rammetric Engineering & 
$3. 
5. Earth science and remote 
and image registration of 
the wavelet transform. In: 
rence "Image and Signal 
ng II", Paris, France, 25-27 
d-Martinet, J. Morlet and P. 
ts: Time-frequency methods 
eal time algorithm for signal 
wavelet transform. Springer 
alage élastique d'images 
Docteur-Ingénieur, Ecole 
élécommunications. ENST- 
finite approach to surface 
S sciences et techniques de 
ls automatic visual obstacle 
f the Sth. International Joint 
ligence, Cambridge, August 
ustice, C. Gurney and J. 
of misregistration on change 
on Geoscience and Remote 
Mangolini, 1997. Fusion of 
rent Spatial Resolutions: 
of resulting images. 
' & Remote Sensing, 63(6), 
ıputing the n-dimensional 
application to Voronoi 
ial, 24. 
A Top Down Strategy for 
Simple Crossroads Automatic Extraction 
Boichis N.* - Cocquerez J.P.** - Airault S.* 
* Institut Géographique National / MATIS, France 
** ENSEA-ETIS / URA 2235, France 
Ph: 4-33 1 43 98 80 69 - Fax: +33 1 43 98 85 81 
Email: nicolas.boichis,sylvain.airault@ign.fr and cocquere@ensea.fr 
Commission II, Working Group 6 
KEY WORDS: Image understanding, crossroads detection, interpretation system, Hough Transform 
ABSTRACT 
The French National Geographic Institute (IGN) is working on problems related to the photogrammetric data 
capture automation. Based on our experience around the road network automatic extraction and on our 
cartographic requirements, we consider that crossroads detection cannot be simplified to a road hypothesis 
grouping problem. The inability for actual system to satisfy the cartographer and the necessity to focus 
our attention on crossroads areas impose the introduction of external data. A basic representation of simple 
crossroads drives the objects detection and set up a precise crossroads model in an image analysis system. 
The respect of the geometric accuracy, topology and object shape will build up our interpretation strategy. 
Intermediate results will demonstrate the interest of our specific modelization and the limit of the actual system. 
The efficiency of the strategy is discussed to propose solutions to the different identified problems. 
RESUME 
L'Institut Géographique National (IGN) travaille sur les problémes liés à l'automatisation des chaines de saisie 
photogrammétrique. Forts de notre experience dans les systémes d'extraction automatique du réseau routier et 
de nos exigences cartographiques, nous considérons qu'il n'est pas possible de simplifier l'extraction des carre- 
fours routier à une simple interpolation entre différentes hypothèses de routes. On a mis en évidence l'incapacité 
des sytèmes actuels à satisfaire le cartographe et on demontrera le besoin de focaliser notre attention sur les 
carrefours. Un modèle adapté aux carrefours simple guide dans un système d’analyse d’image la detection des 
objets et instancie un modèle précis de carrefour. Le respect de la précision géométrique, de la topologie et de 
la forme de l'objet vont constituer les bases de notre stratégie d'interprétation. Des résultats intermédiaires de- 
montreront l'intérét d'une modélisation spécifique et mettront en évidence les limites actuelles de notre système. 
L'efficacité de notre approche sera ensuite discutée pour proposer quelques amélioration indispensables. 
1 INTRODUCTION cartographer but the user interpretation is totally dif- 
ferent. 
IGN is creating a new topographic database and looks According to the state of the art in crossroads 
for methods to reduce operator’s control and to ac- detection, most of existing methods first focus on 
celerate the photogrammetric stereoplotting. The road extraction to create the road network. The 
content corresponds to 1:25 000 maps with a met- detection of crossroads is often realized by percep- 
ric accuracy in the 3 dimensions and its acquisition is tual grouping of road hypotheses. On that account, 
based on stereoplotting of 1:30 000 aerial images (a their local behaviour depends on road geometric qua- 
0.5 m of resolution) . The analysis of aerial images re- lity and a small planimetric mistake can introduce a 
quires a complex reasoning, connected to the inherent large modification of crossroads shape in cartography. 
complexity of images and man made objects (roads, [Groch, 1982] and [Heipke et al., 1995] use a road fol- 
crossroads, buildings). The figure 1 illustrates the lowing method and detect crossroads when the road 
required accuracy and the cartographic final repre- width changes. This method is sensitive to short edge 
sentation. These three extractions are topologically interruptions. [Ruskoné, 1996] proposed a method 
correct and the geometric accuracy should satisfy a 
19