OOo OO O°O”OO9\O©°OO©°)© OO OO OO OO OO OL OO CC 
Ob 00 © 00°” C°0202020 © 090°0°© 900°» 9 02009 © 0 Q 
Q'ODOOOGXooocdocoocoooodooocccuo 
nages, it has been 
e and therefore fast 
'enerally sufficient, 
idated blobs being 
subsequent bundle 
ages on the other 
typically identified 
cess. A shape test 
tion process for the 
st suitable for low- 
the following three 
ow-quality images: 
on via a best-fitting 
nerally require only 
lunder’ pixels from 
because of inherent 
thm. The procedure 
ziation of intensities 
unders those pixels 
mean minus a given 
rocess is needed to 
on for the ellipse fit 
to remove regions 
hot-spots. Whereas 
mogenous reflection 
nsity distribution of 
ally quite inhomo- 
es the pixels of such 
try as dark blunder 
region to the point 
ze testing. 
ellipse to the target 
region. The centre 
ed for the ellipse fit 
been described, for 
| Luhmann (2000). 
a best-fitting ellipse 
or large targets (say 
ions invariably pass 
ry point distribution 
estimation of ellipse 
| of departures of the 
à 
/ 
2.4 Sub-Pixel Centroiding 
Once a target region has passed the validation process its exact 
centre needs to be determined. Though various centroiding and 
template matching methods can be found in the literature (e.g. 
Atkinson, 1996; Luhmann, 2000; Shortis et al. 1994), the 
intensity-weighted centroiding approach is most often adopted 
in automated VM because it is simple and very fast to compute: 
n m x. 
kN : 
Xo | _ i=l j=l Yij 
n m 
Yo 
Ej 
i=l j=1 
@) 
Here, x; and y; are row and column coordinates of pixels within 
the target blob, g; is the corresponding grey value, and (x,, y,) 
are the final centroid coordinates. The main issue of concern in 
any high-precision  centroiding method is background 
thresholding. The threshold has to be chosen as low as possible, 
so it removes all background noise without deleting valid edge 
pixels of the target region. The target edge pixels preserve the 
important geometric information for sub-pixel centroiding, 
which in VM can reach accuracies to 2-3% of the pixel size. For 
today’s large area CCD cameras, this can translate to an image 
coordinate accuracy of close to 0.2 um. 
3. EO DEVICE DETECTION 
Effective automatic detection of the EO device is necessary for 
initial spatial resection of the images forming the VM network, 
which is in turn critical to a successful solution of the image 
point correspondence problem using epipolar geometry. 
Multiple strategies for automatic recognition of target groupings 
within an image can be designed. Theoretically, recognition 
strategies for coded targets could also be employed for the EO 
device, but the main difference between coded targets and an 
EO device is that the points on the EO device have known XYZ 
coordinates within an arbitrary datum, which initially defines 
the reference coordinate system of the full VM network. One 
reason that EO devices are designed to be noticeably larger than 
coded targets is to guarantee sufficient geometric strength for 
the calculation of EO parameters for all images via either 
closed-form resection or, less commonly, relative orientation 
Whereas EO devices in commercial VM systems such as 
V-STARS (GSI, 2002) use recognisable fixed patterns of targets, 
a more flexible recognition and initial EO concept has been 
developed for Australis, where the user has the flexibility to 
define their own EO device under minimal design constraints. 
Figure 4 shows an example EO device suited to Australis. This 
design incorporates one basic constraint, namely that all points 
must be enclosed within a bounding shape, a circle in this case 
(a rectangle, square or triangle would also suffice). To achieve 
better approximations of orientation parameters, one of the five 
targets is usually non-coplanar with the others. Use of the EO 
device requires no more than the user specifying the XYZ 
coordinates of the EO target points. No labelling is necessary so 
long as the targets are asymmetrically distributed, since their 
relative positions are decoded in the recognition process. 
The EO detection procedure can be subdivided into three stages: 
device detection, target recognition and pattern decoding, and 
closed-form spatial resection. For the adopted procedure, the 
pattern decoding and resection stages are closely interrelated. At 
the initial image scanning stage, the primary recognition 
criterion for the EO device is the closed boundary, which need 
not be of a specific shape. The segmentation process treats the 
boundary in a similar way to other targets when forming an 
appropriate segmentation matrix, an extract from which is 
shown in Figure 5 for an EO-device boundary section. During 
the EO validation process it is checked whether the segmented 
region has a ‘hole’, ie an enclosed region within the boundary, 
the test being performed only if the enclosed area contains a 
specified minimum number of pixels. 
  
Figure 4. An example EO device for Australis. 
Following the image scanning, the segmentation matrix for a 
detected enclosed region is tested to establish whether it 
contains others targets. If the number of such detected targets 
matches that for an EO device, the region is flagged as an EO- 
device candidate. The design for Australis can account for the 
detection of multiple EO devices in an image, though this 
feature has not been applied in practical applications so far. 
Once the correct enclosed region and all EO-device targets are 
detected and measured, the pattern decoding and closed-form 
resection are performed. 
0 0 0 0 0 0 0 28 28 28 28 28 28 2B 2B 28 28 28 28 
0 0 0 0 0 28 28 28 20 20 20 20 28 28 28 28 28 28 28 
0 0 0 28 28 20 28 28 28 28 28 28 26 28 28 28 28 28 28 
0 28 28 28 26 28 28 28 28 28 28 28 28 28 26 28 28 28 28 
B 28 28 28 28 28 28 28 20 20 28 28 28 28 28 28 28 20 26 
8 28 28 20 20 28 28 28 28 28 28 28 28 20 28 28 28 28 28 
B 28 28 28 28 28 28 28 28 28 26 28 28 28 28 28 28 28 0 
28 28 28 28 28 28 28 28 28 28 28 28 28 28 20 D 0 0 0 
28 28 28 28 28 28 28 28 28 28 28 28 28 28 0 0 0 0 
28 28 28 28 28 28 28 28 28 20 28 28 28 28 0 0 0 00 
28 28 2B 28 28 28 28 28 28 29 28 28 28 0 0 0 0 
28 28 28 28 28 28 28 28 28 28 28 0 0 0 0 
28 28 28 28 28 28 28 282820 0 0 0 0 0 © 
28 28 28 28 28 28 28 28 0 0 00 
E 2028 20 28 28 28 28 0 0 0 0 00 
28 28 28 28 28 2 
  
  
  
  
  
  
  
  
  
  
   
9 
    
   
0 00 
0 
0 
0 
o 
0 
900 
00 
00 
00 
00 
0 0 
0 
0 
506000000 
ecocooo 
0 
0 
0 
0 
0 
0 
0 
0 
9 
00 
9 0 
0 
...........>.>{ 
oooococooooocoo 
9 90 
0 0 0 474747 0 
0 0 4747474747 47 0 
0 
0 
9 
0 
o 
0 
0 
0 
0 
9 
0 
o 
0 
° 
0 
o 
ooooocoooooocococs 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 0 47 47 47 47 47 47 0 
D 0 47 47 47 47 47 47 0 
0 0 4747 47 47 47 47 0 
0 004747474747 0; 
0000009000; 
e 
ooooooocooocococooooo 
o6o0oo0ooooooooooooo 
o 
0 
0 
[i 
0 
0 
0 
90 
0 
0 
o 
0 
0 
0 
0 
ooooócooon 
50500000 
o 
0 
0 
o 
  
Figure 5. Image of EO device with part of segmentation matrix. 
To determine which image point corresponds to the correct 
object point, a simple but rigorous strategy is employed. An 
algorithm tries all permutations of possible resections to find the 
most likely solution. If the geometric arrangement of EO points 
is favourable (an asymmetric pattern) the sought after 
permutation is the one with the lowest RMS value of image 
coordinate residuals in the resection. This algorithm has proven 
to be very robust in practical applications, even in cases where 
sensor calibration parameters (primarily lens distortion and 
focal length) are only approximately known, or where the lens 
distortion is assumed to be zero and a very course initial 
principal distance is adopted. 
63 —