The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Pari B4. Beijing 2008
3. IMAGE MATCHING AND INVESTIGATION
For the automatic image matching of the DMC-images the
Hannover program DPCOR has been used, which is based on
least squares matching. The approximate corresponding image
positions are determined by region growing, published by
(Heipke 1996). Object coordinates are computed by intersection.
Figure 2 Workflow of matching procedure
Figure 2 shows the work flow of the data handling. It mainly
has following steps:
• Manual measurement of few seed points if not enough
control points already available; the number of required
seed points depends on image similarity and decrease with
smaller time difference of taking the both images of a
stereo model.
• Automatic matching in image space by least squares
method.
• Transformation of pixel to photo coordinates
• Computation of ground coordinates by intersection.
• Generation of Digital Surface Models (DSMs), 3D view,
contour lines, visual inspection.
The point spacing of matching can be specified in the used
program DPCOR. By default every third pixel in line and also
column direction will be matched. A matching of every pixel
leads to high correlation of neighboured points.
Figure 3 shows an example of matched points in an urban area,
which leads to a very dense DSM. The small separated building
with some trees around, did not lead to a clear building shape.
Dark parts, caused by shadows, may lead to gaps.
Figure 3 overlay of matched points to DMC-image
(Matched points = white
dark parts or marked in red = matching failed)
Following situations may cause problems of image matching:
Smooth and sleek surfaces
Poor or no texture
Occlusion
Moving objects
Surface discontinuities
Shadows
repetitive objects
This may cause failed matching or reduced matching accuracy.
Because of expected problems, the investigated area has been
separated into two classes:
• Open area - not disturbed by remarkable vegetation
and buildings
• Area with man-made objects and trees
Forest areas were not included.
Open area:
The first investigation has been made with different settings -
the tolerance limit for the correlation coefficient (R) has been
varied between 0.7 and 0.9; the step width of matching mainly
was 3 pixels in each direction, but also 1 pixel step width has
been used; finally the window size of the sub-matrix for
matching has been changed from 10 x 10 to 5 x 5 pixels.
parameters
R= threshold for accepted correlation
matching success
R = 0.70 , Step width = 3
sub-matrix 10x10 pixels
87%
R = 0.80 , Step width = 3
sub-matrix 10x10 pixels
84%
R = 0.90 , Step width = 3
sub-matrix 10 xlO pixels
75%
R = 0.80 , Step width = 1
sub-matrix 10 x 10 pixels
85%
R = 0.80 , Step width = 3
sub-matrix 5 x5 pixels
62%
Table 1: Matching completeness in open area depending upon
chosen parameters
In open areas with sufficient object texture, the image matching
has no problems. However, some problems are caused by
missing texture, isolated trees and sometimes repetitive objects.
1208