Figure 4: Extracted lines from left subimage of Sydney 
Airport region. 
SPOT subimages over the Sydney Airport region. The 
average intensity of one subimage was approximately 46. 
Normally the images with lower average intensities are 
due to densely vegetated, relatively featureless areas, 
however, this image has low average intensity due to the 
presence of a large body of water. As a result, this image 
is considered relatively dark and so the gradient magni- 
tude threshold, minimum length and region-area thresh- 
olds are kept low to allow more lines to be extracted. The 
actually threshold values used were: 
Gradient Magnitude = 6 grey scale/pizel (1) 
Length Minimum = 5 pizels (2) 
Area Minimum = 5 square pizels (3) 
The process of line extraction is performed on both im- 
ages of each stereo pair. All line attributes from the left 
subimage are stored in one database file, and all from the 
right subimage are stored in a second file. 
Certain lines from the left file are compared with certain 
lines from the right file depending upon various geomet- 
rical constraints and line pairs considered to match are 
determined, as explained in the following section. 
5 Image Matching 
Matching is generally recognized as the most difficult 
stage in stereo vision. Once straight lines are extracted 
from both images and attributes determined (eg location, 
orientation, length, contrast, width and straightness), a 
match function is defined which determines the strength 
of the similarity between a pair of lines. The location of 
912 
  
Figure 5: Extracted lines from right subimage of Sydney 
Airport region. 
straight lines within an image may be determined to sub- 
pixel accuracy. Pairs which are a mutually best match, 
based on the match function and correspondence con- 
straints (epipolar, continuity) are considered correspond- 
ing lines (McIntosh and Mutch, 1988). 
With the continuing advance in computer memory and 
speed it is feasible to produce thousands of lines per 
subimage with arbitrarily strict filters on line quality (eg 
threshold on line length) during the extraction stage and 
arbitrarily strict matching criterion on match quality (eg 
large match function threshold). Due to the large num- 
ber of features examined, a modified implementation of 
the matching algorithm in (McIntosh and Mutch, 1988). 
was required. 
McIntosh and Mutch’s algorithm for matching straight 
lines examines images of indoor scenes and industrial ob- 
jects. This paper also uses Burns’ method for the initial 
stage of extracting straight lines. 
Features in close range environments differ from the dis- 
tant range application under consideration with regard 
to the geometry and distortions in the satellite images. 
Also, each linear feature is more important and there is 
a greater possibility for occlusion of surfaces in the close 
range. 
Nevertheless, the general matching principle of McIntosh 
and Mutch’s algorithm has been adapted and has been 
found quite satisfactory. The matching algorithm finds 
corresponding pairs of lines on a pair of images based 
on a match function. This function combines a set of 
eight descriptive parameters of each line, which may be 
weighted according to their relative importance. 
Corresponding lines are those which have the largest