181 
matching algorithms. Also, the point density in the profiles is automatically adapted to the surface curvature by forcing 
the measured 3D points strictly into a tolerance hull. : 
The internal measuring accuracy is directly influenced by the inherent potential of the area-based matching technique 
being applied by default. If we recall the camera configuration of section 3.3.1 and presume a parallax precision of 
1/10 pixel, the standard deviation of a single profile point amounts to 0.1 pixel * 15 *3 = 0.07 mm. Actually, this is in 
the same order of magnitude than the point cloud precision. Figure 3 shows a concept model measured with profiles 
being parallel to all 3 axes of the coordinate system. 
=] LA 
  
  
   
  
  
  
  
  
  
Figure 2: Point cloud measurement on a concept Figure 3: Profile measurement on a concept model 
model 
3.3.3 Comparison of both measurement options 
At this point we should compare both measurement options. Obviously, there is a clear distinction between the two 
measurement "philosophies". The point cloud generator starts from scratch without any pre-knowledge about the sur- 
face and creates subsequently in each image pyramid a point cloud. The real surface is approximated stepwise and 
gains its best accuracy in the last image pyramid level. The point cloud is almost not affected by poor texture areas, 
reflections or image errors (scratches, wrong pixels) as long as such problem areas are small. Also, the mathematical 
model appears flexible and makes possible to automatically adapt to the surface curavture or to smoothen the point 
cloud, respectively. The surface is described in facets so that specific profiles could be interpolated. The measuring 
precision is gained by a large number of measured 3D points per finite element mesh and hence the system can pro- 
vide reliable statistics like internal precision. 
The profile measurement option starts digitizing the surface from a precise starting point and measures directly on the 
surface along the profiles by keeping problem areas as small as possible. The point density is also adatped to the 
surface curvature. The measuring precision is mainly influenced by the potential of the least squares matching tech- 
nique. The measuring speed of the point cloud generator is twice as good as the profile measurement option (see de- 
tails 5.2). Apparently, techniques of both methods could be combined very efficiently . For instance, measured profiles 
could be filtered by the surface reconstruction part of the point cloud generator, and hence small gaps could be closed. 
4. HARDWARE REQUIREMENTS 
4.1 Workstation 
The current version of the software is designed in first instance for UNIX platforms. The first implementation runs on 
SGI workstations which are well-known for their computing and graphical power. All experiments shown in this paper 
were performed on an Indigo2 with a R4000 MIPS processor. 
4.2 Texture projection and texture slide 
If objects have no natural structure, an artificial texture has to be provided in order to ensure a successful application 
of correlation methods. A well-known technique is the optical projection of a random pattern by a slide projector. Fig- 
ure 4 shows such a typical pattern which includes macro structures still visible in low resolutional levels of the image 
pyramid. 
As already mentioned, digital images are to be provided either by digital cameras or by scanners in combination with 
analog cameras, respectively. For both ways there exists a wide field of products which have to be selected according 
to the project requirements. : 
IAPRS, Vol. 30, Part 5W1, ISPRS Intercommission Workshop "From Pixels to Sequences", Zurich, March 22-24 1995