average elevation). Compared with the original Moravec 
method this approach significantly reduces the time needed 
for generating the Moravec image. As a result a discrete 
Moravec image is derived instead of a continuous image. 
In every sub-area of both images only one interest point is 
selected, the one with the maximum value of the interest 
operator. The implication of this step is that there is no need 
to set a threshold for the value of the interest operator, as is 
done in the second step of the original Moravec method. At 
this stage it is assumed that an interest point selected in a sub- 
area of the left image should match the interest point similarly 
derived in a corresponding sub-area of the right image. 
Investigations have shown that about 60 - 70% interest points 
are matched correctly. 
In order to assess whether a pair of interest points will match, 
a test of similarity has been introduced. For every interest 
point selected in both images, three parameters are calculated 
which characterise the intensity patterns around them. These 
parameters are the ratios between mean gradient slopes in the 
intensity values derived for the four principal directions for a 
particular interest point. The mathematical formulas for the 
similarity test are as follows. 
Assume a window.sized 9 x 9 pixel is placed on an interest 
point. Let the intensity gradients of neighbouring pixels 
within the four principal directions (d- 0,..,3) 
8711-1, (3) 
where: = 0°12 
and I is the intensity value. 
The mean gradient in intensity values is defined by: 
4) 
S4-E(g 
d = 0,...,3, where E(.) is the operator of mean value. 
The ratio of mean gradients is therefore defined by: 
Rz 2 (5) 
  
sd 
where d z 0,1. 
The rejection threshold applied for the measure of similarity 
of interest points on 2 corresponding windows is: 
(6) 
n£ 
  
where n is the number of interest points and RL, RR are 
the corresponding ratios for the pair of interest points being 
matched in the left and right images. 
If for a pair of interest points the average difference in ratios 
is greater then R + 2.0* Gp then matching of the pair is 
considered as incorrect. In the sub-areas of the right image, 
for which the matching is incorrect, a further set of four 
points is selected. These are chosen according to descending 
values of the interest operator for pixels in those sub-areas. 
On this set of additional interest points the similarity test is 
also performed. Matched points are selected accordingly to 
the smallest difference from the average of the ratios of the 
gradients. 
The evaluation of performance of these similarity tests of 
interest points have hsown that the number of correctly 
matched interest points has increased to a level of 90 - 95%. 
The advantage of this similarity test is that the ratio of the 
mean gradients are calculated based on previously derived 
972 
values of gradients in the intensity value, hence eliminating 
recomputation. 
Linear feature based matching will also represent a component 
of this software, creating a lattice of matched features over the 
image. The techniques to be used for this process, which has 
not yet been incorporated into the software, will be described 
by Butler (1992). 
3.2 AREA BASED MATCHING. 
Grey level or intensity based matching is carried out by the 
least squares method, based on 6 affine transformation 
parameters, as has been used by many photogrammetrists to 
achieve high precision matching e.g. Gruen and Baltsavias 
(1985), Rosenholm (1987). 
The least squares matching method has the advantage that by 
its very nature, distortions in geometry can be corrected 
through the resampling process, and in addition, it provides 
information on the quality of the match through the weighted 
sum of squares of the residuals at the pixels. Other forms of 
error detection available in the least squares method can also 
be used. Accuracies of the method are typically 0.3 pixel, 
Gruen and Baltsavias (1985), Rosenholm (1987). 
In this project the least squares matching is performed on a 
window of 21 x 21 pixels. The thresholds for the shift and 
rotation terms in the solution are set at 0.1. Average number 
of iterations needed to satisfy the threshold conditions is 3 to 
4. Of course, any matching computation which does not 
converge in the prescribed number of iterations is discarded. 
An essential element of this computation are the procedures 
for checking the accuracy of the matching and the computed 
elevations. This is divided into 3 stages: 
* checks in the image space. The rotation parameters in 
the least squares solution should be similar within 
certain regions of the overlap area of two images, and 
variations in these parameters will be largely due to the 
effects of parallaxes caused by variations in terrain 
elevations. The standard deviations of the rotation 
parameters derived from the matches within each region 
are compiled during the computation. Any parameter 
which deviates from the mean by more than a certain 
factor times the standard deviation will be discarded as 
an erroneous match. Further, a y-parallax between the 
computed match points greater than 2 pixels will cause 
the point to be discarded. 
* those which check the distribution of the elevations. 
The RMS variations in the elevations within a certain 
region will indicate the characteristic shape of the 
terrain. Therefore, outstanding elevations in a certain 
region of the terrain will be identified as those which are 
greater than the average terrain shape, by a set factor of 
the RMS variations. These points will normally be 
discarded unless further tests indicate that the terrain 
variations are indeed due to a marked change in 
elevation. In addition, when a grid of points is being 
computed, the distance between points is also used as a 
check parameter. 
* visual observations on the. computed data. This test 
could involve viewing the stereomodel in a digital 
workstation where available, but at present, this test 
will be based on the observation of overlaid ortho- 
images. 
4. PERFORMANCE OF THE SOFTWARE 
There are many parameters to be taken into account in 
assessing the performance of the package. 
° the user interface; 
Rega 
point 
use. 
borne 
exam 
draw: 
anotl 
.netwe 
been 
The r 
least : 
pixels 
000 d 
been 
to cor 
hours 
HP73 
It has 
infori 
exam] 
bytes 
suppn 
loops 
of the 
files s 
windo 
block 
can im 
Initial 
compi 
the Wi 
Softw:z 
Observ 
ratio o 
of 1:4, 
did no 
at an i 
elevat 
manua 
points. 
in inte: 
gradua 
A total 
the cor 
observ 
was the 
RMS) 
of disc 
compu 
of 1 m. 
of the « 
the ma 
of 5m, 
7m. FF 
current 
those p 
indicate 
terrain 
for a gi 
The ste 
differei 
probler 
environ 
result i 
dark. A 
column 
these pr 
robust. | 
correct | 
Method