International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B5. Istanbul 2004
Active optical devices are based on an emitter, which produces
some sort of structured illumination on the object to be
measured, and a sensor, which is typically a CCD camera and
acquires images of the projected pattern reflected by the object
surface (Rochini, et.al, 2001). In most cases the depth
information is found by triangulation, given the relative
positions of the emitter and the sensor. The passive measuring
methods function like human vision system. In MEDPHOS, a
combination of active and passive methods has been used.
5. SYSTEM CONFIGURATION
Considering the error propagation from the imaging component
only, the positioning accuracy of object point coordinates is
mainly determined by four factors: the object distance, the base
line, the focal length, and the mean square error of image
coordinate measurements.
The selection of the CCD cameras is critical in terms of pixel
spacing and sensing area specifications. Pixel spacing affects
the accuracy of image coordinate measurements, while sensing
area of the camera eventually influences the parameter setting
related to base line and focal length since the positioning
accuracy is dependent on base line and focal length of the
cameras. The larger the focal length, the larger the image scale.
On the other hand, the use of a shorter focal length will increase
the lens distortion errors. In addition, using a larger focal length
instead of a larger base line will reduce the ambiguities of
image point matching and improve the point measurement
accuracy. However, a tradeoff should be made in order to
maintain a certain field of view (Tao, 1999). If a large sensing
area is used, either the focal length can be increased or the field
of view can be extended. Thus, the base line can also be
extended, provided the same image overlap is maintained. It is
obvious that the large sensor cameras offer better performance
in terms of sensing area, pixel spacing and number of pixels.
Consequently, if large sensor cameras are employed, the
settings of imaging parameters will be more flexible and the
total performance of the system can be improved. However, the
problematic aspects of the low image capture rate and high
storage requirements have to be taken into account. In addition,
the geometric performance of different types of CCD cameras
may vary in terms of electronic noise.
An analysis for optimal configuration shows that the maximum
length of the base line is restricted by the desired overlap
percentage, the overlap percentage is affected by the tiled of
view angle, and the filed of view angle is determined by the
focal length and the sensing arca of the camera. Regarding the
above considerations and the average size of medical objects to
be measured, an overall compromise was achicved to reach an
optimal configuration.
5.1 Multi Camera Concept
Multi image geometrical configurations have been recently
proposed to overcome the limitations of binocular vision
(Faugeras, 2001). The limitations including mainly relatively
poor reliability and low accuracy of reconstruction process. If
third and fourth cameras are added, the geometry becomes
much more richer than that of two camera system. By applying
the trifocal and four-focal constraints, point correspondence can
be found robustly. If the cameras geometries are known,
transfer is done in a straightforward fashion by three
dimensional reconstruction and reprojection (malian, et. al.,
2002).
5.2 Structured Light
Surface measurement of skin involves some problems regarding
targeting especially if the object is dynamic or alive. Structured
light means the projection of patterns onto an object surface. It
provides in some cases the only reasonable approach for surface
reconstruction. Different technologies can be used to produce
structured light pattern: laser emitters, light projectors, slide
projectors, and video projectors. In particular, target projection
is used for any object surface that does not lend itself for target
placement or does not have rich texture. This also avoids the
time consuming placement of retro reflective targets and the
placement and alignment of the corresponding light source
required for retro targets. These targets are detected and
localized by specific image processing techniques. The pattern
used should satisfy some characteristics: it should not be
drastically altered by small variations in photometric and
geometric conditions, the detection, localization and
discrimination of its constituent features from the reflected
image should be easy and accurate and so on. The size of the
object is limited by the possibilities of the projector and the
environment, that is, by the strength of the projector
illumination and environmental light. Project planning must
therefore take into consideration not only the characteristics of
the camera such as depth of ficld and filed of view, but also
those of the projector. This idea is used in MEDPHOS as dot
target projection using a slide projector that serves also as an
active camera with known calibration parameters.
5.3 Prediction
By going beyond the classical binocular vision, the problem we
now address is to predict how the scene would look like from a
third and a forth camera. In other words, given the calibration
information of third and fourth cameras and image coordinates
of an object point in one image, predict the locations of
corresponding image points on other images. The transfer
concept is used in MEDPHOS; The homologous point of a
selected image point lies on the epipolar bands of the other
images. In other words, for one or more image points in a given
image set, the corresponding points in other image sets can be
predicted using Essential and Fundamental matrices. Width and
length of the epipolar band can be restricted with information on
the error budget and approximate depth, the latter is estimated
by applying the MEDPHOS algorithm on a few non-ambiguous
dots distributed within the projected pattern.
6. PROCEDURES
MEDPHOS consists of four digital and calibrated cameras
mounted on a rig that allows required rotations. The cameras arc
activated in a synchronized manner. The base lines of the
cameras can be set in different lengths. The pattern projector is
fixed at the center of the system (Figure 6). It can accept various
pattern types. A total calibration of the system by series of
convergent photography and self-calibration bundle adjustment
(malian, 2000) provides the relative position and attitude of the
cameras and the projector as well as the epipolar geometry for
any image point in any camera. The captured images are
directly fed to the computer where the related software
processes the data in real time. The designed dot pattern is
projected onto the object and recorded by the four camera
system. To reduce the effects of specular reflectance,
homomorphic filtering is applied to the images (malian et.al,
2002). The observed light pattern is then used to detect the
image coordinates of the dots using an optimal thresholding
314
Inter.
techr
appli
perfc
the g
Estal
recor
corre
refer
matc
proje
came
matc
malc
the 1
matc
prop
glob:
cons
mult
cons
In g
inten
can |
degr
stabl
cond
low-
the |
to a
Thes
para
kno
mea:
not
Feat
obj €i
on :
winc
are
assu
VIEW
phot
inter
assu
relat
In N
emp
the i
geor
used
imag
A 1
COI
bo
