OOOO OOD DODO QO OOD OO
Oe OOO OOO COOOL O ODO OOO
o0o00006060900000 0900990909 000 90
oo0000090900090090009090900S930
0900009009 090090900900HH4HHOSS0O0
OIC CD CD DLC uD pd pi iO ONO
OOOO OOO OC OOK Cl ED OHO
QUO QOO ODIO OOD DC OR mt i iO OO
oo0o0o00006060000000#-H"H+ 1H FH HOMO
OOO COCO OOOO OC HF IH HKD OID
OO OOOO 00 000000 O00 OOO OO
OOOO DON OD OO ODO 00 OO 0 O00
O00 CO OOONNNOOOCOOOOONOOOTOO OO
SOOONNNNNO0OD0900990909900090
QUOD ON NININRPKNOOCOCIOIDOOXOOCO0OGOtUo
SOOC0CONNNNNNOO0O0900000000
OO9OO0O9ONNNNNNOOO.O0OO0OGOOOOOO0OO0OC
GCooONNNNNNOOOOOOO0OOOO0OOO
OO9O0009NNN0O90009096S00Q00998SO090
ooce006000090000090990C000900009
QQODO0OO0Oo0o0!oQoOooooooOoooou«oonoocdo
QiOOoOOO0oO0OoOdoOoOoOoOocooogdooodococdxcd
QioocooOoOcoodooocdodocooc0coccooooccoocdo
Figure 3. Test image and its corresponding segmentation matrix after scanning.
An image segmentation matrix and an array of region objects is
built up during the scanning. A region object contains the
connected pixels of a target. This allows a fast validation later
on in the image measurement process. In detail, when the
starting and finishing pixels of a target are detected in an image
line (row or column), a new pixel region is created. This will
contain the two edge pixels and all pixels between. Then,
neighbouring pixel regions from subsequent lines are merged. If
a pixel has been added to a region, the region label is set at the
corresponding position in the image segmentation matrix. This
allows fast detection of new pixels within the neighbouring
rows. Figure 3 shows an example of a small test image and its
segmentation matrix.
2.3 Region Validation
Once the image has been scanned and initially segmented, the
important step of region validation is carried out so that regions
can be classified as targets or non-targets. Various validation
strategies can be employed to eliminate non-target regions prior
to sub-pixel target centroiding. However, strategies which aim
to differentiate target and non-target blobs with 100%
correctness are both very difficult to design and require
excessive computational effort and time. Considering that in
VM we design for high redundancy and optimal geometric
strength of the image station network, gross errors in target
validation and centroiding can be tolerated to a considerable
extent because such observation errors can be detected and
eliminated within the blunder detection process of the bundle
triangulation. Consequently, it is generally sufficient to design
fast validation methods which reject most (say 90-95%) of the
non-target regions, while accepting all legitimate target regions.
A second component of the target validation process is shape
testing. Considering that the perspective image of a circular
target will be an ellipse, an obvious shape verification
mechanism is the best-fitting ellipse. However, simple tests
such as blob size can also be applied to eliminate large bright
areas (e.g. sky) and small reflectance hot-spots. In the procedure
described, attention has been given to evaluation of the image
scanning algorithm in the presence of differing image qualities.
Practical experience has suggested that slightly different target
validation strategies are warranted for high- and low-quality
images (e.g. Figures la & 1b, respectively), even though the
same segmentation process can be employed for both.
In the case of near-binary, high-quality images, it has been
found that image scanning with only a simple and therefore fast
validation process based on size testing is generally sufficient,
with the small percentage of wrongly validated blobs being
rejected as valid observations in the subsequent bundle
adjustment processing. For low-quality images on the other
hand, many more non-target regions are typically identified
(incorrectly) in the image segmentation process. A shape test
via ellipse fitting combined with a normalisation process for the
detected target blobs has turned out to be most suitable for low-
quality images. Within Australis, therefore, the following three
validation processes are performed for low-quality images:
normalisation, a size test and shape validation via a best-fitting
ellipse computation. High-quality images generally require only
the size testing.
The normalisation process removes ‘dark blunder’ pixels from
the target region, these being falsely added because of inherent
weaknesses within the segmentation algorithm. The procedure
simply computes the mean and standard deviation of intensities
within the region and classifies as dark blunders those pixels
whose grey values are below the threshold mean minus a given
multiple of the standard deviation. The process is needed to
determine the correct boundary of the region for the ellipse fit
test. The normalisation process also helps to remove regions
that are small bright spots, e.g. reflectance hot-spots. Whereas
VM targets can be expected to have very homogenous reflection
intensity, as exemplified in Figure 2, the intensity distribution of
small, non-saturated light-spots is generally quite inhomo-
geneous. The normalisation process classifies the pixels of such
regions with strong variations in intensity as dark blunder
pixels. This can then reduce the ‘valid’ region to the point
where it is rejected within the subsequent size testing.
The shape testing process by best-fitting an ellipse to the target
first determines the boundary of the region. The centre
coordinates of boundary pixels are then used for the ellipse fit
computation. Recent ellipse fit tests have been described, for
example, by Fraser & Shao (1997) and Luhmann (2000).
Experience has shown that shape testing via a best-fitting ellipse
is a reliable criterion for region validation for large targets (say
greater than 5x5 pixels). However, small regions invariably pass
this test, since the redundancy in the boundary point distribution
is often too little for reliable least-squares estimation of ellipse
parameters, and for subsequent computation of departures of the
region boundary from the best-fitting ellipse.
z62-—
ee et — MH Do e AY s — e i-i
€) Fri Cc Un € rt o M
Q' Tum Uem GO wu
Qi
Pd me
