Jorge Brito
OCCLUSION DETECTION IN DIGITAL IMAGES
THROUGH BAYESIAN NETWORKS
Jorge Luis Nunes e Silva Brito - Ph. D.
The Military Institute of Engineering (IME), Brasil
Department of Cartographic Engineering - DE/6
jnunes @epg.ime.eb.br
Working Group IV/III.2
KEY WORDS: Occlusion Detection, Digital Photogrammetry, Bayesian Networks, Landscape and Image Maps,
Geometric Modeling and Processing
ABSTRACT
Occlusions are a major obstacle for those dealing with the problem of automatically finding conjugate point pairs in
overlapping images. They lead to mismatches or uncorrelated areas or features in the images. This is particularly critical
in urban areas, where many mismatches generally occur, caused by relief-displacement of buildings.
This research addresses the problem of automatically detecting occlusions caused relief displacement in aerial, central
perspective images.
In a more formal, mathematical approach, an occlusion corresponds to a linear geometric relationship of points in the
image space to points in the object space, involving the image perspective center and two additional points, all three of
which are collinear. The altered order relationship is easily checked in a deterministic approach. However, because
there are errors involved in the computation of the image coordinates of the perspective center and in the projections of
ground points back onto the image space, a simple Boolean test may misconstrue the significance of the apparent order
relationship. These errors or uncertainties in the coordinates propagate along the photogrammetric process and suggest
the use of a probabilistic approach for solving the problem of automatic detection of occlusions in digital images. A
Bayesian Network is suggested, implemented and tested for that purpose.
The tests implemented show how a Bayesian network was successfully applied to the problem of automatically
detecting occlusions in digital images. Three data sets, each of them containing 3D coordinates of point pairs, were
used for testing: stereo-model measurements, simulated data, and digital orthoimage measurements.
1 INTRODUCTION
1.1 Occlusions in Digital Photogrammetry
Occlusions are areas of the earth not visible on remotely sensed images. The occurrence of those areas depends on the
type of remote sensing system being used as well as on the type of energy being recorded.
As far as this research is limited to aerial, central perspective images, occlusions can be understood as areas of the
ground not visible from one or both images (taken from different points of view) of the same geographic area. In those
images, occlusions may occur for two reasons. First, they can be caused by shadows. These shadows can come from
clouds, natural features such as trees, or from man-made features such as buildings. Second, occlusions may also occur
because of the characteristics of the central perspective geometry itself. In the latter case, displacements of ground
features occur along a radial line from the image nadir. These displacements may occlude other features along the same
radial line.
Occlusions are a major obstacle for those dealing with the problem of automatically finding conjugate point pairs in
overlapping images in the context of digital photogrammetry. They lead to mismatches or uncorrelated areas or features
in the images.
The main principle for detecting occlusions in digital images is, at first glance, very simple: an orthoimage should be
free of any parallaxes since it represents the orthogonal (or true) position of a point on the object space. Thus, given a
pair of overlapping orthoimages, any residual parallaxes at a given point on the overlap area may be attributed mainly to
errors in the DEM (which may be caused by occlusions in the original images), or to errors in the orientation parameters
of these images. In fact, a number of investigations have used this principle. For example Doorn (1991) has associated a
non-null matching vector to errors in the DEM. He also defined the geometric condition for the detection of occlusions.
That condition will be used as a basis for occlusion detection using Bayesian networks later in this paper.
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000. 101