Jeffrey Shan
robust bundle adjustment. The approach intends to reach an optimal solution for automatic exterior image orientation
regarding to precision, reliability and efficiency.
The proposed approach starts with selecting a number of image patches in several
standard locations of the aerial image. In order to ensure high reliability in the
automatic process, each patch contains at least 512*512 pixels; and minimum 3*3
evenly distributed patches are selected (Fig.l). Corresponding patches in the
orthoimage are obtained based on a priori initial exterior orientation parameters or
by manual selection. In this way, the search space and amount of computation are
greatly reduced in subsequent processing. Corresponding patches on aerial and
orthoimage will be conducted with feature points extraction, feature
correspondence, feature matching and precision matching.
[2] [3]
© 2
[9]
Fig.1 Patch distribution
—— |
[] ERI [EE
Fig.2 illustrates the workflow of the proposed approach. As is shown, feature
extraction is respectively conducted on theses selected aerial image patches :
and on the corresponding orthoimage patches. Moravec operator and Fórstner Point feature Paint feature
; extraction extraction
operator [4] are used to extract point features. The number of extracted feature
points in each patch can reach upto a few thousands. Such a great amount of |
observations will substantially ensure the quality of subsequent image
| Feature correspondence |
matching and orientation computation.
Consistence check
Feature matching
aerial image Ortho image
Next, feature correspondence is performed among the extracted feature points
by cross correlation within a bounded search window in the corresponding
aerial image and orthoimage patches. In order to avoid false matching, local
topologic and geometric constraints are introduced in the feature
correspondence step. Experience shows that these constraints are necessary for
obtaining reliable correspondence results. Once the feature correspondence is
established, cross correlation and least squares matching [1] are applied to
Resection
further refine the result. Usually, a large amount of feature points (1000-
2000) can be selected for the final bundle adjustment (see Tab.l in next Fig.2 Flowchart of image orientation
section. ). with orthoimage and DTM
DTM data
Exterior orientation parameters are calculated by bundle adjustment with successfully matched feature points in
preceding steps. Since large amount of matched points are involved in the calculation, robust estimation is therefore
introduced in this step to detect and eliminate remaining false matches so that the quality of this automatic orientation
procedure can be further ensured. The covariance matrix of exterior orientation parameters indicates the ideal accuracy
of space resection. Comparing the result of automatic orientation with the one obtained from sufficient manual
measurements does a further evaluation of the proposed approach.
Upon obtaining the exterior orientation parameters, the 3-D location of an image point can be determined by using the
existing DTM data. This is one of the fundamental issues in ortho rectification and database base revision based on
image features. For this purpose, an integrated algorithm for point determination by using mono-image as well as DTM
data is developed and evaluated. A brief derivation of this algorithm is shown below.
The well known collinear equation is written as
X x Xc
Y |1R| y» HIYe (1)
Z —f Zc
where (X,Y,Z) and (Xc,Yc,Zc) are coordinates of a ground point and the camera center respectively, (x,y) are image
coordinates of the corresponding ground point, f is the principal length of the camera, R is the 3x3 orthogonal rotational
matrix whose elements are determined by orientation angles of the aerial image, 1 is the scale factor related to that
image point.
On the other hand, the given DTM can be expressed in general as
832 International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000.
Ba EN FP me ee a em
213
[I
