7A-1-2 
as length, orientation, closeness, etc. Much prior work has been 
done on line extraction. For example, Boldt used the geometric 
symbols grouping criteria and a hierarchical concept in his line 
extraction algorithm (Boldt et al., 1989). Bums addressed the line 
extraction problem by using the global organization of the 
supporting line context (Bums et al., 1986). In this research, the 
existing tool based on Bums’ algorithm has been using for 
extracting straight line features. However only the intersections 
between straight lines and epipolar lines are employed in 
matching process. Fig. 1 illustrates the intersection between an 
epipolar line and extracted lines overlaid on the original images. 
Figure 1: The intersections of extracted lines and 
the epipolar lines 
• Use these attributes to refine (split and merge) clusters. All 
refined clusters are called plateaus. The end points of 
plateaus will be recorded along with their attributes. 
Some useful properties of the plateau end points are recorded for 
the feature matching process. Fig. 2 shows some examples of 
plateau features extracted from the intensity profiles in images of 
the Purdue University campus. 
Spike Extraction 
The spike is another one-dimensional feature utilized in this 
research. It occurs predominately in images of an urban area. The 
spike feature is described as a very short plateau, i.e., sharp peaks 
or valleys, in the homogeneous region of an intensity profile. 
They are extracted from the intensity profile in the epipolar space 
as follows: 
• For each pixel i along an intensity profile, compute the 
average grayscale G(x t ) for the left neighbors and the right 
neighbors (see Fig. 3). 
G(x a )=- ZG(x k ) (4) 
n k=i-p-n 
Figure 2: Step/Plateau Extraction 
Plateau Extraction 
The plateau feature is defined here as a ID region with near 
constant grayscale. It is particularly suited for built up areas or 
highly urban areas. In manual DEM collection, an operator 
would collect at least the end points of this feature, possibly also 
intermediate points inside a plateau. The automatic plateau end 
point extraction method is performed on an intensity profile along 
an epipolar line as follows: 
• Search on an intensity profile for 
\G(x)-G(x-\))\< Threshold (1) 
Where G(x) is the grayscale of pixel x. If the condition is 
fulfilled the pixels will be recorded. Any contiguous pixels 
meeting the criterion is called a cluster. 
• Find attribues of clusters 
AVGPL= 
G(x) 
st end - st 
(2) 
STDPL= 
\_ 
e ^(G(x)-AVGPL) 2 2 
s , end - st 
(3) 
1 i+p+n 
G(x ir )=~ £G(x k ) (5) 
n k=i+ p 
• Determine the average deviation (adm) of the left neighbors 
and the right neighbors of pixel i. 
adm i=——; £ \G(x k -G(xa)i (6) 
n * k=i-p-n 
adm r =—i—- X \G(x k -G(x ir ))\ (7) 
n 1 k=i+ p 
, Where p is the expected width of a spike. And n is the 
required distance of smooth neighbors. 
j i i 
Where st and end are a starting point and an ending point of Figure 3: Spike feature in an intensity profile 
a cluster respectively.