e position of a 
polation of the 
Figure 4). This 
ponding points 
m a new set of 
1] and [1*2] is 
points in [i+1] 
ovili] - 
jon model [i] 
j 
mov [i,i+1] 
n model [i+1] | | 
i 
i 
{ 
eer 
mov [i+1,i+2] 
s 
| its successor 
ecessor [i*n-1] 
ze with the fact 
of this image 
seed points, it 
> can be linked 
rate points and 
a stereo image 
e points is the 
een this stereo 
' the temporal 
e, but both sets 
on between the 
iblished. One 
by using the 
two images [i] 
both images of 
' from earlier 
r field of the 
ence the same 
[left, i+1] and 
reo pair [i] and 
we carried out. 
are transferred 
ages [i+1]. A 
out. From the 
acquired conjugate points of the stereo matching [i+1] suitable 
seed points can then be selected for temporal matching to 
determine the motion model [i+1, i+2]. In this way at every time 
step the motion and stereo model are related. 
6. RESULTS 
Up to now only first analysis of the normal case images has 
taken place (see chapter 3). Because of the poor base-to-height 
ratio a low accuracy in Y-direction is expected. As mentioned 
before, the accuracy can be improved by additional use of the 
two convergent image sequences. 
Figure 5 shows the manually measured seed points of the first 
stereo pair of an image sequence acquired during the first 
measurement campaign. In this case 30 well distributed seed 
points are sufficient. 
  
Figure 5. Measured seed points 
(above: left image, below: right image) 
Using these 30 seed points approximately 36 000 conjugate 
points are determined automatically. The conjugate points are 
superimposed in Figure 6 to the image in white. The visible 
small gaps are areas in which the matching software was unable 
to find conjugate points. 
  
    
Figure 6. Matching of conjugated points 
(above: left image, below: right image) 
Experiments of processing of image sequences with the rigid 
transfer of seed points have demonstrated the potential of this 
procedure. The analysis of a 10 s image sequence acquired with 
a frequency of 5 Hz has been carried out successfully. Figure 7 
shows a photogrammetrically measured surf zone in 3D at 
different time steps. 
  
Figure 7. Sequence of wave surfaces (At 7 1.7 s) 
When applying the motion analysis procedure the vector field 
covers almost the whole image space except small borders and 
regions of low contrast and areas of highly irregular movement 
(e.g. breakers). Figure 8 represents a clipping of such a motion 
model with typical examples of matched and unmatched 
regions. The lines show the movement compared to the next 
image. 
—599—