ISPRS, Vol.34, Part 2W2, "Dynamic and Multi-Dimensional GIS”, Bangkok, May 23-25, 2001
141
A SELF-ADAPTIVE ALGORITHM OF AUTOMATIC INTERIOR ORIENTATION FOR METRIC IMAGES
1 Wanshou JIANG 2 Guo ZHANG 1 Deren LI
Laboratory for Information Engineering in Surveying, Mapping and Remote Sensing
Wuhan University, #129 Luoyu road, Wuhan, P R. China
Tel: +86-139-71187089, Fax: 86-27-87863229, E-mail: wsiws@china.com
2 School of Information Engineering in Remote Sensing, Wuhan University, PR.China
KEY WORDS: interior orientation, self-adaptive algorithm, softassign, deterministic annealing
ABSTRACT
Automatic interior orientation is one of the key problems in photogrammetric processing of aerial images. In past years, it has been
deeply studied, but it is still worth studying, especially when the images with fiducial marks are not good. In this paper, a new
self-adaptive algorithm is put forward. It adopts softassign and deterministic annealing technology combined with affine transformation
to search fiducial marks in metric images. Even when fiducial marks are not good, the Self-Adaptive Algorithm can still work successfully.
Other strategies such as pyramid, automatic determination of searching range are also taken into consideration.
1.INTRODUCTION
Interior orientation is a fundamental problem in photogrammetry.
Nowadays, most of the commercial software of DPS will be
compelled to apply manual interior orientation if the automatic
orientation is failed, which will make the users feel inconvenient.
Interior orientation is usually referred to reestablishing the
relationships between the pixels and the image coordinate
system. We are concerned with the relationship of a set of
(usually six affine) parameters for the transformation from pixels
to image coordinates [1]. The pixel coordinate system of the
digital image is explicitly given through the matrix of gray values.
However, the image coordinate system is only given by the
fiducial marks. Therefore, the transformation between pixel and
image coordinates by fiducial marks has to be accomplished.
The reestablishment of the interior orientation can be considered
as a pattern recognition problem: one has to find the center of
the pattern representing the fiducial marks and ascribe each
found pattern the correct fiducial number.
Listed in Table 1 are several traditional methods of interior
orientation [2][3][4][5][6], In which, interior orientation are divided
into three tasks:
1) Approximate positioning of fiducial marks
2) Subpixel positioning of fiducial centers
3) Computation of transformation parameters
Among these tasks, approximate positioning is the most
important and the hardest. Gray correlation, binary correlation or
other binary image analysis techniques is used in these methods
to approximate position of fiducial marks. For the images
photographed by RC10, RMK and other aerial cameras, these
methods perform well. The fiducial marks of these cameras are
distributed on the edges of image. No objects are imaged around
the fiducial marks. And so no noise or few noises are assumed
in these methods. The target detected by these methods will be
only one. And the target is regarded as the fiducial marks
wanted.
But in close range photogrammetry, images are photographed
with cameras such as P31, UMK etc. The fiducial distribution of
P31 is similar to Fig. 1, in which the fiducial marks usually merge
in the image of objects. Even worse, some parts of the objects
(such as a mesh) may have the similar shape as the fiducial
marks. Approximate positioning the fiducial marks with these
methods may fail in such cases. If similar objects exist near a
fiducial mark, several targets will be detected. Even if no similar
targets exist, it is very hard for binary analysis to detect fiducial
marks in complex gray images. As for gray correlation, the
coefficient on fiducial can be very small. In out tests, some are
smaller than 0.3. So the orientation of close range
photogrammetry remains a problem.
Fortunately, global image matching technique has achieved
great successfully. In fact, interior orientation can be regarded as
a global matching problem. Firstly we can use gray correlation to
find several peaks of correlation coefficient in the predicted
searching area. And then determining the fiducial marks from the
points of peak is a combination optimization problem. Compared
with other image matching problems, searching for the camera
fiducial marks is relatively simple, and an affine transformation
exists.
There are many algorithms such as genetic algorithm, relaxation
algorithm and Hopfield networks, which are usually used to solve
combination optimization problem. Relaxation algorithm and
Hopfield networks generate local minima and do not usually
guarantee that correspondences are one-to-one. Genetic
algorithm is time consuming. To overcome these problems,
softassign and deterministic annealing technology with affine
transformation have been put forward [7][8]. This algorithm solely
makes use of point location information, but it can supply an
access to one-to-one correspondence and reject a fraction of
points as outliers. In our study, softassign and deterministic
annealing are adopted to solve the problem automatic interior
orientation.
In order to be more efficient in locating the fiducial marks with
less effort, several levels of pyramid images and the original for
each patch can be used throughout the template matching
processing. The relationship of fiducial marks can be used in
determining the searching area of the fiducial marks. When such
an area has been determined, we can select several points
whose correlation coefficients are locally maximal as candidates
of the fiducial marks in the searching area.
Table 1.Methods to automatic interior orientation
Reference
Approximate fiducial positioning
Accurate fiducial positioning
Pose estimation
Kersten and Haring (1995)
Modified Hough transform
Least-squares matching
Depends on camera type
Lue (1995)
Grey value correlation hierarchy
Least-squares matching
Manual
Schickler (1995)
Binary correlation hierarchy
Grey value correlation
Automatic
Strackbein and Henze (1995)
Binary image analysis, no
hierarchy
Fitting of parabolas to gray
value function
Manual