ilter Yu., "Shape
gic paradigm and
vol. 2350, 1994.
bject recognitions
rm", IEEE Pacific
itions, Computers
nd, 1989.
vatia R., "Model
- in "Automatic
from aerial and
n at al, Basel:
natic matching of
tomatic extraction
nd space images"/
user, 1995.
AUTOMATIC EXTRACTION OF LARGE BUILDINGS FROM HIGH RESOLUTION
SATELLITE IMAGES FOR REGISTRATION WITH A MAP
V K Vohra and I J Dowman
University College London
Commision III, Working Group III/2
KEY WORDS: Feature Extraction, Boundary Matching, Match Points, Correlation, Registration.
ABSTRACT
Feature extraction plays an important role in extraction of meaningful structure in digital images. It aims at replacing the
iconic representation of the image content by a symbolic image description, representing the essential parts of an image
related to some task. In the work described in this paper the main interest is in extracting large buildings as polygonal
features from high resolution satellite data which can be matched with maps for registration. The emphasis is on
polygons which have a distinctive shape and can be extracted from vector as well as raster data. Algorithms used are
described, for extraction of polygonal features from map and image, and also for matching map and image features to
generate match points. The match points derived are used for determining transformation parameters. The registration of
image to map is achieved by a applying bilinear resampling method using the parameters on the image. Results in terms
of accuracy are given and future work is also discussed for fully automation of image to map registration.
1. INTRODUCTION
The quality of images, the basic models and evaluation
algorithms, and also the way the data is represented are
important parts for the efficiency of the image processing
systems. Image processing is applied in several stages to
extract important image information, to suppress redundant
information and neglect information which is not used in
the processes.
Feature extraction is a very active area of research for
photogrammetry and remote sensing. Forstner (1993),
Schenk (1993) and Sester (1993) have given comprehensive
reviews of the current status and philosophy. Points, lines
or polygons can be extracted as primitives, but work to
identify and extract features is at present object oriented.
Holm et al, (1995) has described a method of map to image
registration for particular flat terrain with many lakes in
Finland. The automation of the relative orientation process
has been described by Haala et al, (1993) and by Hellwich et
al, (1994). The basis of both these methods is the use of
points determined by interest operators within an image
pyramid, the matching of these points and the use of iconic
and geometric constraints to eliminate false matches.
Heipke (1993) has reviewed automation in orientation and
work is in progress to use features such as roads or polygons
instead of points to determine the orientation of images,
either single or stereoscopic. Stevens et al, (1988) have
attempted to match map features directly with images. Lee et
al, (1993) have described an automatic method of registering
an image to a map in two dimensions which uses non-
rigorous methods but claims high speed and accuracy.
Shahin et al, (1994) have been investigating the automatic
detection of roads from SPOT data and matching these to
road co-ordinates determined by GPS.
The present work leads to the automation of registration of
an image to a map which is not yet solved in a flexible and
general way. This is in contrast to the image to image
registration problem which has been solved and can be
applied to images from the same sensor or from different
903
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996
sensors. Automation of the registration processes is based
on methods of feature extraction, as the main operation is
concerned with finding a set of common points i.e. match
points on the map and on the image. The features which are
most suited to this are polygons because they can have a
distinctive shape and can be found in urban and non urban
areas. Suitable points are not easily found automatically
from map data, but polygons can be extracted comparatively
easily from both raster images and vector maps. Therefore,
the extraction of polygonal features are considered in this
paper to generate sets of match points from map and image
for the registration of image with a map. The whole
procedure is divided into three stages:
* Extraction of polygonal features from map and image.
* Matching polygonal features of map and image to generate
a dense network of match points.
* Using match points for registration of image to map and
assessing the results in terms of accuracy for the validation
of the registration system.
2. FEATURE EXTRACTION
Large buildings can be seen very clearly and distinctly on
the high resolution satellite images. Considering this the
extraction of building features is planned for the test and the
area of the Defence Research Agency site at Farnborough,
UK is selected which shows large buildings and airfield
facilities. The data used to extract buildings in polygonal
form is taken from a high resolution Russian DD5 digital
image with approximately 2.5m pixel size and from an
Ordnance Survey 1:10 000 map in raster format.
It is important to bring the map and the image into same
input data level i.e. edge representation for matching
purpose. The map and the image, are seperately prepared
here to get the boundaries of buildings in the map and the
edges of buildings in the image.
