her new and offers
otogrammetry.
rted by the Austrian
>10354-MED.
Image Processing.
ence 751. Berlin
ed Correspondence
ternational Archives
ensing, Vol. XXVI,
r for Detection and
orners and Centres
mmision Workshop
grammetric Data”,
jares Correlation: A
jue, South African
mote Sensing &
- 187.
try: Volume: jl: i=
esses, Dümmler /
1a 1996
AUTOMATIC REGISTRATION OF IMAGES WITH MAPS USING POLYGONAL FEATURES
I J Dowman, A Morgado and V Vohra
University College London
Commission III, Working Group III/2
KEY WORDS: Automation, Matching, Orientation, Registration
ABSTRACT
The automation of the full orientation procedure for images from aircraft or satellites has progressed significantly during the
past 4 years. In particular inner and relative orientation have matured into automatic production processes and aerial
triangulation, with the exception of a general method of identifying ground control points, is well on the way to the same
stage. This paper reports on the development of tools and on an overall strategy which permits automated registration of an
image or 3D model to a map or a digital data set using large polygonal features, stored in raster or vector format. No
manual identification of discrete points is required and the methods takes into account any changes which might have taken
place between the original information being compiled and the current imagery being obtained. The method can also be
used for registering images to images and allows a wide range of image types to be used, ranging from aerial photographs to
synthetic aperture radar (SAR) data.
1. INTRODUCTION
The problem of absolute orientation of aerial photography
can be tackled in a number of ways. Automated aerial
triangulation is a promising method in which photographs in a
block can be joined using a large number of conjugate points
which are automatically determined. A relatively small
number of ground control points then have to be identified
manually, although Gülch (1995) has shown that certain
types of ground control points can be identified automatically.
If the ground points are premarked then the problem is
tractable. The orientation of single pairs or of satellite images
is more difficult because a relatively larger number of ground
control points are required. For high accuracy work at large
scales, high level feature extraction is required and is the
topic of a large amount of research. At smaller scales, and
particularly when map revision is the main application, then
an alternative approach is possible in which the image is
registered directly to the existing map or data base, referred
to as the reference data. A necessary condition of such an
approach is that the matching of common features must be
done in two dimensions. In some cases, for example using
satellite images over flat ground, this might be sufficient, but
is other cases the third dimension must be introduced after
matching. This approach offers a great deal of flexibility and
many of the advantages seen in the matching strategies used
for automatic digital elevation computation. For example the
determination of many redundant matched points and the
ability to tune the system for different types of image.
The system described here is designed as a flexible generic
system which will allow a range of image types to be
registered with a reference data set. The system will be fully
automatic but will offer the user a number of tools to employ
to optimise the solution for the particular data being used and
to validate the end result.
The method is based on the selection of conjugate polygons
which can be uniquely identified in the image and the
reference data. A number of well established methods are
available for segmentation and edge detection for use in
defining homogenous areas on an image. Polygons are more
easily identified uniquely in vector data than are points or
lines, hence polygons are highly suited to this task.
Recognition of the possible distortions due to tilt and relief
must be made but these are generally easily modelled.
The paper first outlines the strategy to be used and then
describes a number of algorithms which have been developed
and tested. Three examples are given in the paper. New
work which carries out the full absolute orientation process
on a pair of aerial photographs using a 1:10 000 map is
described in some detail. Previously published results
showing the matching of woodland from attributed layers ofa
1:50 000 map and a Landsat Thematic Mapper image, and
matching of large buildings on a 1:10 000 raster image and a
1:10 000 map are summarised.
2. ASTRATEGY FOR IMAGE TO MAP
REGISTRATION
2.1 Preparation of the reference data
The strategy is shown in figure 1. Each stage will be
discussed in outline but reference should be made to the
papers cited for further detail. The method involves the
separate preparation of the image and the reference data. The
preparation of the reference data will differ according to the
source. If the data is available in digital form, with
attributes attached, than the process of extracting polygons of
a given type is straightforward. A paper map by itself
provides a greater challenge and some manual operations are
still necessary to identify particular features. If the original
layers used in the printing stage are available then this will
enable features such as woods or water bodies to be easily
isolated. If such layers are not available then manual
identification may be necessary. One approach to this is
described by Vohra and Dowman (1996) in which polygons
have been identified semi-automatically and then only those
exceeding a specified area are retained.
The degree of processing on the image depends on the
method to be used. If general polygons are to be used then
edge extraction and segmentation are needed. Generally
homogeneous areas are looked for and a multispectral
classification may assist if features such a woodland or water
are used.
139
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996
