hina 
on-metric 
producing 
cted from 
| 1D least 
ted to get 
vith direct 
surement. 
tential in 
ected arcs 
000) with 
automatic 
ws a well 
n the case 
measuring 
and non- 
tegy is to 
accurately 
netry and 
data and 
is used to 
tial values 
)-designed 
pology of 
Matching 
section 2. 
strial parts 
d arcs and 
in be used 
s. System 
section 4. 
[ING 
racteristics 
computer 
lifference" 
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B5. Istanbul 2004 
is chosen as criteria, image matching equation is S vv = min - 
If only random noises are considered, error equation of image 
matching will be V= 8, (5) - g., y). This is the basic 
principle of least-squares template matching (Gruen 1985, 
Schenk 1999) which is widely used in digital photogrammetry. 
Line template matching is a two dimensional technique. Length 
of image window is usually more longer than 10 pixels. Image 
patch is usually rotated into horizontal to facilitate matching. As 
shown in Figure 1, the level rectangle represents the standard 
template for matching, while the dashed rectangle represents 
the image patch to be matched. Displacement along the line is 
not important in image matching, but the angle between 
template and image must be eliminated, so two unknowns dy, 
and dy, are essential to fit the small rotation angles between 
image patche and template. Note that real template should be 
generated according to image patch before image matching. 
  
  
  
[EM eee 
TT €— M " 
Hera e | 
(Orione eis 
[mn EE I dy; 
| ue i 
  
  
Figure 1. Two unknowns in line template matching 
Grid points are detected as the intersection of two matched lines 
fitted to each corner, as shown in Figure 2. The black crosses 
are the predicted image corners, and the white crosses are the 
matched ones. The precision of image matching results is 
higher than 0.05 pixels. 
  
Figure 2. Initial projections and matched results of points 
Left of Figure 3 shows the initial projections of line segments of 
the part. The initial image lines are usually several pixels away 
from the real image edges. Although many rusts exist on the 
part and can be seen clearly, the matched image lines are well 
fitted to the real image edges (right of Figure 3). Actually, the 
matching precision is also higher than 0.05 pixels. 
  
Figure 3. Initial projections and matched results of lines 
2.2 One-dimensional Point Template Matching 
As we know, a line can be represented by a group of colinear 
small segments. If image window of line template matching is 
subdivided into small segments, cach with a length of 2-5 pixels, 
named “point segment”, the rotation angle between standard 
template and small point segment can be neglected. Matching 
between the “point segment” and template is called point 
template matching in this paper. Different from line template 
matching, there is only one unknown for one-dimensional point 
template matching, the vertical shift dr between template and 
image, as shown in Figure 4. Angle is assumed to be not existed, 
since the length of point segment is usually very short. Figure 5 
shows the initial projection and matched result of a circle by 
one-dimensional point template matching. Although there are 
many rusts, the matched circle is well fitted to the image. 
  
  
  
  
  
  
Figure 5. Initial projection and matched result of a circle 
3. 3D RECONSTRUCTION OF INDUSTRIAL PARTS 
The topology of CAD data of sheetmetal part is assumed to be 
correct since it is designed on computer and often checked 
many times before the part is produced. But the geometry of 
sheetmetal part is usually not the same as CAD-designed data 
because of mis-operation during producing or deformations 
after a period of usage. Detailed approach of how to reconstruct 
the correct geometric model with the designed data and 
information extracted from imagery will be discussed. 
In this paper, world coordinate system is chosen the same as the 
one of grid. Generally, coordinate system of the industrial part 
defined in CAD-designed data will not be identical with the 
world one, there are at most six elements of rotation and 
translation to convert the CAD coordinate system into world 
coordinate system (grid coordinate system). 
Sheetmetal parts are mostly composed of line segments. This is 
the reason that we choose line photogrammetry to reconstruct 
and measure them. As shown in Figure 6, the image line pq, 
space line PQ and the projection center S should be coplanar, 
while p and P, q and Q are not necessarily correspondences, 
which is the most important advantage of line photogrammetry 
(Debevec 1996). A line in the image can be parameterised in 
several ways. Although two parameters are sufficient for 
representation of an image line, four image coordinates of two 
end points are used here, because it is singularity-free and easy