type as 1 because two orientation vectors define a cylinder. It 
is obvious that fewer objects have to be used to register scans 
when cylinders are being used for the registration. 
4. TEST RESULTS 
Firstly results are shown from the used fiting algorithms. 
Cylinder fitting is still in the stage of finding approximate 
values for its parameters. Still the parameters found are a very 
good approximation of the cylinder. This can be accounted for 
by taking into account the huge number of points present on 
almost all of the cylinders modelled. Laser scanners, 
nowadays, deliver data sets with a point density of one point 
every 5cm, which results in a large amount of points on the 
objects to be modelled. Figure 4 shows a point cloud of a 
cylinder with 2543 points. Points are distributed nicely on both 
sides of the fitted surface of the fitted cylinder as seen in 
Figure 5. Tests have been conducted with the same point cloud 
reduced by a factor 20 (170 points left). The algorithm was 
still able to find the parameters but with a lower accuracy. 
  
Figure 4. 2543 Points on a cylinder 
  
Figure 5. The calculated cylinder (white) together with the 
original laserpoints 
Processing time to find the parameters to find the cylinder 
consisting of 2543 points shown in Figure 4 is 0.7s on a 
pentiumIV 
Alhtough the fitting algorithm is not yet finished for cylinders 
the initial values are of sufficient accuracy that they can be 
used as initial values for the iterative least squares adjustment. 
The fitting of the plane is finished. The algorithm's output 
consists of a normal vector and a distance as described section 
2.1.1. Figure 6 shows an example of the derived normal vector 
multiplied by the perpendicular distance of the plane from the 
origin. 
  
Figure 6. The white line shows the vector ln of the vertical 
plane located on the left 
The test to register two scans described in this section is 
performed on laser data from one scan. The points in scan one 
are a selection of the total scan and the points in scan two are 
other points from the same scan. Furthermore the points in 
scan two have been rotated relative to the points in scan one 
with angles specified by the user. The disadvantage of creating 
two scans from one scan is that the scans perfectly math using 
the right transformation parameters. However, when 
registering the images with objects measured in different scans 
this only happens if the same points were used in the object 
fitting stage. This is not the case in the registration described 
in this section, as the objects used for registration are only 
partly visible in either scan. The results will therefore reflect 
the results that can be achieved using real data. 
Figure 7 and 8 show the two laser scans that are registered, 
both from their own scan angle. 
  
Figure 7. Scan 1 
  
Figure 8. Scan 2 
To do the registration three corresponding planes were 
measured in both scans. Table 1 shows the deflections from 
the values defined by the user and the transformation values 
found and the final result is shown in Figure 9. The 
transformation between the two systems was sufficiently small 
—16— 
ess  — = o f éa im = - = LL 
a 
  
Eine 
uma alti EEE 
  
FE e; ooo mo vua 
  
  
  
  
Th 
fac 
difi 
froi 
assi 
resi 
diff 
Scal 
To 
con 
achi 
mea 
poir 
para