1e entire proc-
ference frame
are fully auto-
iages (Nieder-
| the internal
ined with the
e points have
mined with a
ed targets
e images of the
the acquisition
e least squares
f area around a
ite and the oth-
| image is mod-
ation, rotation,
are varied by
algorithm finds
d of the select-
' the sum of the
levels in these
> least squares
xels. The black
white box rep-
search image.
gorithm
jendently from
nts of the proc-
ess, approximations for a few corresponding points (about
10) have to be manually selected in the five images. A least
squares algorithm is applied to find their exact location in
the pictures. Starting from the seed points, the stereo
matcher automatically determines a dense set of corre-
sponding points. The process is done separately for the left
and right side of the face. For the left side the images taken
by the cameras 1, 2.3 are used and for the right side the im-
ages taken by the cameras 3, 4, 5 (Figure 1). The images 2
and 4 are used as template images. The stereo matcher
searches the corresponding points in the two search images
(1 and 3 for the left side, resp. 3 and 5 for the right side)
independently. At the end of the process, the data sets are
merged to become triplets of corresponding points (points
matched in the three images 1, 2, 3 for the left side, resp. 3,
4, 5 for the right side). The resulting set of corresponding
points is a mixture of triplets and stereo pairs, because of
the presence of points matched in two images only (1 and
2, 2 and 3 for the left side, resp. 3 and 4, 4 and 5 for the
right side). The process can therefore be defined as quasi-
multi-image-matching (but without geometrical con-
straints).
To define the regions between the different seed points, a
Voronoi tessellation is done in the template image. The
picture is divided into polyhedral regions according to
which of the seed points is the closest (Figure 8). The
boundaries are perpendicular to lines joining pairs of
neighbouring seed points.
Fig. 8: Seed points Voronoi tessellation
of the template image
The search strategy of the stereo matcher is the following:
the process starts from one seed point, makes a horizontal
shift in the template and in the search image and then the
least squares algorithm is applied in the shifted location. If
the quality of the match is good, the shift process continues
horizontally until the boundaries of the region are reached.
The entire polygon region of a seed point is covered with
subsequently vertical and horizontal shifts (Figure 9). If
the quality of the match is not satisfactory, the algorithm
works adaptively by changing parameters (e.g. smaller
shift, bigger size of the patch). The normally used value of
the shift is 1 pixel but it can be defined as a subpixel value
in the cases where the match has not given satisfactory re-
sults.
© —— 6 —— 0 — 0 N scgión
boundary
8 <— © ——— ® —— 6 — 6 —0 x
N
© <— 6 «4— 0 -«4—()—»- 0 —»- 0 —» 0
© -«4— 0 -«— 6 —»- 6 —»- 0 —- 06
O seed point
Fig. 9: Search strategy for the establishment
of correspondences between images
The search process is repeated for each seed point region
until the whole image is covered. At the end of the process
it is possible that holes of not analysed areas do appear in
the set of matched points. The algorithm tries to close these
holes by searching from all directions around.
Different tests have shown consistent results in the match-
ing process: the mean number of matched points on the
face is about 15000 and the mean precision of the match is
about 0.05 pixel in x- and y-directions in the picture.
2.4. 3-D model of the face
The 3-D coordinates of the matched points are determined
by forward intersection using the calibration results. The
results show a mean standard deviation of about 0.3 mm in
the sagittal direction and about 0.1 mm in the lateral direc-
tion, which corresponds to about 0.2 pixel in the image.
With the set of 3-D points a triangulated surface is then
generated using the Delauney method (Figure 10).
Fig. 10: Triangulation mesh
405
