Beijing 2008
AN AUTOMATIC CONTOUR BASED DETECTION OF TERRESTRIAL OBJECTS
FROM AERIAL IMAGERY DATA
Yuri B. Blokhinov a , Dmitry A. Gribov b , Sergei Y. Zheltov c
State Research Institute of Aviation Systems (FGUP GosNIIAS), Moscow, Russia -
a blokhinov@gosniias.ru, b gda@gosniias.ru, c zhl@gosniias.ru
Commission, WG IV/3
KEY WORDS: Automation, Feature Extraction, Image Segmentation, Contour Matching, Dynamic Programming, Object Detection,
Recognition, Aerial Imagery
ABSTRACT:
The original approach to image matching is proposed. The method itself can be classified as contour matching, based on boundary
pattern analysis technique. Contours extraction technique is reinforced by combined information from independent contours and
boundaries of segmented regions. In this process some general assumptions and simple rules are effectively used. Boundaries of the
regions under investigation are extracted from digital images. Then 1-D representation of contours is constructed for further analysis.
Features of 1-D representation are used for contour matching. For robust extraction of contour features the special procedure, based
on dynamic resampling is used. In order to avoid ambiguities the dynamic programming technique for curves matching was
elaborated. All calculation procedures are invariant to images rotation. Finally, the developed approach is applied to the problem of
automatic image registration for aerial images.
1. INTRODUCTION
Image matching or image fragment localization is the actual and
important problems, aroused in many different applications of
machine vision. Ecological monitoring, forestry, monitoring of
urban areas, automatic search of remote sensing data in vast
geospatial databases - to mention just a few applications leading
to the problem of image matching as a problem of patch
identification inside large images. At the same time, the
situations when accurate information on scale an orientation of
images is not available and distortions are not known a priory
are not uncommon. The authors, based on various ideas,
proposed many approaches to such matching problem but up to
present day, this problem is far from final decision. On the
other hand, the task under consideration should be of interest
for wide range of specialists.
In general, image can be represented as a set of spatially
distributed features. Each feature is described by set of digital
parameters to be unique. In all feature based methods (FBM)
matching is done on extracted features as points, edges, or
regions (Zitova, B., Flusser, J., 2003). In contrast to the area-
based methods, the feature-based matching works with image
information of higher level. This property makes feature-based
methods suitable for situations when illumination conditions or
image geometry changes are expected.
'
Feature based matching procedures consist of three steps:
extracting distinct features (points, edges or regions) in the
images separately,
building up a preliminary list of candidate pairs of
corresponding features based on a chosen similarity
measure,
deriving a final list of feature pairs satisfying the set of
criteria.
branch, many interesting results was obtained (Heipke, C., 1996,
Woozug, C., 1996), each optimal to use in its specific domain.
Here we try to develop some special type of contour based
matching invariant to image rotation and some distortions. This
method was elaborated mainly for automatic monitoring of
large areas from aerial images. The characteristic property of
the method under investigation is that dynamic programming
(DP) technique is used for proper curves matching. All
intermediate steps of the method are illustrated on two pairs of
images, one demonstrating natural objects and one
demonstrating artificial objects.
2. FEATURES EXTRACTION
2.1 Outline of the Registration Method
Among the features used for matching edges (or contours) are
of special interest mainly because of two reasons: contours are
highly stable with respect to various kinds of geometric
distortions and they are natural visible features for most type of
objects in images. In view of difficulties arising in extraction of
contours from real images, the authors tried to develop
approach, combining primary information from both types of
features, edges and regions.
The line features can be the representations of object contours,
coastal lines, rivers and roads in aerial imagery. Standard edge
detection methods, like Canny detector and detector based on
the Laplacian of Gaussian (Nademejad, E., Sharifzadeh, S.,
Hassanpour, H., 2008) are employed for the line feature
detection. However, these techniques usually do not give
contours of required quality for real images because of presence
of discontinuities and considerable dependence on illumination
conditions.
Considerable efforts was done by the investigators in this