TOWARDS AUTOMATED HOUSE DETECTION FROM DIGITAL 
STEREO IMAGERY FOR GIS DATABASE REVISION 
Zhongchao SHI 
Ryosuke SHIBASAKI 
Institute of Industrial Science, The University of Tokyo 
Email : shizc, shiba@shunji.iis.u-tokyo.ac.jp 
Commission IV, Working Group 3 
KEY WORDS : GIS, Urban, Database Revision, Automation. 
ABSTRACT 
To keep GIS database current, it is crucially important to capture the changes of man-made structures, especially houses, 
from digital urban imagery. In this paper, we describe a method aiming at extracting houses automatically from stereo 
imagery of urban areas for GIS database revision. We begin with the discussion on feature extraction based on wavelet 
analysis. Image segmentation techniques involving region segmentation and line feature extraction are then discussed. 
Region-based and line-based stereo matching algorithms for the purposes of disparity estimation, house detection, and 
change detection of houses for GIS database revision are addressed, respectively. Experimental results are demonstrated at 
the end of this paper. 
1. INTRODUCTION 
One of the most challenging approaches in the domain of 
GIS database revision is the automated 3D structure 
extraction, especially the house and road extraction, from 
complex photographs of urban or suburban scenes. 
Since the pioneering work of Nagao and Matsuyama 
(1980) in which structural analysis is applied for 
establishing fully automated system, a multitude of 
automated structure extraction methods have been 
proposed and tested such as shadow-analysis based 
algorithms(Liow and Pavlidis, 1990) and information- 
fusion based systems(Mckeown, 1991). Many of them, 
however, employed image shadows as a obvious 
information for predicting the presence of houses and 
estimating the height of houses. Moreover, hypotheses 
are often exploited in these existing studies, such as the 
roof of each house should have corners of 90 degrees(i.e., 
the even edges of a house should be parallel with one 
another and similar for odd edges, while the even edges 
have to be perpendicular to the odd edges), there should 
exist shadows on photographs, and so on. The question 
is that : when there are no explicit or complete shadows 
or/and there exist very complex and irregular roofs, can 
these algorithms work perfectly ? It seems that most of 
them are not robust enough to resolve such problems. 
780 
This is the major motivation for the research to be 
introduced later in this paper. 
In order to solve above problems, we believe that the 
knowledge of depth of the scenes in aerial images is 
significant in the absence of shadows, the presence of 
irregular roofs and occlusions. In practice, the depth 
information is generally obtained through stereo 
analysis(Marr and Poggio, 1977; Grimson, 1985). It 
indicates that stereo matching will be the major technique 
for developing our system. In addition, for the purpose of 
man-made structure extraction, such as houses and roads, 
from complex aerial imagery in urban scenes, we also 
believe that the feature-based stereo-matching techniques 
are superior to the area-based ones because of the higher 
matching efficiency and accuracy are expected. 
Moreover, we assume that house change has the state of 
either emergence or demolition. It indicates that two 
kinds of house change detection algorithms corresponding 
to emergence and demolition detection will be studied in 
this research. 
According to above considerations, a practical system 
was developed which involves mainly following three 
parts : feature extraction, stereo matching and house 
extraction, as well as house change detection for GIS 
database revision. 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B4. Vienna 1996