International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B5. Istanbul 2004 
distance (m) 
distance (m) 
(b) rms distances of points from their corresponding surfaces 
Figure 2: Two kinds of registration errors of simulated 
point clouds with different levels of noise. c is the stan- 
16 
1.4 
T2 
087; 
0.6 F* 
Qr x 
0.2 f 
  
Tr 1 
sigma 0.00 
sigma 0.03 
sigma 0.06 
* Xx 
Xx wx 
x 
  
  
iteration 
(a) rms distances between corresponding points 
  
  
  
  
0.35 r r y x ; d 
sigma 0.00 
0.3 sigma 0.03 
sigma 0.06 
0.25 | 
0.2 |} 
0.15 + 
0.3.1. 
0.05 | x 
0 À L i 1 
0 5 10 15 20 25 30 35 
iteration 
dard deviation of zero-mean Gaussian noise. 
scale 
scale 
Figure 3: 
of (a). 
40 
  
1.4 
1.3} 
12. 
0.9 | 
08, 
07% 
0.6 | 
  
0.5 
1.005 
1.004 
1.003 }- 
1.002 | 
1.001 + 
0.999 
0.998 | 
0.997 + 
0.996 
0.995 
sigma 0.00 
sigma 0.03 
sigma 0.06 
iteration 
(a) 
x 4 
  
40 
  
T 
T 
  
  
L_ derriere 
5 10 15 20 2 3 35 
iteration 
(b) 
The scale of selected corresponding points in 
each iteration of the registration of simulated point clouds 
with zero-mean Gaussian error. (b) is the magnified figure 
  
(d) 
Figure 4: A Buddha statue scanned by Riegl LMS-Z210. 
(a) and (b) are before the registration. (c) and (d) are after 
the registration. 
real point clouds are much smaller than the point spacings 
of point clouds defined as the average distance from a point 
from its neighbourhood. The registration errors of the two 
real point clouds are the order of centimetre. In the cases 
of the building and trees captured by the Mensi GS200, 
registration is successful as indicated by the registration 
error, e, despite the difference of the point spacings of two 
point clouds being about the order of 10cm and the pres- 
ence of many trees, which hinders the registration of the 
point clouds. 
  
  
  
  
  
  
  
  
  
  
ni k 1 t € di 
n2 (sec) (m) d» 
Cube 2640 40 7 3. 0.000040 0.119 
cg —0.0 4048 0.118 
Cube 2640 40 39 21.0 0.00915 0.119 
cg — 0.01 4048 0.118 
Cube 2640 40 39 21.0 0.0267 0.119 
o = 0.03 4048 0.118 
Cube 2640 40 39 16.0 0.0504 0.119 
o =0.06 4048 0.118 
Ayutlaya | 39268 30. 40. 620 00235 0.043 | 
4393 0.061 
building 139665 lO 49 323.0 0.0388 0.194 
(Qi 217377 0.361 
building 139665 10-1249: 602.0 0.0238 0.194 
(2+3) 325870 0.371 
  
  
Table 2: Results of experiments with simulated and real 
point clouds. m; is the total number of points of point 
cloud C. k and i are the numbers of the neighbourhood 
of a point and total iterations, respectively. ¢ and c are the 
execution time and the registration error. d; is the point 
spacing which is defined as the average distance of a point 
from its neighbourhood. 
4 CONCLUSION 
A method for the registration of two partially overlapping 
point clouds from different locations without good a pri- 
ori alignment was proposed and tested with a simulated 
226 point cloud with different levels of Gaussian noise and two 
   
   
    
   
   
   
   
  
   
    
   
    
  
   
  
  
  
   
   
   
  
   
   
    
  
   
   
   
  
  
     
   
    
   
  
    
   
   
  
  
  
  
     
    
  
   
    
    
   
Internati 
Figure 
(a) and 
tively. 
real poi 
from a | 
error m 
point c 
that of 
deviatic 
Several 
of exec 
ferent r 
distribu 
and the 
properti 
method 
the sele 
corner [ 
ric prin 
they cai 
additior 
indicati 
5 ACI 
This re: 
Council 
thanks t 
REFEF 
Anders 
J., Don, 
ling, S., 
Users' ( 
plied M