International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B2. Istanbul 2004
dkirc
- Yokohama (bottom) dataset.
Parameters Waldkirch Yokohama
Coord. system WGS84 Japan. grid
Acquisition date May 2002 October 2003
Camera SPI SP2
No. strips [paral./cross] 4/2 3/0
Ground pixel size [m] 0.20 0.20
Sensor pixel size [um] 6.5 6.5
Radiometric quality Good Average-poor
No. GCPs 8 5
o0 [uum] 2.5 7.2
Flying height [m] ~2000 ~1944
Table 2. Acquisition and bundle adjustment parameters.
2.2 Systems
In SS, the adaptive method or AATE (Adaptive Automatic
Terrain Extraction) was used. Adaptive matching can use more
than two images, can generate regular grids or triangulated
irregular networks (TINs), changes some of the strategy
parameters based on an “inference” engine, and computes the
mean terrain inclination in small neighbourhoods. Based on this
inclination and image exterior orientation the two best ones out
of all available images are selected. This selection is preferred
(e.g. Bacher, 1998, Baltsavias et al., 2001) and can lead to better
results compared to the non-adaptive as problems due to
occlusions and large perspective differences can be reduced by
an appropriate choice of images. In some cases AATE produces
severe errors at image borders, i.e. the terrain is flattened. The
matching method, utilized in SS, uses area patches, which lead
to smoothing of surface discontinuities. The TIN method is
inherently based on the grid matching approach utilised in SS
(no interpolation is performed at the last stage, for grid points
that have not been successfully matched).
The AIM method is based on a combination of area and feature
based matching techniques. Different types of primitives (area
patches, single edgels that belong to contours, edges) are
combined based on the type of the terrain (rugged, steep, flat).
However, since AIM is still an experimental system under fine-
tuning, several parameters are set by the user, according to the
area and terrain type. The description of the algorithmic
approach exists already in the literature (Pateraki and
Baltsavias, 2003b), and below only a brief overview of AIM is
given. Two types of matching strategies can be utilized, namely
single and multi-template strategy. The first is applied in case of
relative flat terrain, whereas the second in more complex areas.
Multi-resolution levels are employed in a doublet approach
(Pateraki and Baltsavias, 2003a) in order to acquire approximate
values. More than two images are matched simultaneously,
geometrical constraints are enforced by means of quasi-epipolar
curves, and 3D position is computed only from the good rays,
following correlation and blunder detection. In the upper levels,
a surface approximation is derived by matching of grid points
(favorable for faster processing) and which is subsequently
refined in the lower levels by inclusion of linked contour points.
Initial positions at each level are derived by a multi-patch
approach, utilizing cross-correlation and three masks of
different size. Least squares matching (LSM) with geometrical
constraints is further used for verification and refining the
matching solution and is applied for straight edges and single
points (edgels and grid points). The main reason for extending
LSM to straight edges is to improve modeling of discontinuities
and minimize surface smoothing (Pateraki and Baltsavias,
2004).
As automatic matching in each system is based on different
strategies, the assessment is focused on the quality of the final
product, the DSM respectively. Alternatively, an analysis on a
different level, namely forcing the systems parameters to be
relatively similar, would not be realistic for SS as it has certain
limitations for full control of the matching strategy and blunder
detection. For AIM, modifications would be feasible in terms of
implementation, to a certain extent (to adapt some of its
parameters to the ones of SS, e.g. using area-based grid
matching). However, this would be less favorable as the AIM
method takes into consideration several characteristics of
ADS40 (Pateraki and Baltsavias, 2003a) and uses different
primitives for an optimal matching strategy, in contrast to SS. In
both cases, the pyramid levels and the initial mask sizes have
been set to equal values and the same number of images has
been used. The three stereo panchromatic channels and the
Green channel taken out of one strip have been used as input in
all systems. In SS, the TIN version without additional filtering
(elimination of tress/buildings/other objects) has been used.
Similarly for AIM, additional smoothing has been excluded and
raw matched data have been used in the analysis (irregularly
distributed points). Table 3 lists the basic strategy parameters
used in each system.
Parameter SS AIM
Primitives Area patches Area patches/
contour points/
edges
No pyramid levels/ 6/8 6/6
matching passes
No of images 4 4
Type of matching Image matching Image matching
using two "best" all images
images simultaneously
Table 3. Matching parameters.
2.3 Reference Data
In order to check the matching accuracy, reference datasets were
derived from ADS40 images. Mass points and breaklines have
been manually collected in stereo mode in SS with an estimated
404
Intern
accura
Yokot
for th
measu
groun
points
close
arising
model
extrac
or ma
Points
corres
areas
was d
impre:
Morec
datase
the sh
severa
degrac
Road
Jess |
buildii
separa
3.
DSM
shado
Wald
region
(a)) w
existe
good 1
includ
of tre
region
includ
discor
radior
reliabl
The p:
the ch
22, |
combi
W105
was u
the gr
densei
the de
the st
strate£
W105
pyram
interv.
levels
only i
The re
areas,
' The 1
region
Waldk
that th
the are
