ment. As 
his paper 
nted. The 
ciples for 
lysis and 
acking of 
| images 
mages as 
on to be 
surement 
1ages via 
solved. 
urface is 
'eoscopic 
(Linder, 
on cross 
'd region 
a three- 
a digital 
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B5. Istanbul 2004 
The point-wise matching algorithm runs as follows: 
Approximate three-dimensional coordinates for the seed point 
P, the maximum height variation AZ in object space and 
orientations of the images are needed as input data. À straight 
line is then defined through the centre H of the camera base 
C'C"' and the seed point P (see Figure 1). Also the uppermost 
point U, which lies AZ/2 above P, and the lowermost point L, 
which lies AZ/2 below P, lie on this line. Between U and L 
several points are defined in a way that their distance in image 
space amounts to approximately one pixel. 
Using the collinearity equations all these points are then 
projected into image space yielding several point pairs (P', P"). 
Square windows of a predefined size are set up around each 
position in the left image. The window in the right image is 
defined by projecting the four corners of the window in the left 
image into object space and then into the right image, in both 
cases using the central perspective transformation. For each pair 
of windows the cross correlation coefficient p is computed. The 
window pair with the maximum coefficient pg, is considered 
to be the pair of conjugate points corresponding to the point S 
in object space (see again Figure 1), provided that pmax lies 
above a pre-specified threshold value. Otherwise the point is 
rejected. 
In order to exclude incorrect correlations, due to small contrast 
for example, pmax 1s also checked for uniqueness: From the 
neighbouring five correlation coefficients on either side, the 
minimum value Pmin is selected. If the difference between Pmax 
and Pmin is smaller than a value of 0.5, the object point is 
rejected. 
The described principle of point-wise correlation can be 
considered as a variety of the method of vertical line locus 
(Bethel, 1986). The difference is that points are selected on the 
line HP rather than a vertical line through P. 
22 Region Growing 
A three-dimensional point cloud is generated continuing the 
point-wise correlation over the entire model area by region 
growing. Region growing is divided into two parts. 
First, the original images are down-sampled. Then, rays in the 
XY-plane are defined starting from each seed point into the 
eight main directions. Using a constant step size in X and Y 
direction, points on these rays are selected. The step size is 
taken to be equivalent to the grid size of the DSM to be 
eventually generated. Starting from the results of the seed 
points, point-wise correlation on the reduced image resolution is 
carried out for each new point on the rays, always using the Z 
value of the previous point as initial height value. Region 
growing in each direction and for each seed point continues 
until the correlation fails. 
In the next step a regular DSM is interpolated from the resulting 
three-dimensional points, and point-wise correlation is repeated 
for each grid node using the original images. Finally, the results 
are low-pass filtered to eliminate gross errors. 
3. IMAGE SEQUENCE ANALYSIS 
The basic idea of processing image sequences in our approach 
is that in the area of non-breaking waves the change in height of 
the DSM from one image to the next is very small. This value 
obviously depends on the recording frequency and must be 
chosen accordingly. It is then possible to start the process of 
image matching using only a few manually measured seed 
points (Santel et al., 2002). Our method is able to find the 
needed seed points of the following stereo pairs automatically. 
Matching 
   
  
Figure 2. Determination of wave surfaces from image 
Sequences 
In the following the analysis of image sequences is described in 
more detail (see Figure 2). The matching procedure is executed 
for the first stereo pair at time step [i]. This leads to a large 
number of object points. Because of the small wave motion, the 
object points of the time step [i] can be utilized as seed points 
for the following time step [i+1]. In order to reduce the 
matching effort only a pre-specified amount of regularly spaced 
points generated at step [i] is used as seed points at [i+1]. 
Matching of the stereo images [i+1] is carried out. Then the 
results are used in the same way for the stereo images [i+2] and 
so on. 
4. EXPERIMENTAL TEST 
In order to test the described method for the envisaged 
applications we acquired four stereoscopic image sequences of 
the coast of Northern Germany using digital video cameras. The 
selected area is a groyne field seawards Norderney Island. The 
size is approximately 200 by 200 m°. 
The background for the test, and indeed the project, is a 
cooperation with the Institute of Fluid Mechanics and Computer 
Applications in Civil Engineering, University. of Hannover, 
with the aim to test the applicability of digital photogrammetry 
for the application at hand. The project is being carried out 
together with the local administration responsible for coastal 
management and protection.