In: Wagner W., Szekely, B. (eds.): ISPRS TC VII Symposium - 100 Years ISPRS, Vienna, Austria, July 5-7, 2010, IAPRS, Vol. XXXVIII, Part 7B
A METHOD FOR ROBUST EXTRACTION OF CONTROL POINTS ON
HIGH-RESOLUTION SATELLITE IMAGES
J. Gonzalez“, V. Arevalo’ 1 '“, C. Galindo“
“Dept, of System Engineering and Automation, University of Malaga,
Campus Teatinos, 29071 Malaga, Spain - (jgonzalez,varevalo,cipriano)@ctima. uma.es
Commission VII/7
KEY WORDS: Geometry, Extraction, Registration, DEM/DTM, Imagery
ABSTRACT:
This paper presents a procedure to robustly distribute control point (CP) pairs in high-resolution satellite images as a preliminary step
for accurate image registration. The proper distribution of the CPs is achieved by means of a quadtree decomposition of a coarse digital
terrain model (DTM) of the sensed region. This technique parcels up the image according to its relief variance yielding almost planar
pieces of land. A comer detector is then employed to identify key points in the reference image and an affinity-based feature tracker
that searches for their corresponding comer in the target one. This search is executed in every parcel, selecting (at-least) one CP,
ensuring thus denser distributions in rugged regions than in flat ones. Additionally, robustness to mismatches is attained by exploiting
the intrinsic affine epipolar geometry of the two images. The proposed method has been successfully tested with a broad variety of
panchromatic high-resolution images of the city of the Rincon de la Victoria (Malaga, Spain).
1. INTRODUCTION
Image registration is the process of spatially fitting two images of
the same scene acquired on different dates, from different view
points, and/or using different sensors. Image registration is re
quired in a variety of applications, like, image fusion, 3D scene
reconstruction, and multi-temporal analysis (i. e. natural disaster
monitoring, urban change detection, etc.). See (Schowengerdt,
2007) for a comprehensive survey.
Image registration is typically accomplished by (automatically
or manually) identifying common features, called control points
(CP) pairs, in the involved images. Through such CPs it is pos
sible to estimate the underlying geometrical transformation bet
ween the considered images, which is used to spatially transform
(register) the target image. The accuracy of this process is, then,
strongly tied to:
1. the type of geometrical transformation considered for the
registration (affine, projective, piecewise linear, thin-plate-
spline, etc.), which should account for the relative geometric
differences between the images, and
2. the distribution of CPs over the images, which should take
into account the nature of their differences.
A correct combination of both aspects is crucial to guarantee the
accuracy of the registration: while only two pairs of CPs suffice
to perfectly overlap images of a flat terrain (since they may only
differ in shift, scale and rotation), a large number of them will be
necessary to capture the geometric difference between images of
high-relief surfaces acquired from different viewing angles, re
quiring, also, complex (so-called elastic) transformations. While
elastic transformations have been broadly studied in the remote
sensing field (see (Arevalo and Gonzalez, 2008), for example),
the proper distribution of the CP pairs has not been addressed
indeed. This paper focuses on this issue. *
* Corresponding author.
In the absence of information about the relief of the imaged sur
face, the more effective (but surely not more efficient, see (Fon
seca and Kenney, 1999) for an interesting control-point assess
ment for image registration) approach is the straightforward so
lution of distributing regularly as many CPs as possible all over
the images (Arevalo and Gonzalez, 2008). However, when some
information about the terrain profile is available, a more elabora
ted algorithm can help us to decide their appropriate distribution
on the images.
This paper presents an automatic method to distribute CPs for the
accurate registration of high-resolution satellite images. Exploi
ting the terrain profile information provided by a coarse digital
terrain model (DTM) of the imaged scene, our approach genera
tes a minimal distribution of CPs, achieving significant speedup
in the CPs extraction, without jeopardizing accuracy in the regis
tration.
Our method is intended to be applied to basic high-resolution
satellite imagery, that is, products that are only featured with co
rrections for radiometric distortions and adjustments for internal
sensor geometry, optical and sensor distortions. As the effect
of the terrain is not compensated, two images of a rugged re
gion acquired from different viewpoints may present severe local
geometric differences. Main providers, as it is the case of Geo-
Eye (http://www.geoeye.com, accessed 1 Jun. 2010) or Di-
gitalGlobe (http://www.digitalglobe.com, accessed 1 Jun.
2010), distribute several of these products, as the Ikonos Ortho
Kit, QuickBird Orthoready, etc., which are significatively chea
per than geometrically corrected ones.
The rest of this paper is organized as follows. In section 2, we
describe in detail the proposed method. In section 3, some ex
perimental results are presented. Finally, some conclusions and
future work are outlined.
2. DESCRIPTION OF THE PROPOSED METHOD
The proposed method combines techniques adapted from the com
puter vision field to divide the images according to their estimated