AN ATTRIBUTE-DRIVEN APPROACH FOR IMAGE REGISTRATION
USING ROAD NETWORKS
Caixia Wang, Peggy Agouris, Anthony Stefanidis
Center for Earth Observing and Space Research
Department of Earth Systems and Geoinformation Sciences - (cwangg; pagouris; astefani)@ gmu.edu
George Mason University, Fairfax, VA 22030, USA
Commission IV, WG IV/2
KEY WORDS: Georeferencing, Matching, Imagery, GIS, Feature, Transformation
ABSTRACT:
Geospatial analysis is becoming increasingly dependent on the integration of data from heterogeneous sources. In this paper, we
present an automated, feature-based approach for geometric co-registration using networks of roads (or other similar features). This
approach is based on a graph matching scheme that models networks as graphs with embedded invariant attributes. The main
advantages of our approach reside in its ability of using both geometric and topological attributes to reduce the ambiguity in search
space for inexact matching as well as its invariance to translation, rotation and scale differences (through the use of appropriate
attributes). Furthermore, our approach requires no user-defined threshold to justify local matches. Using the information derived
from this matching process, the registration of two datasets can be accomplished.
1. INTRODUCTION
Geospatial analysis is becoming increasingly dependent on the
integration of data from heterogeneous sources. The geometric
co-registration of these datasets still remains a challenging and
crucial task, especially given the emergence of novel data
capturing approaches, like the use of unmanned aerial vehicles
(UAVs) to capture long image sequences. In this context,
registration may refer either to the registration of images to
images, to generate for example long mosaics, or to the
registration of images to maps, to identify their orientation
parameters. This registration problem becomes increasingly
complex when we consider differences in coverage, scale, and
resolution as corresponding objects in two datasets may also
differ to a certain extent.
Road networks usually are common features in areas of interest.
Unlike points or point-like features e.g. manhole covers
(Drewniok and Rohr, 1996) or building comers (Rohr, 2001),
road networks contain inherently substantial semantic
information in their structure (e.g. topology and geometry), and
thus are considered robust entities for matching in our approach
based on graph matching. A great deal of effort has been
devoted to graph matching by the computer vision community.
In the work of Barrow and Popplestone (1971), relational graph
matching was first studied where a relational graph is designed
to represent scene structure for matching. After that, it has been
widely adopted and developed for matching problems. Two
major approaches can be identified. One involves the
construction of structural graph model where geometric
attributes of components are not taken into consider. Matching
techniques are developed solely based on structure pattern, like
the graph and sub-graph isomorphism approaches (Shapiro and
Haralick, 1985; Pellilo, 1999; Bunke, 1999; Jain et. al., 2002).
The major drawbacks in these graph-theoretical methods are
their computational complexity and inability to handle inexact
matching due to noise or corruptions in the graph. Later works
in Wilson and Hancock (1997) and Luo & Hancock (2001)
exemplify some enhancements based on pure structural graph
model. The second approach to the problem appreciates the
measurements of network components and represents networks
as attributed relational graphs. Matching techniques are
developed to compute graph similarity based on these
measurements and network relational structure, such as
relaxation labelling algorithm (Rosenfeld et. al., 1976; Li, 1992;
Gautama & Borgharaef, 2005), information theory principles
(Shi and Malik, 1998), Markov Random Field method (Li,
1994). In these approaches, invariant measurements of network
components are essential for the matching as they can reduce
ambiguities in local similarity and the corresponding searching
space. But due to different scope of computer vision
applications (e.g. face recognition, content-based image
retrieval) research has addressed geometric and topological
attributes of the network in a rather limited manner, focusing
instead more on performance metrics (e.g. faster convergence).
In this work, we develop an efficient algorithm of inexact graph
matching using invariant attributes (geometry and topology)
included in geographic networks and is based on the relaxation
labelling introduced by Hummel and Zucker (1983). The
challenges we are facing include the computational complexity
of matching network components (i.e. junctions and polygons),
as well as errors in feature extraction due to the presence of
noise in scenes, like building-induced shadows and occlusions.
In this paper, the utilization of point networks and revised
relaxation labelling provides the ability to utilize structures and
geometric attributes derived from the network to improve the
matching algorithm and thus achieve relatively efficient
computation. The process is fully automatic in terms of no input
needed from users. These unique advantages serve both as the
motivation for our work and constitute the main contributions
of this paper.
The remainder of the paper is organized as follows: Section 2
describes the formal representation of road networks in terms of
attributed relational graph. The attributes developed for
193