| of 
rto 
na 
car 
Ory 
Ire- 
ilot 
Ihe 
av- 
)ec- 
ere 
and 
Tas 
for 
7 of 
ia a 
be). 
The image format was set to 726 x 568 pixel with a pixel 
spacing of 11 x 11 um. 
Z-axis 2 
e 
e, 
CCD-cameras 
3 
  
       
  
   
   
Y-axis M 
Turntable 
  
Front view 
CCD-cameras 
2 1 
Turntable E E 
Y-axis 
Plane view 
Figure 1 Measurement configuration of pilot test. 
3 PRE-CALIBRATION 
Precise values for the elements of the interior orientation, 
the parameters of the pixel-to-image coordinate transfor- 
mation, and the additional parameters were determined 
before the test via a system calibration using a three-di- 
mensional testfield. The accuracy requirements of the 
test correspond to approximately 1/ 10' of the pixel spac- 
ing. The calibration of the cameras was thus not designed 
for utmost accuracy but to fulfil the requirements. 
Four cameras were calibrated with their lenses focused at 
infinity. Initial values for the interior orientation and the 
parameters of the pixel-to-image coordinate transforma- 
tion were derived from the data sheets of camera and 
frame grabber. These values were improved via the fol- 
lowing additional parameters: 
* location of the principal point in x and y 
* camera constant 
* scale factor in x 
* first two parameters for radial symmetric distortion 
The pixel-to-image coordinate transformation is defined 
as (Beyer, 1992): 
x = (x -Xy) Xpsx (1) 
ye (yyry)xpsy (2) 
    
     
    
   
    
   
   
    
   
  
   
   
    
    
   
  
    
  
    
   
   
   
   
     
   
   
     
   
    
    
  
   
  
    
with: 
XT cevercoantrice image coordinates 
XV erer-ezetter pixel coordinates 
X p: y p location of principal point in pixel 
PSX, PSY ceversoer pixel spacing in x and y 
A testfield spanning 4.2 x 2.0 x 1.1 m? with 186 targets 
was used (see Figure 2). Latter were black circles with a 
diameter of 20 mm on white background. Images were 
acquired with 0 and 90 degree roll at each of a 6 station 
network. The image coordinates were measured with 
Least Squares Matching (LSM, Gruen, 1985). The image 
scale varies from 1:460 to 1:760. The targets are thus im- 
aged onto 2.4 to 4.0 pixel in diameter. The inner preci- 
sion for these targets is 0.02 to better than 0.01 pixel. The 
ideal diameter of these targets in the imagery was deter- 
mined to be at least 6 pixels (Beyer, 1992). 
  
Figure2  Testfield used in pre-calibration of CCD- 
cameras. 
The pre-calibration was performed with separate bundle 
adjustments for each of the four cameras. Thirty control 
and 124 check points on the wall of the testfield (the 
front structure was displaced shortly before the test and 
no reference coordinates were available) were used. The 
accuracy in image space is between 1/1 09 and 1/20™ of 
the pixel spacing. In a re-evaluation of the imagery all 
image coordinates were re-measured and the data was in- 
troduced into one combined bundle adjustment. The ge- 
ometry is very strong with every object point imaged in 9 
to 48 images with 33.5 images on average. The accuracy 
in object space is (after a spatial similarity transforma- 
tion) 0.22, 0.10 and 0.56 mm in X, Y and Z respectively 
(X and Y within the plane of the wall, Z towards the 
cameras). The accuracy in image space corresponds to 1/ 
30 of the pixel spacing. The accuracy is severely de- 
graded by local illumination gradients across the targets 
due to shadows. Figure 4 shows a plot of the differences 
to reference coordinates for the points on the wall of the 
testfield. The horizontal displacement for points in a 
number of columns is apparent. Figure 2 b shows a part