to- AUTOMATIZED UPDATING OF ROAD DATABASES FROM 
hii SCANNED AERIAL PHOTOGRAPHS 
rall 
hic 
ac M.E. de Gunst & M.J.P.M. Lemmens 
eci- 
ac- 
by Lab. of Photogrammetry & Remote Sensing 
rn- Faculty of Geodetic Engineering, Delft University of Technology 
and Thijsseweg 11, 2629 JA Delft, The Netherlands 
Commission IV 
ap- - 
ffi- 
hic ABSTRACT 
The aim of this paper is to present our approach and discuss preliminary results on automatizing the updating of a road 
database using scanned aerial photographs. 
Image understanding, easily performed by human operators, is the bottle-neck in automatizing photogrammetry. The 
ific complexity of aerial photographs requires the use of context information for object recognition. We show that the concept 
te - of a multi-stage approach can create a context that influences or guides the computer interpretation. The multi-stage 
nce approach embeds a two level approach for image analysis. The low level contains several image processing techniques for 
segmentation. The high level guides the segmentation by inference. It integrates knowledge sources, e.g. a priori 
information present in an existing database. 
We discuss in detail the verification whether a road segment in the database is present in the aerial image at the 
IT hypothesized location. It illustrates the huge complexity of aerial image interpretation. 
ES Keywords: Image analysis, Data Base, Change Detection, Map Revision. 
ote 
ser 
ind 
la 1. INTRODUCTION knowledge of a human operator and implement it in 
do digital photogrammetric systems. For image analysis 
The ultimate aim of our present research activities is and change detection we also use a priori knowledge 
SM to design a system that is able to update a road data- present in an existing database, or more generally, GIS 
base at a high degree of automation. (Semi)-automatic knowledge. (Lemmens, 1988; Lemmens, 1990) 
ng. updating of topographic databases is highly desirable 
ote since the need for high quality geo-information at all Our research domain is the updating of road databases. 
levels of geo-management is evident. However not the Roads express the principal structure of areas and their 
in production of aerial photographs and satellite images extraction from aerial photographs and satellite images 
In: forms the bottle-neck in the geo-information supply, is the subject of more research (e.g. McKeown and 
on but the photogrammetric processing of the image data. Denlinger, 1988; Cleijnenbreugel et al., 1990; Fischler 
91, In its present form it is time consuming and labour et al., 1981; Gunst et al., 1991, to mention a few). In 
intensive. contrast with this previous work we use a priori 
ted Characteristic for digital photogrammetry as opposed knowledge from a topographic road database. 
en, to analytical photogrammetry is access of the com- 
puter to the contents of the image, because it is stored At present we emphasize the concepts for the design 
ind in a digital form. This creates the possibility to use of the system, especially the integration of knowledge 
ner digital image processing tools for the development of sources and digital image processing techniques at the 
nd fully automatized photogrammetric systems. conceptual level. So implementation aspects, such as 
Geometric tasks, such as aerotriangulation and data structures, storage and manipulation of knowled- 
LS. orientation, can at present nearly be solved ge, communication of the control mechanism with the 
de- automatically by transferring experience from data and the image processing routines are not consi- 
OT analytical to digital photogrammetry. However tasks dered. 
nt, involving interpretation capabilities of human 
operators are very difficult to solve by computers. This paper is structured as follows. The next section 
OT Ima derstanding is the research field of computer discusses the complexity of interpretation of aerial 
ge understanding e p plexity rp 
nd vision and artificial intelligence. The updating problem images. This leads to the definition of concepts for 
can be approached as an image understanding this interpretation in section 3. Section 4 illustrates the 
al problem. A strategy is to formalize domain-specific presented concepts by a casestudy. 
477