5.1 Rigid Transfer of Seed Points
In the first stereo pair the seed points for matching have to be
measured manually. Then the matching procedure is executed.
This leads to a large number of conjugate points. Because of the
small wave motion the conjugate points of this pair (called pair
[i] in the following) can be utilized as seed points for the
following time step [i+1] (see also Figure 3). In order to reduce
the matching effort only a pre-specified amount of seed points is
used. Well distributed points with a high correlation coefficient
are selected. Then, the matching of the stereo images [i+1] is
carried out, the results can then be used in the same way for the
stereo images [i+2] and so on.
/ X
. ew [ seedpoints
A N
matching ^
| N\ leftfi] - rightfi]
leitüfl| —— 7 rie
emm T
selection
Y
seedpoints
| / matching ^. 1
| eftfi+1] - right[i+T, b Ret
left [it y
right i+ 1}
¥
Zn
[Goniugate points
arit
selection
i
Figure 3. Principle of transfer of seed points
The analysis of image sequences using the rigid transfer of seed
points does not take the movement of the sea surface into
consideration. Thus a number of seed points for matching the
pair [i*n; n21] does not lead to successful results. However this
loss is compensated by the large amount of available seed
points.
5.2 Motion Analysis
It is also feasible regarding image sequence analysis to take the
movement of the sea surface into account, for example based on
a vector field modelling motion from one time epoch to the next
(Jähne, 1991). The vector field provides information about the
direction and velocity of the movement.
5.2.1 Motion Models: Assuming a successful matching
between an image [i] and its successor [i+1] of one image
sequence taken by one camera, a large set of conjugate points in
both images is available. Since these images represent one view
at consecutive time steps, the position difference of any of the
matched conjugate points in image space can be interpreted as a
displacement or movement vector of that point. All
displacement vectors of all conjugate points form the vector
field, which represents a model of measured pixel movements.
Since the time interval between image [i] and [i+1] is not only
the same as between image [i+1] and [i+2] but also very small,
the assumption of a linear approximation of the water mass
movement during the period [i] to [i+2] is assumed to be valid
for a significant number of points. Therefore the position of a
point at time [i+2] can be predicted by extrapolation of the
movement measured between [i] and [i+1] (see Figure 4). This
results in a large variety of potentially corresponding points
between the images [i+1] and [i*2]. Out of them a new set of
seed points is chosen. Then, the matching [i+1] and [i42] is
performed. This yield a new set of corresponding points in [i+1]
and [1*2].
À pre [i+1] :=
j predicted * pos [i] * mov [i-1,i]
{positions [i+1]
| seedpoints | Vet
mov [i] =
pos [i*1] - pos [i] d
4
| matching
motion model [i]
| [fir] |
* | pre [i+2] :=
i | predicted! — pos [i*1] * mov [ii 1]
| (positions [ir2T)
^ image lit]
|
(o . “| seedpoints )
matching | 9
[i+1] ~ [i+2]
mov [i+1] :=
pos [i+2] - pos [i+1]
i
FE
motion model [i1] | |
: E pre [#3] := |
Nt I | predicted] POs [i+2] + mov [i+1,i+2]
à . : | | posit 3r |
Gr magelix2l™- 1 i
Figure 4. Principle of motion models
It is possible to match any image [i+n] with its successor
[i+n+1] as long as the matching with the predecessor [i+n-1]
was successful as well. Combining this knowledge with the fact
that the very first matching step [0] to [1] of this image
sequence can be done with manually measured seed points, it
becomes obvious that all images of this sequence can be linked
together by matching. Therefore a set of conjugate points and
motion models can be derived for that sequence.
5.2.2 Stereo Models: If a matching between a stereo image
pair of time [i] is carried out, a set of conjugate points is the
result. There exists no functional context between this stereo
point set and the sets of conjugate points of the temporal
matching within the left and right image sequence, but both sets
refer to the same image space.
After every motion model calculation a connection between the
corresponding stereo images must be established. One
opportunity to solve this problem is provided by using the
motion model [i, i+1] for a stereo sequence of two images [i]
and [i+1] and the subsequent stereo matching. In both images of
the stereo pair [i] seed points are available from earlier
computations. For the calculation of the vector field of the
motion model in the left and right image sequence the same
seed points have to be determined in the image [left, i+1] and
[right, i+1]. The matching procedures for the stereo pair [i] and
the motion models [left i, i+1] and [right i, i+1] are carried out.
Then, the conjugated points of the stereo pair [i] are transferred
by the use of the motion models into the images [i+1]. A
matching of the image sequence [i+1] is carried out. From the
—598 —
acquir
seed |
detern
step th
Up to
taken
ratio :
before
two ct
Figur:
stereo
meast
points
Usi
poit
sup!
sma
to fi