual