Another big area of application lies in medical research
projects. For ergonomic investigations the ability to
model the spatial movements of feet, ankle and knee
during the process of walking is of special interest. In
facial medicine the doctor wants to be able to analyse the
movements of the upper and lower jaws while the patient
is chewing.
That is why it seems useful to connect the advantages of
high precision spatial measurement of static objects with
the visualisation of movements as they occur in films.
2. SET-UP OF THE SYSTEM
Up to four video cameras (CV-235 monochrome, with a
resolution of 752x582) and optional two pattern
projectors are used in this project in order to register a
dynamical process. The cameras and the projectors are
connected with a frame-grabber (IC-P with AM-CLR from
Imaging Technology). During the recording a pattern can
be projected on to the object (head, knee) so that surface
descriptions can be generated in a fully automatic way by
using matching algorithms. Besides, targets are fixed on
the object in order to provide an easy possibility to detect
identical anatomical points in different images.
Over a certain period of time each of the four cameras
records a sequence of 2D-images (-slices). This
sequence can be interpreted as one 3D-image (= Space-
Time-Cube, STC) with the image co-ordinate axes X, y
and the time-axis T. The STC in figure 1 represents a
hand which is clenched to a fist and then opened again.
Concerning the axes the special meaning of the T-axis
should be taken into account. Different to the Z-axis of
conventional voxel-arrays where all three axes are equal,
the T-axis has a special meaning because the recorded
movement of points can only occur in the direction of the
positive T-axis.
The four STCs can now be used to generate a motion-
model with the axes X, Y, Z and T. The slices at different
moments T = constant can be taken to generate surfaces
using matching algorithms. All these surfaces should
take into account the existence of correlation between
slices adjacent in time.
For the following computation process first each STC is
analysed separately. Then the 4D-points of the motion
model are determined by spatial intersection.
3. COMPUTATION PROCESS
When computing the spatial points of the motion-model
two different methods are being tested.
3.1 Point Tracking method
In the first approach one arbitrarily chosen slice is taken
as initial position. In this slice points of interest are
determined with the help of feature extraction operators
(e.g. Fórstner-Operator) [Fórstner, 1987]. To select
appropriate parameters for the thresholds of the feature
extraction algorithm in a fullautomatical way, mean value
and variance of the specified slice are used.
Next, corresponding points have to be found. This is
done by tracking each of the feature points through the
STC. A least-squares-approach [Jáhne, 1993] is used to
determine the direction to the next corresponding point.
On the assumption that the direction of the gray value
gradient. changes in a local 3D-neighbourhood both
components o1 and o2 of the optical flow can be
determined (Fig. 2).
4 T-Axis
«C
a2 | 2 r
y-Axis T
x-Axis
Fig. 2: Components a1 and o2 describing the local
direction of a moving object element. The right
object element is static.
Fig. 1: The position in the STC of the xT-slice on the left
is marked by the vertical white line in the last
recorded image, while the top face of the cube
shows the yT-slice marked by the horizontal line.
136
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996
The equati
where gx, ¢
values. L
neighbourt
adjusting |
method of
The comp
position of
Then the I
by Least-S
size of the
size of th
variance o
high then
When the
goes back
slices have
In that w
object poir
found. Th
STCs. Th
checked ir
In the cou
curves are
Finally tl
correspon
where the
YT), Z(V)]
3.2 Point |
In the sec
is applied
classify ct
slice are «
The comp
between
motion ci
determine
After this |
following
method. T
the STCs
of the mot
If we mak
methods «
concernin
do not re
each of tl
different p
of the obj
curve hav
determine
the light rz
This con
correspon
introduce
Correspoi