Figure 5. One of the corresponding polygons in the map and 
the image and the matched points. 
5. REGISTRATION OF BUILDINGS ON DDS DATA 
TO 1:10 000 MAP 
A further test was carried out on the extraction of large 
buildings from satellite data with a 2.5m pixel size. In this 
work the buildings were registered with a 1:10 000 scale 
raster map. The work is described in detail by Vohra and 
Dowman (1996). 
The preparation of the map involves considerably more effort 
than in the two examples discussed above, as a lot of 
unwanted information has to be removed. Buildings are first 
identified manually and then assigned a unique DN value. 
Objects below a given size are removed and the edges of the 
remaining objects are extracted. In order to carry out the 
polygon patching the edge strength and direction are needed 
and these are also extracted at this stage. The image is 
prepared in the way described in 2.1 above. 
The matching of the polygons produced 923 matched points 
and root mean square residual after transformation with an 
affine transformation of 2.0 pixel. Inspection of the residuals 
clearly showed the effect of perspective distortion of the 
buildings and a arbitrary figure of 0.5 pixel was adopted as 
the acceptable limit. This left 64 points with a rmse of 0.3 
pixel. 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996 
Once again the area is flat and only a two dimensional set of 
points are determined. At this scale the perspective view of 
the buildings has a significant effect and the residuals on the 
matched points show systematic displacements which are 
related to the camera look angle and the orientation of the 
buildings. Future work will look for methods whereby the 
perspective effects can be identified and removed using a 
suitable model of the distortion. Shadows also have an effect 
which must be removed. 
6. CONCLUSIONS 
The three examples described illustrate a number of important 
component of the automated registration system. There is 
however still a considerable amount of work to be done to 
remove manual operations, refine the algorithms and 
introduce new algorithms to allow other data to be used. The 
results show that the core algorithm, the polygon edge 
matching, is robust and flexible. It allows matching to take 
place despite distortion and differences in the actual 
polygons. The results allow an analysis to be made which 
indicates that errors due to relief and perspective can be 
modelled and corrected for. They also show that a large 
number of conjugate points can be extracted and provided 
that the strategy used ensures that there is a good distribution 
of points, sufficient points should be obtained to allow three 
144 
    
   
  
  
  
  
  
  
  
   
  
  
  
  
  
  
  
  
  
  
  
  
  
  
   
   
  
  
  
   
   
  
   
   
  
   
   
  
   
  
  
  
  
  
  
   
  
  
  
   
  
  
   
   
  
   
    
dimensiona 
registration. 
The metho 
Union 4th | 
London, Tt 
of Stuttgai 
Corporation 
will concen! 
automated t 
new algoritl 
Abbasi-Dez 
st 
A 
3( 
Gülch E, 19 
Pl 
H 
Holmes A, 
cl 
M 
Maitre H, V 
fc 
Ol 
A 
Morgado A 
D 
2 
Newton, W 
a] 
al 
Ir 
R 
Vohra V ar 
b 
Ie 
P