one row of pixels in the right image, instead of seaching in two 
dimension. 
The correlation itself is carried out on a coarse-to-fine 
resolution basis. For this reason an image pyramid with 
decreasing image resolution is built on top of the normalized 
images. The computation of the image pyramid is combined 
with the normalization of the images and appears to the user as 
a one step process. 
2.2.4 Generation of grid ordered point clouds 
This method of producing surface data uses feature based 
matching to generate a grid ordered point cloud of 
measurements. 
So called feature points which are generated from the 
normalized digital images using the Fórstner operator serve as 
input into a two step process. The first step is to match the 
interest points generated for the left image to their right image 
partner. The second step takes the intersected 3D points as 
input into a robust finite element adjustment. 
An imaginary plane lying parallel to the normalization plane of 
the images is taken as reference for the grid representation of 
points. Points on the parametric lines on this reference plane at 
regular intervals define the corners of the finite elements. The 
third dimension results from the perpendicular distance of the 
grid point to this reference plane. The advantage of choosing 
this 22 dimensional approach is the very efficient way at 
which the resulting systems of equations can by formulated 
and solved. In addition, it is possible to introduce weighting 
factors for curvature and torsion between grid points 
influencing the smoothness of the calculated grid. 
The disadvantage of this representation of points is, that it is 
not possible to represent ambigious surfaces on a single grid 
storage unit, because the data structure is not truly three 
dimensional. Such surfaces are on the other hand highly 
unlikely to capture on a stereo pair. 
The grid generator produces one set of grid points for every 
stereo pair present in the project. The point clouds are usually 
measured in the various stereo pairs in such a way to produce 
overlapping point clouds. 
Owing to the preference of profile data the profile generator 
may be used to interpolate point coordinates from the grid files 
generated by the grid generation process. 
2.2.5 Profile generation 
This mode of operation can generate surface data by the well 
known method of least squares matching. The required input 
data include the two normalized images making up a stereo 
pair and parameters like average point density. In addition it is 
alternatively possible to generate points by interpolating 
coordinates from the previously measured grid ordered point 
clouds in which case the grid files serve as input instead of the 
image files. 
The advantages the least sqaures matching method are the high 
accuracy of a single measured point and the possibility of 
generating the surface normal of the matched point in addition 
to its coordinates. The designers today still tend to prefer 
profile data as input to the CAD system for some reasons. One 
might be that some CAD systems are not capable of handling 
the massive amount of data from unsorted point clouds, 
another reason may be that it is easier to imagine the shape of 
the object on a computer screen through profile data. 
The generation of surface points is performed along predefined 
planes of intersection with the object. The algorithm requires 
478 
as input the parametrized intersection planes in at least two 
coordinate directions and at least one starting point coordinate. 
In the case of least squares matching the starting point has to 
be digitized accurate enough to fall within the pull-in range of 
the least squares matching, which is about 2 pixels. 
The profile generator will select the intersecting plane closest 
to the starting point. A first preliminary match is performed to 
determine the direction towards this plane, the program will 
step froward from here towards the plane until this is reached. 
If planes are defined in at least two coordinate directions, 
measuring a profile along one plane will generate intersections 
points, called nodes, with cross profiles as well as regular 
points. The nodes serve as new starting points for measuring 
the cross profiles, which in turn generate even more nodes until 
the complete stereo pair is covered. 
Point sets from neighbouring stereo pairs can be generated to 
coincide if the profile planes are defined such that their 
parametrization is equal between stereo pairs. 
It thus becomes possible to produce coinciding point sets from 
grid ordered point clouds by interpolation. 
2.2.6 Conversion of measured data into CAD usable format 
  
Figure 2 - Example of measurements of a steering wheel 
All data produced by the previously described measurement 
modes are stored in binary format, optimized for speed during 
the measuring process. This format is generally not suitable for 
CAD packages. Once because the program lacks the interface 
to directly access the computed data, and secondly because the 
format may not be efficient for the following task. Thus there is 
a conversion tool available to produce ASCII data from the 
binary files. 
The conversion of grid sets results in profile like point strings, 
where profile planes are defined by the parameter lines of the 
grid reference planes. 
The conversion of profile data results in point strings lying on 
the defined sectional planes. Since the measured or 
interpolated points usually lie on coinciding planes it is 
possible to combine the profiles measured in more than one 
model into one global data set. 
The output data can be formatted into VDA or DXF files which 
in turn can be read by many CAD applications. 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B5. Vienna 1996 
2.2.7 Accu 
The typic: 
leads to a 
apporxima 
scale here 
0.5 mm! o 
The grid g 
mm depen 
accuracy Q 
The typicz 
along the | 
profile pla 
yields an : 
0.15 mm ( 
2.3 Can 
The came 
have a fi 
cameras tl 
CCD sens: 
the lenses 
characteri: 
Since the 
purposes * 
here much 
under unf 
larger ap! 
focusing r: 
Thus, one 
greater di: 
for a speci 
This requ 
system. T 
desired foc 
scale mot 
1.5 m and 
enough de 
still sharp, 
The lense 
16 differe 
calibration 
markers oi 
automatic: 
into a bun 
model the 
positions. 
The result 
and the f{ 
measuring 
The calibi 
constant . 
calibratior 
focusing d 
this could 
Since the 
generator . 
of 1/10th 
orientatior 
Experienc 
results bet 
of the sens