)JECTION
Calgary, AB,
ejeon, 305-350,
ce high-resolution
quate size that is
n of the physical
terior and exterior
1ercially available
nation process is
on has emerged as
| the mathematical
ion of resampling
x in the resampled
Such requirements
n of such plane to
>, the parameters
concealed by the
other approximate
model, the direct
direct linear
nd Hu, 2001; Ono
rara, 1971).
1 projection model
ing of linear array
sses the epipolar
es the linear array
>] projection model
| Section 3. Section
sampling approach
ntal results using
Finally, Section 6
1s for future work.
S
ng is to generate
ate property that
pipolar resampling
(Cho et aL, 19923.
time of exposure
resampling process
s onto a common
plane, whose orientation is determined by the orientation
parameters of the original images. Moreover, original and
normalized scenes share the same perspective centre. In this
case, an object point, P, is projected onto the left and right
normalized images as p and p', respectively. Although the
normalization process is straightforward when dealing with
frame images, it is not possible using the rigorous modelling of
linear array scanners as will be discussed in Section 2.3. But
first, some terminologies for linear array scanners as well as
their geometric modelling will be introduced.
Right original image
Left original image
Left normalized image
P
Figure 1. Epipolar resampling of frame images
2.2 Differences between Image and Scene
It is important to distinguish between the two terms “image”
and “scene” throughout the analysis of linear array scanners,
Figure 2.
va A
|
x =“
I. ff
(a)
I
, Cotumn 5
! ;
or time
Figure 2. À sequence of 1-D images (a) constituting a scene (b)
An image is defined as the recorded sensory data associated
With one exposure station. In the case of a frame image, it
contains only one exposure station, and consequently it is one
complete image. In the case of a linear array scanner, there are
many 1-D images, each associated with a different exposure
station. The mathematical model that relates a point in the
object space and its corresponding point in the image space is
the collinearity equations, which uses Exterior Orientation
Parameters, EOP, of the appropriate image (in which the point
appears).
In contrast, a scene is the recorded sensory data associated with
one (as in frame images) or more exposure stations (as in linear
array scanners) that covers near continuous object space in a
short single trip of the sensor. Therefore, in frame images, the
image and scene are identical terms, while in linear array
scanners, the scene is an array of cons” «tive 1-D images.
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B3. Istanbul 2004
2.3 Rigorous Modelling of Linear Array Scanners
Rigorous modelling of linear array scanners involves the
scanner interior orientation parameters, IOP, and the exterior
orientation parameters, EOP, for each image in the scene. Lee
and Habib (2002) explicitly dealt with these parameters in pose
estimation problems. Alternatively, other researchers reduced
the number of involved parameters using functional
representations of the EOP (Lee et al., 2000; Wang, 1999;
McGlone and Mikhail, 1981; Ethridge, 1977). The underlying
reason is that EOP do not abruptly change between consecutive
images within the scene. For indirect estimation of the EOP
using GCP, instability of the bundle adjustment exists,
especially for space-based scenes (Wang, 1999; Fraser et al.,
2001). This is attributed to the narrow Angular Field of View
(AFOV) of space scenes, which motivates the investigation of
the parallel projection model. Such a model will be discussed in
more detail in the next section.
Before proceeding with the analysis of the parallel projection
model, one has to note that rigorous modelling of linear array
scanners does not allow straightforward epipolar resampling.
One of the reasons is that it results in non-straight epipolar lines
(Morgan et al., 2004a; Kim, 2000). However, in frame images,
epipolar lines are straight in the original and normalized images
(Cho et al., 1992). For this reason, a different model, parallel
projection, will be sought for the analysis in Section 3.
3. PARALLEL PROJECTION
This section discusses the rationale behind selecting the parallel
projection model. Linear and non-linear forms of the parallel
projection model are presented. Finally, a prerequisite prior to
handling scenes according to parallel projection is presented.
3.1 Motivations
The motivations for selecting the parallel projection model to
approximate the rigorous model are summarized as follows,
Figure 3:
e Many space scanners have narrow AFOV — e.g., it is
less than 1? for IKONOS scenes. For narrow AFOV,
the perspective light rays become closer to being
parallel, Figure 3a.
e Space scenes are acquired within short time — e.g., it
is about one second for IKONOS scenes. Therefore,
scanners can be assumed to have the same attitude
during scene capturing. As a result, the planes,
containing the images and their perspective centres,
are parallel to each other, Figure 3b.
e For scenes captured in very short time, scanners can
be assumed to move with constant velocity. In this
case, the scanner travels equal distances in equal time
intervals. As a result, same object distances are
mapped into equal scene distances, Figure 3c.