7 A-1-3 
'xy; 
Figure 4: Spike point (o) detection 
A point is considered as a spike candidate if admi and 
adm r are less than a threshold value, and if the grayscale 
value of the point is considerably different from 
G(x u ) andG(x ir ). 
Only the candidate with the highest grayscale difference 
from its neighbors will represent a spike of its region. 
strong geometric features as straight lines, circles, comers, etc., 
but it is still possible to extract edges. Therefore, in this research, 
edges are the most useful feature for rural scenes. We need the 
descriptive information from edge points to use in the matching 
process. Strong edges and low noise are preferred. Actually any 
well-defined detectors based on the first derivative can satisfy 
these objectives (Roberts, 1965), (Prewitt, 1970), (Davis, 1975). 
Both gradient magnitude and direction which are very useful 
information for matching can be obtained from these gradient 
operators. Edges are extracted from 2D images. But only the 
intersected points with epipolar lines are used in matching 
process. 
3 Feature Matching 
In order to convert the search problem into one dimension, the 
images will be resampled into epipolar geometry. For any feature 
found in one image, it is guaranteed that we need only look along 
a single corresponding line in the other image in order to find a 
match (discounting occlusions). With known parameters of 
interior and exterior orientation (or at least relative orientation) 
the epipolar resampling can be accomplished by generating two 
fictitious pictures with parallel view directions. 
MtslhitMj Besait 
m 
! 
'V 
i «iti 
Figure 5: Result from feature matching 
In an image, spike points represent both point features such as 
manhole covers and 2D features such as edges and straight lines. 
Based on this interpretation, it is possible that some spike points 
may be redundant with other features in a profile. The monotonic 
constraint and the optimization policy of dynamic programming 
in the feature matching process should take care of this situation. 
Fig.4 illustrates some typical spike points in real intensity profiles 
from the left and the right images. 
■■■ ■■ mmmm ■■ 
■■■■ ■»■■■■■■■ 
■■■ mmmmmmm ■■■ 
axiiiniiauit 
■natnaiii) 
■«■■■■■■■a a 
aaaaaaa aaa 
■■aaaaafi aa mm 
me a aaaa mmmmm 
aaaaaaa aaaaa 
a aaa afiraaa 
The inevitable occlusions in large scale photographs lead to 
requirement for robustness in the matching process, and further 
argue for a global rather than a local cost function for the match 
problem. In this matching application, the feature points on an 
epipolar line of one image represent stages, whereas the feature 
points of an epipolar line of the other image represent decision 
variables for each stage. The descriptions of features are 
employed to determine a cost function (Eq.8). 
f(x n ) = min 
d„ 
c n (x n -i,dn)+ min (c l (x 0 ,d l ) + ‘ 
= min [c„ (*„_!, ) + /(*„_!)] 
d„ 
(8) 
111 ■ II 
Figure 6: An elevation profile on cost matrix 
Edge Detection 
Edge segments are commonly used as primitive features in stereo 
matching. In the images of rural areas, it may be difficult to find 
An ordering constraint of the matching process must be enforced. 
It dictates that the optimum path chosen must be a monotonically 
decreasing function. 
The solutions may be different if the features from the right image 
represent stages and the features from the left image represent 
decision variables, instead of the reverse. This situation may 
occur when the set of extracted features on the left image is 
different from the set extracted from the right image. To handle 
this situation, a revised strategy is introduced. The optimal paths 
will be determined twice, considering each image of the pair as 
the reference. The final solution would then be one that was 
consistent with each of these provisional solutions.