plane 
> that 
with 
>M is 
high 
'neral 
is are 
been 
r and 
scale 
ition” 
M to 
ROIs. 
gs in 
es. If 
other 
While 
rable 
alters 
nages 
ise in 
m is 
as are 
imum 
naller 
Xf the 
ystem 
ed in 
Bofang Zhao 
reconstruction process. Extracting shadows in aerial images may be done simply by thresholding in simple situations 
(Huertas and Nevatia, 1988; Irvin and McKeown, 1989). In order to handle complex aerial images, the method 
proposed in (Liow, 1990) has been extended by incorporating shadow intensity information. As illustrated in Figure 2, 
under non-vertical sun illumination a building casts a dark shadow strip. Since in many cases the shadow line opposite 
the building does not always appear, due to the interference of other objects nearby, only the building shadow line is of 
interest in this study. The building shadow line is extracted using three criteria: average gradient magnitude along the 
shadow line, relative orientation between sunlight and gradient direction, and average intensity of the shadow line's 
neighbouring region in the sunlight direction. The Sun's orientation and elevation are automatically derived from 
longitude, latitude and imaging time. 
4. BUILDING EXTRACTION 
The extraction process is carried out in a data-driven, bottom-up approach: from the derivation of primitive features to 
the formation of building structures and building models. The complete extraction and reconstruction process involves 
these steps: primitive feature derivation, building segment extraction, segment matching and plane construction, and 3D 
building modelling, of which the first two steps are to be discussed in this section. As the detected building regions, i.e. 
ROIs, provide focus of attention for the extraction process, all these steps, except the first one, are executed locally on 
each individual detected building regions. 
Corresponding to the employed generic polyhedral model of buildings, Burns’s “straight line algorithm” (Burns et al., 
1986) is applied to generate the primitive linear features. The algorithm extracts linear features by finding regions of 
similar gradient direction. For each extracted feature, a set of attributes is defined and calculated for subsequent 
processing. Due to the complexity in the images stated in Section 1, the primitive features extracted from aerial images 
are usually highly fragmented, especially for those derived from large-scale aerial images in urban areas. Hence they 
need to be grouped. The grouping is based on the perceptual grouping method (Lowe, 1985; Mohan and Nevatia, 1992), 
and is achieved through a series of operations (Trinder and Zhao, 1998). After grouping, those features located within 
individual detected building regions are selected as building segments for the 3D matching and reconstruction process. 
All extracted building segments are stored in a graph, in which segments and their geometric relationship are 
represented as nodes and arcs. Accurate individual building boundaries are constructed by traversing the graph using 
constrained depth-first search to find closed loops in the graph. 
  
Figure 3. Original Images (a) Left Image (b) Right Image 
5. TEST AND RESULTS 
A pair of aerial images of Mildura, New South Wales was used for the tests (Figure 3), each with about 1600 x 1650 
pixels. The image scale is 1:4,000 and the pixel size is 24 x 24 cm” on ground. The area covered by the images is a 
mixture of industrial and residential areas, with many trees, trucks and roads around some of the buildings requiring 
  
  
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000. 1029