free network adjustment, with a single distance observation
providing scale to the network.
For providing texture on the measurement surfaces, both a
standard slide projector and overhead projector were tested for
stability of the projected patterns. It was found that the slide
projector was unstable (due to temperature changes and vibrations
of the slide and projector due to the fan), whereas the overhead
projector provided a relatively stable projection after allowing for
a warm up period (one and a half hours in this case).
3. MATCHING PROCEDURE
Multiple images (usually four or more) of an object are acquired
with a pre-calibrated camera. The matching of surface points in
the multiple images is started once the exterior orientations of the
various images has been computed. This orientation is achieved
through the use of signalised points on a reference frame
surrounding the object.
The following procedure was adopted for the image matching and
surface measurement:
1. In a reference image, find a dense distribution of well
textured reference patches on the object.
2. Determine provisional values for all image coordinates of
the patches and the 3-D object coordinates of their
associated surface point.
3. Perform MPGC matching for each set of patches to
determine accurate image correspondences and 3D surface
position.
4. Monitor the matching results with a run-time blunder
detection process and also perform a post-measurement
blunder detection.
In the first step it is possible to use any favoured interest
operator, such as the Forstner Operator (Forstner and Gulch,
1987) or simple edge operators. The second step is the main
subject of this paper, for which a combination of nearest
neighbour extrapolation and MIC has been investigated. The third
step is well documented as a high accuracy image matching and
3D position estimation technique. The blunder detection
technique is also an important element of the procedure; here
various parameters such as the a posteriori standard deviation of
unit weight, the average correlation coefficient and the number
of iterations (from the MPGC matching) are compared to
absolute and relative thresholds to detect and eliminate blunders
automatically at run-time and post-measurement stages
respectively. For more details on all the aspects of this matching
procedure see Van der Vlugt (1995).
The task of determining all provisional parameter values needed
for the MPGC matching can be seen as the primary matching
problem in that it is this task which actually determines the
correspondence between the reference patch and the search
patches, as opposed to the MPGC algorithm which provides the
fine matching. The provisional values needed by the MPGC
matching for a single match are: X,Y,Z object coordinates of the
surface point; x,y image coordinates of the reference patch and
all search patches and affine parameters for all search patches.
The affine parameters are usually set to one for the two scales
and zero for the shifts and shears. The other parameters can all
be computed from the reference patch position (known) and only
one other coordinate (image or object), using the known camera
orientations and distortions. It is advantageous to use the depth
coordinate (often defined by the Z object coordinate axis) from
which to calculate all the others. The Z-coordinate is thus loosely
defined here as the depth coordinate or height, which is more or
less perpendicular to the general object surface. The provisional
value problem thus reduces to finding the correct Z-coordinate
given the reference patch position, much like adjusting the height
of the floating dot onto the terrain surface in an aerial image
stereo-pair.
DEM image
fi
(x,y,Z)
"n
and provisional,
patch positions / ^
(X, )
Figure 1. The DEM image and computation of
provisional values for MPGC matching
The MIC search procedure has been developed for estimating the
correct Z-coordinate. MIC is a reliable method of determining
provisional values when multiple images are used, so it has been
adopted as a back-up method for a less computation intensive
(and less reliable) algorithm using surface height extrapolation.
An initial surface match is obtained using MIC and subsequent
surface point provisional Z-coordinates are then computed using
the surface extrapolation technique until an MPGC matching
failure occurs. When a matching failure occurs, the controlling
routine assumes that an incorrect provisional value set was passed
to the MPGC sub-routine. The MIC search is then called and a
new set of approximations generated which are passed to the
MPGC algorithm for a second matching attempt.
The height extrapolation is incorporated into the matching
procedure as follows. A "DEM image" pixel-linked to the
reference image is set up. Each pixel in the DEM image would
ideally contain the height of the object point imaged by the
corresponding pixel in the reference image. However as not
every pixel in the reference image is matched and as the DEM
image is still growing at any one time, the pixels in the DEM
image either contain a height value indicating a successful match
or a value defined as "no height" for a large number of
unmatched pixels. When a reference patch with strong texture is
extracted from the reference image, the Z-coordinate of the
centre of this reference patch is calculated by extrapolating the
500
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B5. Vienna 1996
height :
image.
regular
this sim
Once a
remaini
Firstly,
arrange
orientat
second
projecte
coordin:
then trai
the kn
paramet
matchin
is store
more re
for the |
Multi-ir
grey lev
normali:
between
problem
search p
1. E
Figı
CO
to tl
When r
procedu
search i
known,
