(4) 
t kind of 
Im cameras 
In the case 
ransfer and 
he following 
lefined with 
the point of 
s where the 
ical axis. In 
position the 
parameters 
related with 
of interior 
tment it is 
tion to the 
measured 
'esiduals of 
n becomes 
  
lage 
  
ortion 
Figure 5 and Figure 6 show the original video image with 
large radial distortion values and the corrected image 
after self-calibration. 
3.3 Testfield Calibration 
The CCD-camera has been calibrated by 13 convergent 
images of the testfield. Due to a large shift of the principle 
point and extreme values for radial distortion (see Figure 
7), the observations (image coordinates) have to be 
corrected iteratively by the calibrated parameters of 
interior orientation. With the corrected observations the 
adjustment process has been started again until the 
system reaches convergency. In comparison to the first 
step of calibration the modified parameters lead to 
variations in image space of about 20pm. This effect is 
much poorer for lenses with larger focal length and less 
distortion. 
CCD camera 1/2", f=4.8mm 
100 7 
50 4 
  
0,5 1 15 2 
-50 4 
image radius r' [mm] 
radial distortion dr' [ym] 
-100 4 
  
  
-150 - 
Figure 7: Radial distortion 
The | measurement accuracy after projective 
transformation of a flat (AZ < 5mm) surface (size 1.5m x 
4m) has been determined to a standard deviation of 
sxy=t1.2mm using the calibrated CCD camera. This 
result is equal to a relative accuracy of about 1:4000 and 
meets the requirements. 
4 IMAGE MEASUREMENT 
4.2 Point Measurement 
The measurement of circular targets with high sub-pixel 
accuracy has been solved for more then ten years now 
(e.g. Luhmann 1986). The algorithm used in this 
application is based on a course point detection by a blob 
analysis algorithm which searches for point-wise patterns 
In the image. In a second step these approximate 
positions are used for a precise point measurement 
where the point center is calculated from a best-fit ellipse 
of the extracted contour points. The average point 
accuracy of well-defined targets is better than 3/100 of 
the pixel size. 
4.3 Edge Measurement 
The actual processing task is the extraction and 
Measurement of non-signalized object edges. In the case 
of precast concrete ceilings the edges are disturbed by 
production failures (e.g. overflow of concrete), changing 
light conditions and different observation angles with 
respect to the camera (example in Figure 8). In order to 
improve the edge detection process a CAD data set of 
each part is transformed into approximate edge position 
values in image space. 
  
Figure 8: Edges on concrete part 
Edge detection is performed by a self-adaptive line 
following process. Given a starting point and direction 
(from CAD data) the operator searches for the most likely 
edge in the neighbourhood of the current position. Using 
the n previous edge points the next edge candidate is 
predicted. For each successful edge location a set of 
edge attributes (e.g. edge width, curvature, contrast etc) 
is stored. The n previously determined attributes are used 
for a learning scheme in order to select the desired edge 
point out of many possible edge candidates. 
  
o predicted edge position 
& precise edge position 
A: starting point d: offset 
p;: image profile 
s: step width 
Figure 9: Line following 
According to Figure 9 the operator adapts to the current 
line curvature. In the case that a profile offset d exceeds a 
given limit the step width s is automatically reduced in 
order to generate a denser point distribution for higher 
edge curvatures.