nds is 
as the 
as the 
ermed 
racted 
, The 
which 
or the 
these 
ed by 
arallel 
| with 
> then 
rizon- 
region 
is of 
itude, 
[ntosh 
), also 
re not 
oid of 
; form 
imizes 
e per- 
d and 
square 
ie. dis- 
(o, 9) 
s fora 
) with 
ths up 
r very 
noise. 
n pro- 
. The 
ength. 
deter- 
d line 
; used 
higher 
ensely 
» dual 
m the 
> ana- 
image 
ed of 
Tr any 
olume 
same 
d fea- 
1s dis- 
'acted 
  
  
  
  
  
5000 hlens Ift 2750 3250 
4000 
3000- 
20007 
10007 
  
  
  
  
0 5 10 
pixel units 
Figure 2: Frequency histogram of line lengths for typical 
subimage. 
from two-dimensional images. Any increase in cultural ac- 
tivity is expressed by more features per unit area. This 
selection of scale was useful and a general trend was found 
which was used as a basis for adjusting threshold param- 
eters. 
It was found from graphs such as Figure 3 that the value 
on the y-axis of an z-value > 50 pixels is approximately 
the same as that of the next line above for an z-value 
< 50 pixels. That is, lines of shorter length may be in- 
cluded when dealing with darker subimages to produce a 
similar number of total lines with respect to the brighter 
subimages. 
4.5 Remaining Line Attributes 
All the line attributes to be extracted are: 1) orientation, 
2) location, 3) length, 4) width, 5) contrast, 6) steepness, 
7) dark edge and 8) light edge. A straightness attribute 
has been derived but has not been implemented during 
the matching stage. 
The width being defined to be the area of the line sup- 
port region divided by the length of the line. The light 
edge and dark edge values are the average of the light- 
est and darkest 10% of the pixels in the region, respec- 
tively. Contrast is the difference between the light and 
dark edge values (McIntosh and Mutch, 1988). Steepness 
of the intensity surface of the edge was the average gradi- 
ent magnitude of the region in McIntosh and Mutch, but 
was evaluated here as the contrast divided by the width. 
An additional parameter describing the shape was used 
as extra information for the matching. The straightness 
parameter was estimated by determining the longest line 
contained in the region and ratioing the distance of the 
centroid from this line over the length of the line. 
4.6 Linking Lines 
An optional step at this point is to link lines with similar 
attributes and a close pair of endpoints. In some cases 
Average intensity versus Number of lines on Sqrt Scale left hlens 
3000 
4-5 
1000— o 
7 
500 
7-8 
Je 
eee dg -10 
10074 fo 
Hi 
50] 14-15 
195 SZ 
14 
Tr ET TTY ete 
911 
  
  
  
  
  
Figure 3: Average intensity versus number of lines pro- 
duced for an East-West strip of left subimages. Note ver- 
tical axis has a square root scale. Each line represents the 
number of lines produced for a particular range of lengths. 
this is important as a single long line in one image may 
correspond to two line segments in the other image and 
an opportunity for a match may be lost. However, the 
aim of the project is to obtain as many precisely located 
features as possible and over-linking of lines might reduce 
the density of points in the network after matching has 
been performed. In (Zhou et al., 1989), lines are linked 
based on certain criteria and some are extended to meet 
at corners. 
Line linking has been implemented, if endpoints from two 
lines are within 5 pixels, and if after the lines have been 
mapped into points in (p, 8) space using the Hough Trans- 
form their proximity is less than 2 units and the remaining 
parameters are similar then the lines were linked. 
4.7 Filtering Lines 
Lines are filtered as a final step in the extraction pro- 
cess. Lines with very short lengths may represent noise 
or texture and ignoring these lines allows removal of noise 
without resorting to preprocessing the image using digi- 
tal image processing techniques. Short lines derived from 
regions of small area will have attributed to them a set of 
unreliable parameters. Width, steepness, contrast etc are 
all more representative the longer the line and the larger 
the region from which they were obtained. 
Also, lines with near horizontal orientations are not good 
for matching as they are nearly parallel to the epipolar 
plane. If different portions of the same horizontal linear 
feature are extracted from each image of the stereo pair, 
then all parameters may be similar and the epipolar con- 
straint satisfied but an incorrect height obtained. 
Figures 4 and 5 show the result of extracting lines from