on of roof 
he useful 
W. In this 
proposed 
ment, 3D 
pints from 
ec matched 
nages. To 
of points, 
hnique is 
tly, edges 
ing to the 
surface is 
will be 
xtraction, 
me testes 
[TON 
building 
d and the 
1g simple 
ne use of 
buildings 
relies on 
be data- 
strategy, 
features 
tures are 
g images 
tures. À 
tegy and 
in Mayer 
traints. It 
1 number 
cet has a 
ines such 
y will be 
. It starts 
imagery 
  
  
and the features extracted from different images are matched to 
create 3D features. 3D features are then grouped to form roof 
primitives — facets. Finally they are grouped together to create a 
complete roof structure. 
3. AUTOMATIC EXTRACTION OF BUILDINGS 
3.1 Feature Extraction 
Based on the building roof model, two types of features, i.e. 
points and edges, are extracted from both vertical and oblique 
images for reconstruction of buildings. There are numerous 
operators for extraction of points and edges and each has its 
advantages and disadvantages. In this study, a simple point 
operator - the modified Moravec operator is used to extract 
point features for the efficiency of processing while edges are 
extracted by using the Canny operator. The Canny operator is 
famous edge detector and it has good localization accuracy and 
less sensitive to image noise. An example of feature extraction 
by using Moravec and Canny operators is shown in Figure 1. 
As can be seen, most features are extracted correctly, but some 
false features are detected as well. These false features should 
be removed before image matching in order to derive reliable 
3D features. 
  
Figure 1. Extracted points (green) and edges (red) 
3.1.1 Preprocessing of Features One common problem 
of the point operators is that they cannot extract all the feature 
points from the image due to the complexity of aerial images 
and at the same time some of the extracted feature points may 
not be true features. Thus, false feature points should be 
eliminated and missing point features should be added before 
generating 3D features. In addition, the extracted edge pixels 
should be traced to form edge segments. A roof corner is the 
intersection of two or more than two roof lines. Most false point 
features can be removed by applying this constraint to the 
extracted feature points. The extracted edge pixels are traced 
based on their proximity and similarity in edge orientation. 
Smooth and straight edge segments are generated by using a 
split-and-merge operation to the traced edge segments. Most 
noisy edge segments can be removed by this operation. 
To find missing point features, close line pair segments are first 
selected and their spectral properties in their flanking areas are 
then compared. They belong to the same roof facet and their 
intersection is computed if they have similar spectral property. 
  
  
    
   
  
    
  
   
  
  
  
  
  
  
   
  
  
  
  
   
  
  
   
   
  
  
  
  
  
  
   
  
  
  
  
  
  
  
  
  
  
  
  
  
   
   
  
  
  
   
   
   
  
  
    
   
    
  
   
   
     
If the computed intersection is not in the list of extracted point 
features, it is then added to the list. 
3.1.2 Removal of Gutter’s Effect Gutter is an important 
feature of a building roof. It is visible in a high-resolution aerial 
image and has two parallel edge segments with approximate 
equal length in the image, as shown in Figure 2. The existence 
of double edges at roof gutters may yield mismatch points and 
thus reduce the reliability of roof reconstruction. Therefore, 
only one edge should be used in the reconstruction. In order to 
keep the accuracy of computed roof area and pitch, the inner 
line should be used and the outer one needs to be removed. To 
eliminate the outer edge of gutters, all parallel edges with 
opposite edge direction near roof boundary are examined and 
only those with the distance between two parallel edges within 
a given threshold are selected. Their distances to the center of 
the building are then computed and the edge with larger 
distance is removed. 
  
(a) Image edges before removal of gutters 
  
(b) Image edges after removal of roof gutters 
Figure 2. Removal of roof gutters 
3.1.3 Computation of Feature Properties and relations 
Both feature properties and their relations play an important 
role in feature matching and grouping of 3D feature points. 
Once roof corner and ridge points are extracted, their properties 
are computed for matching to create 3D features. The properties