B5. Istanbul 2004
arts
he first task is to
cameras center of
This process is
the object and for
libration process is
;h of the cameras.
meters (Xt, Yt, Zt)
eters (phi, epsilon,
he parameters, an
| points of known
ne of the cameras.
sed to estimate the
il et al. 2001)
cameras’ extrinsic
»bject is calculated
vo images and the
o overcome the
pattern is projected
pondence problem
id in the literature
, Wust and Capson
methods to use a
jects may be multi-
(1998) analyze the
ystem. This model
LCD projector, the
and the surface
is that it assumes a
in the three RGB
showed that when
n captured in RGB
e channel contains
'o channels contain
a noise immunity
jetween a and the
ossible strips in the
im accuracy under
ent.
nted in the present
s the different RGB
'D sensors changes
d patterns can be
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B5. Istanbul 2004
squeezed or extended in order to include the regions where the
system can efficiently differentiate between color levels.
The colors and texture of chicken filets are relatively constant
among filets, and are also uniform across the filet itself. The
system presented in this work addresses the problem of
matching points in stereoscopic images for objects with
uniform color and texture. The main idea is to project a non-
uniform varying color pattern. In some wavelengths the
spectrum of the reflected color pattern will change rapidly, but
in some ranges the color spectrum will change slowly. This
change depends on the object, on the imaging system and on
the projection system.
Projection of a monotonic linear increasing hue value pattern
on the object does not result in the same linear increasing
pattern on the captured image. This is due to the object's
chromatic characteristics and the features of the capturing /
projection systems. Analysis of the captured pattern showed
that the hue values, which are absorbed by the object, look
similar on the captured pattern. Wavelengths, which are
reflected by the object body, are well detected as changing on
the captured pattern. The ability to detect small amounts of
change depends on the projection / capturing systems. The
systems sensitivity is also not constant and may change from
one wavelength to another.
Assuming that the best-fit projection will be the one that 1s
captured as linearly changing across all the hue values, the
specific objective of this work was then narrowed down to
develop the methodology for constructing a projection pattern
that will be captured as linearly changing across all the hue
values. For a given projection system, image capturing system
and chromatic / textural object characteristic a projected
pattern that leads to a linear captured pattern can be found.
To find the best projection pattern an iterative method was
implemented. Let i be the iteration number. Let L be a pattern
with linearly varying hue values. Let Pi be the projected hue
pattern at iteration i. Seti 2-1 and P1 = L.
The stages of the iterative method:
1. Project (PD).
2. Capture the pattern of the hue values as reflected by
the object (Ci).
3. Compute the differences between the captured
pattern, Ci, and the linear pattern, L. Denote the
difference pattern as the delta pattern Di where Di =
Ci-L.
4. If all elements of D; are smaller than a small value
€ then stop.
else Pıy = P; - @ D; , where Q is the learning
rate, go to 1.
7, EXPERIMENTAL RESULTS
The method was initially tested on a white board plane. A
linearly varying hue pattern was projected on the surface of the
board and two cameras captured the scene. Figure 2 shows the
result of the reflected color pattern after the first iteration. The
X-axis is the pixel number across one cycle of the varying hue
pattern (from red too violet) and the Y-axis is the hue value. In
the first iteration L = P1 which means that the projected pattern
is the desired linearly varying pattern. C1 is the resultant
pattern captured by the imaging system. P2 is calculated
according to P2 = P1 - & D1. The learning rate was set to be
0.7. In the next iteration the projection pattern was P2. Figure 3
shows the results at iteration 11 after the process converged.
One can see that the captured pattern almost coincides with L,
which means that the hue of the captured pattern varies almost
linearly.
08} =
0.8 / 1
|
0.7 {+
Captured mb |
ed, pattem 7 Projected
ss | (C1) /
/ pattern (P1)
04 y^ and L
03} iier 1
tT |
a y om Corrected 1
7 s t |
"E pattern (P2) |
0 rs NO tpi]
0 20 40 60 80 100 120 140 160
Figure 2. Iteration lof the iterative process
os. Captured
pattern
(CID)
06 s
07r
os! p Projected
A pattern
56,100 150. 200 250 300
{A) iteration no.i
0 100 200 300
18) iteration no.5
100 200 300
{C} iteration no.11
Figure 4. Error rate of phases 1, 5 and 11 of the iterative
process
Figures 4 A, B, C show the DEM profiles, generated at
iterations 1, 5 and 11 respectively. X-axis is the DEM profile
and Y-axis is the pixels height in millimeters. The
improvement of the DEM accuracy of the flat white board
along with the iterations is clearly seen: as the iterative process
