International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B5. Istanbul 2004
control information, usually not all the additional parameters
(APs) are recovered.
3.2 Matching process
In order to recover the 3D shape of the static human figure, a
dense set of corresponding image points is extracted with an
automated matching process [D’Apuzzo, 2003]. The matching
establishes correspondences between triplet of images starting
from some seed points selected manually and distributed on the
region of interest. The epipolar geometry, recovered in the
orientation process is also used to improve the quality of the
results. The central image is used as template and the other two
(left and right) are used as search images (slaves). The matcher
searches the corresponding points in the two slaves
independently and at the end of the process, the data sets are
merged to become triplets of matched points. The matching can
fail if lacks of natural texture are presents (e.g. uniform colour);
therefore the performance of the process is improved with
Wallis filter to enhance the low frequencies of the images.
3.3 3D reconstruction and modeling
The 3D coordinates of the 2D matched points are afterwards
computed with forward intersection using the results of the
orientation process. A spatial filter is also applied to reduce the
noise in the 3D data (possible outliers) and to get a more
uniform density of the point cloud. If the matching process
fails, some holes could be present in the generated point cloud:
therefore a semi-automatic closure of the gaps is performed,
using the curvature and density of the surrounding points.
Moreover, if small movements of the person are occurred
during the acquisition, the point cloud of each single triplet
could appear misalign respect to the others. Therefore a 3D
conformal transformation is applied: one triplet is taken as
reference and all the others are transformed according to the
reference one.
Concerning the modeling of the recovered unorganized 3D
point cloud, we can (1) generate a polygonal surface with
reverse-engineer packages or (2) fit a predefined 3D human
model to our 3D data [D' Apuzzo et al., 1999; Ramsis].
3.4 Results of the modeling of a static character
The presented example shows the modeling of a standing
person with a digital still video camera Sony F505 (Figure 2).
Figure 2: Four (out of 12) images (1200x1600 pixels) used for
the 3D static human body reconstruction.
The automatic tie points identification (section 3.1.1) found
more than 150 correspondences that were imported in the
bundle as well as four control points (measured manually on the
body) used for the space resection process and the datum
definition. At the end of the adjustment, a camera constant of
8.4 mm was estimated while the position of the principal point
was kept fix in the middle of the images (and compensated with
the exterior orientation parameters) as no significative camera
roll diversity was present. Concerning the distortion parameters,
only the first parameter of radial distortion (K1) turned out to
be significant while the others were not estimated, as an over-
parameterization could lead to a degradation of the results. The
final exterior orientation of the images as well as the 3D
coordinates of the tie points are shown in Figure 3.
ë & 989
$ .
imma
SAN ES
|
i
|
i
: . : LE bk
Figure 3: Recovered camera poses and 3D coordinates of the tie
points (left). The influence of the APs on the image grid, 3
times amplified (right).
Afterwards, the matching process between 4 triplets of images
produced ca 36 000 2D correspondences that have been
converted and filtered in a point cloud of ca 34 000 points
(Figure 4). The recovered 3D point cloud of the person is
computed with a mean accuracy in x-y of 2.3 mm and in z
direction of 3.3 mm. The 3D data can then easily be imported in
commercial packages for modeling, visualization and animation
purposes or e.g. used for diet management.
Figure 4: 3D point cloud of the human body imaged in Figure 2
pre and after the filtering (left). Visualization of the recovered
point cloud with pixel intensity (right).
is
"m
4. MODELING A MOVING CHARACTER WITH A FIX
CAMERA
Nowadays it is very common to find image streams acquired
with a fix camera, like in forensic surveillance, movies and
sport events. Due to the complicate shape of the human body, a
fix camera that images a moving character cannot correctly
model the whole shape, unless we consider small part of the
body (e.g. head, arm or torso). In particular, in the movies, we
can often see a static camera filming a rotating head. Face
modeling and animation has been investigated since 20 years in
the graphic community. Due to the symmetric forms and
geometric properties of the human head, the modeling requires
very precise measurements. A part from laser scanner, most of
the single-camera approaches are model-based (requiring fitting
and minimization problems) while few methods recover the 3D
shape through a camera model. Our solution tries to model the
head regarding the camera as moving around it. Therefore it can
be considered as a particular problem of the previous case. We
have only to assume that the head is not deforming during the
movement.
An example is presented in Figure 5. The image sequence,
found on the Internet and with a resolution of 256x256 pixels,
shows a person who is rotating the head. No camera and scene
information is available and, for the processing, we consider the
images as acquired by a moving camera around a fix head.