International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B3. Istanbul 2004
Input of image data and approximate EO of
images and creation of image pyramid
Extraction of feature points at top-level of
image pyramid
Y
Image matching of feature points at top-level
and point tracking by LSM
v
Bundle adjustment at intermediate-level and
creation of DTM at this level
Extraction of feature points and matching of
points at intermediate-level
v
Point tracking by LSM and removal of
shadow points
v
Final bundle adjustment with extracted tie
/pass points
Figure 1. Automatic Generation of Tie/Pass Points in Z/I
Imaging’s ISAT
Automatic generation of image tie/pass points consists of three
major steps: (a) feature extraction and matching at top-level of
the image pyramid, (b) bundle adjustment at intermediate-level
of the pyramid for obtaining better EO parameters of images
and generation of a rough DTM, and (c) feature extraction and
matching at intermediate-level with computed EO and DTM,
and point tracking by LSM. The purpose of performing feature
point matching at the top-level is to get better EO parameters of
images and create a DTM for better image matching at the
intermediate-level. The use of DTM created from the matched
points can increase matching speed and improve the reliability
of image matching at low levels of the pyramid, especially for
mountainous areas. In ISAT, feature points are extracted by
using the Forstner operator (Forstner and Giilch, 1987), a well-
known interest operator in photogrammetry. It extracts points
by examining the gradients of image intensity values around the
points. By using the Forstner operator, most distinct points in
the image, such as corners, can be extracted. Some shadow
points may also be extracted at the same time. Some shadow
points may survive from image matching and may reduce the
reliability of final bundle adjustment, especially when strips in
the block are very long. In order to improve the reliability of
derived EO parameters of images, a process for removal of
shadow points has been added in image point tracking in ISAT.
3. AUTOMATIC DETECTION OF SHADOW POINTS
Shadow points have some distinct properties in the image,
which differentiate them from other features. The most
distinctive properties are: (a) the image gradient around the
shadow edge is usually large and (b) there is an abrupt change
in elevation in the surrounding area since shadow is cast by
high objects such as trees or buildings. A method for detecting
shadow points in images has been developed based on these
properties.
3.1 Generation of Candidates of Shadow Points
The shadow points are detected in two steps and the first is the
selection of candidates of shadow points. Since shadow usually
has larger intensity value than its neighbouring objects, points
on the edge of shadow have large intensity gradients. The
gradients in four different directions are computed as shown in
Figure 2 and the maximum is chosen as the gradient of the
point.
M
Figure 2. Intensity gradients in x, y and two diagonal directions
G = max { G; }, 1 = 1, 2, 3, 4. (1)
i in the above formula represents direction the intensity gradient
is computed and includes x , y and two diagonal directions.
Once the intensity gradient of a feature point is calculated, it is
compared with the given threshold and the point is selected as a
candidate of shadow points if it is larger than the threshold.
3.2 Detection of Shadow Points
Shadow is cast by high objects such as trees or buildings. There
is an abrupt change in elevation around trees or buildings that
cast the shadow. Thus, shadow points can be detected by
detecting the change of terrain slope in a local area around the
selected candidate points.
3.2.1 Generation of Local Digital Surface Model (DSM):
Before the generation of a local DSM, an area around the point
should be defined on the overlapping images. The area should
be big enough to cover the object casting the shadow. Once the
area is defined, all image points in the area are matched by
image matching. Since the feature points in the area may not be
dense enough, an area-based image matching is used. The
elevation of the matched points is then computed by forward
intersection with the EO parameters of the images from first
bundle adjustment.
3.2.2 Fitting A Surface to Local DSM: After a local DSM is
created around a feature point, a surface is fitted to the created
DSM. It is assumed that the terrain surface in a small area is a
smooth surface in most cases and there is an abrupt change in
elevation in the area where breaklines, trees, buildings, etc.
occur, as shown in Figure 3. In this study, a first-order surface
is fitted to the local DSM by using the least-squares method.
International Archiv
THICI Gear CH ET
The standard deviat
and the fitted surfa
change of terrain sl
is larger than the &
there is probably a
removed as a shado
Eleva
The developed appr
ISAT and some tes
project presented in
Figure 4. Automatic
extracted t