International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B3. Istanbul 2004
same sampling frequency of image data with the on-board
GPS/IMU recording time so that cach image line has its exterior
orientation elements. Meanwhile, we will convert these exterior
orientation elements into the mapping coordinate system such
as UTM.
Preprocessing of Image
Y
Generation of Image
<—
Determination of GCPs
*-
Approximate Triangulation
<—
Feature Extraction
<—
Accurate Triangulation
Y
Combination of Feature-based
M atching and Grid-Based
Y
DEM Interpolation
Figure 2. Concept of the algorithm for DEM
generation with PRISM imagery
Also, the removal of stripping and the contrast enhancement
should be performed with great care with respect to possible
changes in the location of edges and other features.
2.2 Generation of Image Pyramids
As conventional aerial images, one scene of PRISM image
corresponds to as much as several hundreds Megabytes data.
For processing and managing this large amount of data
efficiently, coarse-to-fine strategies based on image pyramids
are at the very heart of every modern image matching algorithm.
These strategies also further reduce the necessity for accurate
initial values for the points to be matched. It was therefore
straightforward to adopt a coarse-to-fine strategy for our
approach. Figure 3 shows one example of generated image
pyramids from PRISM simulated image.
2.3 Determination of Ground Control Points
Although on-board GSP/IMU can provide high precision orbit
and attitude data for each PRISM image line, a certain number
of ground control points are still necessary for removal of
systematic errors such as offsets between GPS antenna and lens
center, alignment between PRISM camera and IMU, and drift
errors of IMU in later generalized bundle adjustment.
GCPs could be obtained through ground survey, aero
triangulation, large scale existing vector maps and DEM. 2D (X,
Y) or ID (Z) can be used in our algorithm. It should be stressed,
however, that the resulting absolute accuracy of the object point
coordinates heavily depends on the quality and distribution of
the control points.
2.4 Feature Extraction and Determination of Tie Points
Features in an image are distinct points, edges (or lines), and,
areas. In. our approach, we deal with point features only,
because points can be used as tie points in triangulation and
random points for DEM generation. In this subsection we only
give one method to extract feature points from PRISM
simulated images and the matching algorithm for extraction of
tie points and random points for generating DEM.
2.4.1 Feature Extraction: Interest operators which locate
point features in an image can be found in the literature
(Forstner, 1986). Work on comparing these operators has shown
that the Moravec and the Fórstner operators perform best for
real images in matching applications (Luhmann and Altrogge,
1986). The Fórstner operator has some theoretical advantages
compared to the classical Moravec operator (e.g., rotation
invariance, and the potential for subpixel accuracy). We briefly
describe the Fórstner operator in following and select it in our
approach.
The Férstner operator is based on the summary values of square
gray value differences along the two diagonal directions over a
window of specified size. Using these values we can calculate
the rotation invariance and weight of the window center. Given
a MxN window with center on (i + M / 2, j + N / 2) in an image,
the rotation invariance and weight of the window center are
calculated as follows:
M-IN-I
Nom Y [gi c k 1, j- 1 D - gi * k, je DT
ae
V Y Ig e k 1, jD- gti ek, je E DT
Sens
[g(i+k+1,j+1+1)-g(i+k, j+1)]e
k=[|=|
[gG+k+1,j+1+1)-gG+k, j+1+1)]
uae Vor)
(V, * V)
NV V?
S x V.
M
where, R is the rotation invariance of the window center,
and if the R is larger than certain value it is more possible
to take the window center as an ideal feature point; W is
the weight of the window center, and if the W obtains the
highest value in an area on the image we take the pixel as
feature point . R and W can be used to control the density
of the extracted feature points from PRISM nadir image.
2.4.2 Feature Matching: We used a combined matching
modules which are: (a) cross-correlation feature point
matching based on region growing strategy, (b) grid point
matching based on relaxation, (c) least square matching
with or without geometric constraint, and (d) semi-
automatic feature point matching. The grid matching
International Ai
based on the r
on a PRISM s
direction of the
relaxation ma
single point
function and
procedure. Wi
areas in the it
surface varies
Figu
Feature matchin,
level using the
exterior orientat
based on a few C
was used to pa:
pyramid level fo
our matching ap]
conjugate
candidate
Figure -
To find matches
constrain in the
currently valid or
the search area a
feature point in tl
the conjugate [