on Pattern 
9-256. 
3. Springer- 
88. A knot 
tting based 
oximations. 
y d. Gand 
d matching 
f Computer 
tching and 
grammetry 
g & Remote 
f randomly 
of B-spline 
27 pp. 663- 
. Springer- 
imensional 
ection in a 
| Report no. 
New South 
by splines. 
ind Remote 
mization in 
'asurement 
p. 267-290. 
sic Theory. 
'egistration 
1al Journal 
Rectangular Building 3D Reconstruction in Urban Zones 
Thierry QUIGUER 
ONERA-DES/STD, BP 72 
92322 Chatillon CEDEX, FRANCE 
quiguer@oncra.fr 
Commission III 1992-1996 
Theory and Algorithms 
KEY WORDS : Photogrammetry, Urban, Vision, Reconstruction, Algorithms, Edge, Stereoscopic, Three-dimensional 
ABSTRACT : 
Our paper concerns 3D reconstruction of buildings in urban and sub-urban zones by stereovision using vertical aerial 
images (resolution is in 40 cm rango). Our images arc well registrated, image lines correspond to epipolar lines. We 
limited our investigations to rectangular buildings because it is not an obvious problem. We propose a semi-automatic 
method in order to avoid major drawbacks of low-level processes. In effect, in low-level vision algorithms we need to 
introduce a priori knowledge (i.c. thresholds). So, in many cases we have to adapt thresholds to images. In order to 
overcome this particular unpleasant aspect, we focus our works on high-level process and we propose an original 
method to recognize building in an image. Our algorithm is semi-automatic because we select manually a corner then 
we apply our high-level algorithm. Results are very interesting because we obtain a good precision of detection and 
reconstruction. We compare our results with BDTOPO® (TOPOgraphic Data Base of French National Geographic 
Institute) which are truth data. 
1. INTRODUCTION 
Our paper concerns photogrammetry which consists in 
computing object dimensions by mcasures rcalized on 
perspective views of this object. We can find a large 
collection of papers concerning this domain, basic 
notions being available in the manual of photogrammetry 
[PhotoG 80]. Photogrammetry is a vast research domain 
so we deliberately restricted our investigation 10 
rectangular building reconstruction which is not an 
obvious problem, see Figure l'in order to illustrate this 
assertion. The size of this image is 2000 by 2000 pixels. 
  
  
    
(© French National Geographic Institute) 
657 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996 
Recent papers tackle this very difficult problem [Dang 
94] [Dissart 95] [Gabet 94] [Huertas 88] [McKweon 93] 
[Maitre 92] [Mohan 88] [Shufellt 93] and show that this 
problematics still stays a subject of interest in the 
international community. A common characteristic, about 
all these algorithms and about vision algorithms in 
gencral, can be pointed out: results of high-level process 
and consequently of complete process are dependant of 
low-level one. With this assertion two communities 
appear: those who neglect low-level process and consume 
time computational during high-level treatment, and 
those who try to have a perfect detection and 
consequently develop easy high-level technics. We think 
that an intermediate position will be better. Any detection 
process is perfect even if you provide several a priori 
knowledge. Thus we think it is important to overdetect 
primitives in image in order to provide all pertinent 
elements to the recognition level process. The job of 
high-level will be to separate good detections from false 
detections. We suppose low-level process provide 
weighted detection, weight qualifying quality of an 
element. This quality measure helps us during the 
high-level process. 
Nevertheless, in order to be sure that all pertinent 
elements will be detected, we have to choose between 
several a priori knowledge and interactivity. We choose 
the second option because we hope to climb automation 
ladder (see figure 2) when detection problem will be 
resolved. In figure 2 we qualify our approach using clas- 
sical critical systemic parameters used in literature. So, 
interactivity overcomes low-level problems and then we 
decorrelate some behavior parameters like automation 
and complexity of a priori knowledge.