track horizontally and vertically, and maintain the planned
flying speed relative to the ground.
The digitizing computer is connected to the RDGPS card of
the laptop. The clock of the digitizing computer is synchro-
nised with the GPS signal using a synchronisation card. The
coordinates of the position of the aircraft during the last few
seconds prior to the digitization are stored with correspond-
ing time stamps in the header of each image. Using this in-
formation attached to the images the position of the camera
during the frame grabbing can be estimated with an accu-
racy of 3 — 7 metres. These coordinates are used as initial
values in the global matching, where more accurate coor-
dinates are computed. The in-flight system is shown in fig-
ure 1. Missing from the image are the tilt-sensors giving the
approximate attitude information for each video-image.
RDS antenna GPS antenna
DGPS card and RDS receiver
Laptop with navigation software
operated by the pilot or co-pilot
|
[]
4
[T rris camera
PC with large harddisk, synchronisation card, digitizing card
and software operated by the cameraman or co-pilot
Figure 1: The on-board digitizing and navigation system
3 DESCRIPTION OF THE MATCHING APPROACH
First, a geometric and a radiometric model in object space
are introduced (Figure 2). The geometric model consists of
agrid DEM. The grid is defined in the XY-plane of the object
surface with grid nodes X,, Y, and grid heights Z(X,, Yn
Zi. The mesh size depends on the roughness of the ter-
rain. A height Z(X, Y) at an arbitrary point is interpolated
from the neighbouring grid heights. In the radiometric model
object surface elements of constant size are defined within
each grid mesh. The size is chosen approximately equal
to the pixel size multiplied by the average image scale fac-
tor. An object intensity value G(X, Y) is assigned to each
object surface element. The centre P of each object sur-
face element is projected into the different images using the
collinearity equations. Subsequently image intensity values
g at the corresponding locations x, y in pixel space can be
resampled from the original pixel intensity values.
As the assumptions of constant illumination parameters and
perfect Lambertian reflection are not rigorously met in the
imaging process, a radiometric image transformation T is in-
troduced to compensate at least partially for the deviations.
This simplification does not hold in general, but all image
matching algorithms without prior knowledge about the ob-
ject surface reflectance properties have the same problem.
In the following, the grid heights Z, ;, the parameters p for
the exterior orientation of the images, the intensity values
G(X, Y) of the object surface elements, and the parame-
ters of the radiometric transformation T are treated as un-
knowns. They are estimated directly from the observations
g(x,y) and control information in a least squares adjustment.
Thus, g(x, y) depends on Z, ; and p. For each object surface
332
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996
DEM grid point
with height Z
Object surface
element with
intensity value G
Figure 2: The geometric and radiometric models, and the
transformation from object to image space
element, as many values g(x, y) can be computed as there
are images, and as many observation equations of the fol-
lowing type can be formulated:
v7 G-Tlg(x(Z.p).y(Z.p)). (1)
where
v is the residual of the observation T[g]
G is the unknown intensity value of the object surface
element
T is the radiometric transformation
g is the resampled image intensity value
x,y are the pixel coordinates
Z are the unknown heights of the surrounding grid
points
p are the unknown parameters for the image
orientations
The system of observation equations is completed by intro-
ducing control information with appropriate standard devia-
tions. Since the observation equations are nonlinear in Z
and p, the solution of the least squares adjustment is found
iteratively.
4 IMPLEMENTATION ON PARALLEL HARDWARE
In order to enable the operational usage of the method within
practical computation times the global object reconstruc-
tion algorithm is being implemented as a massively paral-
lel MIMD (multiple instructions - multiple data) computing
application. Due to the complexities resulting from paral-
lelization characterized by scores of tasks interacting with
numerous message types, a need to model and design the
application with a uniform, structured and formal notation
soon became self-evident. The Object Modelling Technique
(OMT) (Rumbaugh et a/. 1991) was chosen for this purpose
because it was seen to be both a powerful general-purpose
modelling technique and particularly well-suited for design
of parallel applications.
In spite of massive parallelization the application has been
designed for portability. This is being achieved by coding
the application with the ANSI-C-language and using only
the widely available public-domain Parallel Virtual Machine
(PVM) library (Geist & Beguelin 1994) as a tool for paral-
lelization. The application has also been designed to mini-
mize both the memory requirements and the communication
betwe
streng
The aj
a TN3
ode, F
68 no
of a si
mance
sor ha:
Mbyte:
In the |
norma
Even if
reduce
unkno
plies a
proble
1:
2,
5
Them
lution :
Here,
the de
The s
dencie
each c
one of
of eac
corres
Here,
plies tl
and th
only re
sweep
a sem
the di:
The s
which
the or
pende
In the
matrix
