g
r
e
y
v
a
|
u
e
occlusion
b= ft
ars of interior
ints.
2K CCD frame
are digitized
order to avoid
e CCD sensor
Nhich are later
era control and
interface. The
> which gives a
rgets (Godding
patterns
96
Figure 3: Images for testfield calibration
The technical description of the acquired images is as
follows:
Camera Rollei 6008 / ChipPack
Imaging device 2048 x 2048 pixels
Sensor elements 15um x 15um
Image format 30mm x 30mm
Lens Distagon 40mm
Size of testfield 80cm x 80cm
Mean image scale 1:25
Figure 3 displays the set of 8 testfield images.
4 RESULTS
41 Synthetic Point Patterns
Theoretical investigations and practical experiments have
shown that the position of a circular target can be
determined with an maximum accuracy of 1/1000 pixel.
Basically the resulting accuracy d depends on the size of
the target s and the contrast k between target area and
background. The precision has been determined
emperically to
d= nai (1)
s-k
This equation shows that accuracy improves with
increasing size and/or contrast. The factor 4 results from
differences between synthetic and real image data
(Broers & Hemken 1996). Theoretical investigations by
FORSTNER have shown similar results (Fórstner 1984).
Figure 4 shows some of the results obtained by the test
participants on the delivered imagery. In the case of
IMAGE1 the best results have been achieved by method
A, B and G (all ellipse operators). These methods show
327
mean differences to the nominal values between 0.003
and 0.009 pixels. The results of center-of-gravity
operators (F,l) and the line following approach (D) are
between 0.015 and 0.03 pixels. Even better results (0.002
pixels) have been produced with IMAGES (ideal patterns).
0,16
0,14
0,12
0,10
0,08
0,06
0,04
0,02
0,00
Image Image Image Image Image
1 2 3 4 5
Pixel
—H—B
—e—G m
A
F
——
3
Figure 4: Results on synthetic images
IMAGE2 shows slightly worse results caused by noise
added to the image greyvalues. Additive noise of 3
greylevels leads to reduced accuracy if ellipse edge
operators have been applied.
IMAGE4 (additional occlusions) show equal results for the
best operators in the range of 0.05 pixels. Although
occlusions should be handled easier by structural
approaches (i.e. ellipse methods) a higher number of
points could not be measured.
In IMAGES a variation of the background greylevel has
been simulated. This effect leads to a decreased
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B5. Vienna 1996