bul 2004
egarding
ructured
irface. It
r surface
produce
rs, slide
rojection
or target
oids the
and the
t source
ted and
¢ pattern
not be
tric and
on and
reflected
ze of the
and the
projector
ng must
ristics of
but also
S as dot
Iso as an
)blem we
ce from a
alibration
ordinates
ations of
transfer
oint of a
the other
n a given
ts can be
Vidth and
nation on
estimated
mbiguous
cameras
meras are
es of the
'ojector is
pt various
series of
djustment
ude of the
metry for
nages are
software
pattern is
ir camera
flectance,
lian et.al,
detect the
resholding
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part BS. Istanbul 2004
technique. Then, a connected component labeling procedure is
applied to identify the points. Watershed transformation is also
performed to isolate the overlapping dots and find the center of
the gravity of each dot.
Establishing correspondences is the key problem in 3D
reconstruction from multiple images. The goal of
correspondence is to assign matches to each point in the
reference image. The essential problem in geometric image
matching lies in establishing a correspondence among the
projections of the same physical location in each of the four
cameras. Conceptually, the correspondence process in image
matching consists of two stages : local matching and global
matching. In the local matching stage, for every feature point in
the base image, an attempt is made to find a set of candidate
match points in other images which have similar local
properties and which satisfy the four focal constraint. In the
global matching stage, a scheme for imposing the global
consistency among the local matches is used to disambiguate
multiple local match point candidates through the solving of a
consistent labeling problem.
In general, image matching algorithms can be divided into
intensity-based and feature-based. Intensity-based algorithms
can produce dense depth information but are very sensitive to
degradation in illumination and contrast and hence are not
stable. Their distinguishability is poor especially in textureless
conditions. Furthermore, most intensity-based techniques have a
low-pass filer effect on the derived object surface as a result of
the large matrix sizes. On the other hand, smaller matrices lead
to a rapid reduction of accuracy and reliability. Moreover,
These techniques require good approximations for the unknown
parameters. That is, the solution of the problem should be
known a priori. Regarding the typical situations of medical
measurements, it is obvious that intensity-based approaches are
not reliable enough for users in medical and health sectors.
Feature-based algorithms first select some salient points of the
object and perform the matching process on these points based
on some similarity measures in particular the correlation
windows centered around the selected features. These methods
are of course more robust but still rely on the Lamebertian
assumption that states image intensity is independent of the
viewing direction. This is not usually the case in close range
photogrammetry especially in for medical objects. So, applying
intensity-based techniques for an object without Lambertian
assumption being met, can produce a lot of points that are
related to homologous elements.
In MEDPHOS, a new strategy for image matching has been
employed that needs a minimum amount of information about
the intensities of the pixels and object space conditions. Robust
geometric constraints that exist in a multiple view system are
used to establish robust correspondences among different
images of the object.
A number of difficulties arise in the establishment of
correspondences that must be taken into account:
1. False matches can occur due to
e Photometric differences and specular reflectance
e Lack of texture
e Incorrect camera calibration
* Discrete nature of a digital image
e Noise
2. Missing matches can occur due to
e A point becomes invisible when the line of sight
is interfered with by another object
e Part of the scene not being in the field of view of
the other cameras
e A matching element is too weak in one of the
images and, therefore, is discarded as noise in the
feature extraction process
In MEDPHOS software, the image point correspondence
problem is solved using maximum weighted bipartite technique
(malian, et.al., 2002).
After the establishment of all correspondences, three
dimensional object points are computed and the quality control
criteria is done. Beside the determination of three dimensional
surface points, a blunder detection and correction strategy is
used to resolve the following problems:
e Gross errors due to image point detection and
localization
e Gross errors due to the ambiguity in point
correspondences
e Gross errors due to incorrect three dimensional
reconstruction.
An integrated multiple image forward intersection is carried out
to monitor the quality of the results. The resulted point cloud is
classified into different reliability levels based on the
information stored in the cost functions used during the
reconstruction procedure and the distance to next best answer
for each matching candidate. Finally, the desired quantitative
medical parameters are computed and a global surface fitting
technique is used for three dimensional surface reconstruction.
D
CON
M
=
Front View
ir
Hee A
Top View Side View
Figure 6. MEDPHOS: Hardware Plan
