164 
with a diameter of 8 mm. Some of these targets were observed geodetically to provide the scale of the 
photogrammetric block. In addition, check points on some of the object curves were measured, too. 
In order to get digital images, the photographs were scanned with a resolution of 15 um using a Zeiss PhotoScan 
PS1. Due to the large image format we obtained digital images with 50 MB each. 
3. IMAGE ORIENTATION 
For inner orientation, the fiducial marks were matched using an area based algorithm. A synthetical template is 
moved in the search image, and the cross correlation coefficient of the grey levels is calculated at each position. The 
position of the fiducial mark corresponds to the position of the maximum of the cross correlation function. Subpixel 
estimation is performed by a polynomial approximation of that function and by determination of the position of the 
maximum of the polynomial. Thus, the fiducial marks could be located with an accuracy of +1/3 pixels (Rottensteiner, 
1993). 
The measurement of the targets was done using a "digital mono comparator". The targets are interactively identified 
on screen and their final position determined by calculating the weighted centres of gravity of the gray levels within a 
small window (Fórstner, 1986). The accuracy of the position of the targets was estimated to be about £1/2 pixels. 
In addition to the targets, 10 non-targetted points distributed symmetrically on the car had to be identified interactively 
in the images, thus enabling the definition of an object coordinate system whose xz-plane is identical to the symmetry 
plane of the car body. 
The geodetical observations together with the image coordinates of both the targetted and the non-targetted points 
provided the input for bundle block adjustment. Altogether 3491 observations were used to determine 1347 
unknowns; 90 observations were eliminated as gross errors by a robust estimation technique. The r.m.s. errors a 
posteriori of the targets and the non-targetted points were estimated by +6.5 um and +27 um, respectively. This 
corresponds to an average 3-D r.m.s. error of 0.5 mm of the object points, the greatest error being 1.1 mm. The 
accuracy of the adjustment was mainly limited by the uncertainity of the symmetry assumptions. As a result of bundle 
block adjustment, the orientation parameters of the images were obtained. 
4. EDGE DETECTION 
Many smooth complex curves were visible on the car which had to be extracted from the digital images using some 
edge detection algorithm. First the images are smoothed using an edge preserving filter (Abramson and 
Schowengerdt, 1993). The smoothed images are convolved with a modified LoG kernel of 5x5 pixels. Candidates for 
edge points are found at zero-crossings of the convolved images. Our algorithm delivers these candidates with a 
resolution of 0.5 pixels; it also provides a thinning of the candidates in order to avoid thick edges. In a second step 
  
     
  
   
e ur 
~~ extracted edge 
4 aia node of adjusted 
curve 
Figure 4: representation of the extracted edge by 
adjusted nodes (right back door) 
neighbouring candidates are connected to form 
segments which should be as long as possible. In this 
step, gaps which are smaller than 2 pixels are bridged. 
Figure 3: results of edge detection in one of the images. The Finally, edge segments which are likely to belong to 
regions outside the car body have been masked. the same object line are joined together (Kerschner, 
1995). In order to reduce the computational effort 
required for edge detection, irrelevant regions of the images were masked and only pixels in "interesting" regions 
were used. The results of edge detection in one of the images can be seen in figure 3. 
  
IAPRS, Vol. 30, Part 5W1, ISPRS Intercommission Workshop "From Pixels to Sequences*, Zurich, March 22-24 1995 
As CC 
edge 
Tests 
also 
appe 
curve 
and s 
point: 
previ 
5.1 E 
Figur 
in (F: 
secti 
three 
runni 
imag 
  
certe 
have 
equ: 
In e 
PO € 
The 
obs: 
Not 
poir 
AS : 
(K), 
orie 
1A