A Bayesian 
Netherlands, 
Cover Using 
1g Countries, 
ate the Initial 
| Visions and 
t 4, Stuttgart, 
dy of Phrao, 
e analysis in 
16-20, ITC, 
onvention of 
Data Access. 
nmission III, 
Agricultural 
Proceedings 
ammetry and 
Improvement. 
\pplication of 
] University, 
onference on 
ge Parameter 
ration of GIS 
$93), Tokyo, 
Proceedings 
epts, Methods 
rged Edition, 
iges in Land 
ian, A., 1999, 
idan Province 
epresentation 
Tal Abramovich 
  
A QUANTITATIVE MEASURE FOR THE SIMILARITY BETWEEN FEATURES 
EXTRACTED FROM AERIAL IMAGES AND ROAD OBJECTS IN GIS 
Tal ABRAMOVICH and Amnon KRUPNIK 
Department of Civil Engineering, Technion — Israel Institute of Technology, Haifa, Israel 
{tabramov, krupnik} @tx.technion.ac.il 
Commision III, WG I11/3 
KEY WORDS: Feature extraction, GIS, Modeling, Recognition, Fuzzy 
ABSTRACT 
Automated approaches for extracting roads form aerial images, and for verifying the correctness of extracted 
roads, usually involve matching of graphic entities from different sources. Of particular interest is the matching 
of GIS data to information extracted from images. This paper presents an approach for matching road segments 
extracted from aerial images to road objects in an existing GIS. In order to match different entities with a 
minimum number of errors a model to quantify the similarity of road segments has been developed. The method 
presented here is based on fuzzy logic and conceived as a better way for sorting and handling vague, imprecise, 
or incomplete information. The paper describes the motivations, outlines the method and presents preliminary 
results. 
1 INTRODUCTION 
Updating is a very important task required for improving and preserving the quality of the data in Geographical 
Information Systems (GIS). Updating is usually done by verifying new information with respect to existing data 
and by extracting new objects. Common sources of new information are aerial and satellite images, while the 
actual extraction is usually performed by a human operator. Such a procedure obviously requires extensive and 
tedious work. Performing this task automatically with a high level of detail requires comprehensive prior 
knowledge. Many studies have been carried out in the past two decades concerning extracting man-made 
features in general, and roads in particular (see Heipke ef al., 1997 for review). However, automatic extraction 
of man-made features is still in its research stage. 
One way of introducing knowledge into automatic extraction of man-made objects is to exploit existing data. In 
particular, existing GIS data are shown valuable. These data provide knowledge about the cartographic objects 
in the scene and a description of their properties. A general framework for considering existing GIS data in the 
updating process consists of three steps: 
l. Extracting image features that may be hypothesized as road segments. The basic assumption is that 
a road appears in the image as a set of parallel edges. Additional properties are also used to 
perceptually group raw features into objects that are then hypothesized as roads. Roads are also 
extracted from the GIS data, and each segment is assigned a set of properties. Some of these 
properties appear in the database, while others are derived from the geometry. 
Comparing objects from the database to the hypothesized road segments from the image. The 
problem is to match graphic entities that appear differently in two overlapping sources. Such 
differences can be observed in the geometrical and topological descriptions. At the end of this step, 
there exist three groups of results: (i) matched objects; (ii) objects that appear in the database but do 
not appear in the image; and (iii) features in the image hypothesized as roads but do not exist in the 
database. Features of the latter group are either non-roads or new roads that should be verified 
carefully in the third step. 
3. Classifying image features that were hypothesized as roads segments, and were not matched in the 
second step. This process is based on a learning procedure in which properties of matched roads are 
used to verify road hypotheses. At the end of this step new road segments are determined and linked 
to existing roads. 
N 
This paper deals with the second step mentioned above. It defines a method for quantifying the similarity 
between a road entity from a GIS, and a hypothesized road feature extracted from an image. The method 
  
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000. 17