The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B5. Beijing 2008
791
defined techniques measuring the similarity of the local area
surrounding a feature. The image matching technique was
intended to precisely locate tie points during image-to-image
registration and estimate the accuracy of the local
image-to-image transformation. The choice was based on
Normalized Cross Correlation (NCC), which returns the
similarity of two windows of pixels (Equation 1). Two highly
similar windows centered on comers indicate that the comers
match. The matching result is the position in the test image
with the highest correlation score.
Y(I(x,y)-IXJ(x,y)-J)
(1)
Where:
A*>y)
and -j
Mx,y)
I(x, y) is the value of a pixel from window 1.
J(x, y) is the value of a pixel from window 2.
N is the total number of pixels in window 1 or window 2.
2.3 Reconstruction
A shorter baseline between stereo images is better for reducing
occlusions and increasing the reliability of matching. However,
because a smaller baseline produces larger errors along depth
direction, a larger baseline with an intersection angle near 90
degrees is needed to improve accuracy. Ideally, the system
would combine the advantages of short baseline and large
baseline. Our hypothesis is as follows. First, intermediate
images between two images with an appropriate baseline for
best accuracy should be used for better matching. Second,
tracking conjugate points through all images and reconstructing
a point intersected by conjugate light rays from stereo images
with an appropriate baseline can achieve an automatic
matching procedure for generating accurate output of a
reconstructed surface model.
By using stereo imaging, 3D object points can be derived by
the intersection of conjugate light rays which are defined by the
conjugate points, the IOP of the camera, and the EOP of the
image. A set of randomly distributed points can be obtained by
using the intersection process. A well-reconstructed 3D facial
model established from these random points requires an
interpolation method. Thus, Thin Plate Spline is used in this
stage. Thin Plate Spline (TPS) is an interpolation method that
finds a "minimally bended" smooth surface that passes through
all given points. The thin plate spline is the two-dimensional
analog of the cubic spline in one dimension and is the
fundamental solution to the biharmonic equation. The TPS can
represent the surface through a mathematical function where
the facial surface can be resampled and modeled with regularly
spaced points.
2.4 Registration
Cheng and Habib (2007) introduced an automated surface
matching algorithm for registering 3D geographic datasets
constructed relative to two reference frames. The Modified
Iterated Hough Transform (MIHT) is coupled with the Iterative
Closest Patch (ICPatch) algorithm to improve the convergence
rate of the matching strategy as it relates to the nature of
acquired surface models. This algorithm can be used to model
surfaces with randomly distributed points when it is unknown
how they correspond with each other.
Considering the characteristics of collected surface models, 3D
points can be used to represent the first surface (Si) while
triangular patches can be used to define the second surface (S 2 ).
3D similarity transformation is used for describing the
mathematical relationship or mapping function between the
reference frames associated with the two surfaces. Seven
parameters are involved in 3D similarity transformation,
including three translations, one scale, and three rotational
angles. A coplanarity constraint (Figure 3) is used here as the
similarity measure in this algorithm. The enclosed volume of a
transformed point and the corresponding patch should be zero
if the point belongs to the same plane as the patch (Equation 2).
V =
X,
Y,
z,
R
R
R
x n
y
z nn
pa
pa
pa
X p b
Z p b
X
r
z
pc
pc
pc
V
z ,
are
= 0
(2)
point from Si, and p a , p h , p c denote the coordinates of the
three vertices of the conjugate patch from S2
Figure 3. Similarity measure for relating conjugate primitives
in two facial models.
The MIHT approach is a voting procedure that derives the most
probable solutions of the transformation parameters needed for
the best alignment of two surface models by considering all the
possible matches between points in Si and patches in S 2 . The
MIHT also determines the correspondence between conjugate
surface elements in the involved facial models. To improve the
performance of the MIHT, the ICPatch is used to fine tune the
estimated transformation parameters and corresponding
elements in the involved facial models by using only matched
point-patch pairs obtained from MIHT. This algorithm does not
deform the surfaces and can perform registration and matching
in a one-step procedure.
3. EXPERIMENTS
The proposed approach employs a low-cost imaging system to
capture overlapping imagery, which are then used to derive a
3D facial model. The generated 3D model is then registered
and matched with available 3D models in a central database for
personal verification or identification purposes. A Canon EOS
Digital camera (8 mega pixels; pixel size: 6.5 micrometers) was
used in this study to capture images of two persons. The camera
was accurately calibrated, checked for stability and mounted on
a tripod. A test field with a set of 3D points which were
previously measured was constructed to compute the position
and orientation of the camera at the time of exposure. To reduce
the motion caused by human operation, image capturing was
remotely controlled. A pattern was projected onto the face
during image acquisition by using a Sony projector to enable
easier identification of a dense set of points. For each person,