2004
M
, See
1ation
n the
more
04). .
laser
:gular
erates
re are
is to
point
econd
e real
or the
2 SCH
where
nding
tional
' and
image
e SCI
. This
led by
th the
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B4. Istanbul 2004
3. MULTISENSOR IMAGE FUSION
In many image processing applications it is necessary to
compare multiple images of the same scene acquired by
different sensors or image taken by the same sensor but at
different times or from different locations. This section
describes the process of matching multi-source data of the
same scene acquired from different viewpoints and by
different. The purpose of multi-sensor image matching in this
paper is to establish the correspondence between RCI and SCI,
and to determine the geometric transformation that aligns one
image with the other. Existing multi-sensor image matching
techniques fall into two broad categories: manual image and
automatic image matching the implementation and results of
manual multi-sensor image matching, which includes, interior,
relative and absolute orientations using two different types of
software for comparison purposes, has been discussed in
Forkuo and King (2003). The manual measurement was
necessary to understand the key issues such as geometric
quality, both spatial and geometric resolutions of the generated
synthetic camera image.
3.1 Automatic Multisensor Image matching
Once the 2D intensity image has been generated from the 3D
point cloud, the location of corresponding feature in the
Synthetic Camera Image (SCI) and the Real Camera Image
(RCI) is determined. The most difficult part of the automatic
registration is essentially the correspondence matching: Given
a point in one image, find the corresponding point in each of
the other image(s). Although the automatic correspondence is
not a problem for vertically oriented images, it is still a
problem in the terrestrial case and it is even much complex in
terrestrial multi-sensor case. It can be observed that, since both
image types are formed using similar mechanisms, the location
of many objects are identifiable in cach image. However, there
are differences in illumination, perspective, reflectance as well
as lack of appropriate texture (Milian et al, 2002) between
these images. Also, images from different sensors usually have
their own inherent noise (Habib and Alruzoug, 2004).
Furthermore, the automatic registration problem can be
complicated, in our case, by differences in image resolution
and scale, and low image quality (especially with the SCD.
One approach to automatically overcome the correspondence
problem is both area and feature based approach was used
(Dias et al, 2002). The first step for correspondence matching
or simply pairwise matching is the extraction of features,
generally interest points from both images using Harris corner
detector. Initial correspondence between these points is then
established by correlating regions around the features. The
similarity is then judged by the accumulated development of
corresponding interest points in the two images (Rothfeder et
al, 2003). We have discussed the matching algorithm which
consists of feature extraction process followed by the cross
correlation matching in Forkuo and Bruce (2004).
3.1.1 Automatic Feature Detection and Extraction
The automatic registration problem requires finding features
(edges, corners) in one image and correlates them in another.
For this paper, Harris corner detector as proposed in Harris and
Stephens (1988) is used detect and extract corners in both
images. This operator has been widely used and it has been
shown to be robust to viewpoint changes (i.e. image rotations
and translations) and illumination changes (Dufournaud er al,
2004; Rothfeder er al, 2003). However, the Harris point
detector is not invariant to changes in scale (Dufournaud er al,
2004. It uses a threshold on the number of corner extracted
based on the image size. The number of corners detected in
images is variable (Rothfeder er al, 2003) and in figure 4, the
two images are shown with the detected corners features. Once
feature points are extracted from image pair, correspondence
matching can be performed.
3.1.2 Correspondence matching
This section concentrates on determining corresponding
between two sets of extracted interest points that were detected
with Harris corner operator. To match these features
automatically, the zero mean normalized cross correlation
(ZNCC) measure, which is invariant to varying lighting
conditions (Lhaullier and Quan, 2000) is used. This method
uses a small window around each point to matched (this point
becomes the center of a small window of gray level intensities),
and this window (template) is compared with similarly sized
regions (neighborhood) in the other image (Rothfeder ef al,
2003). In other words, the ZNCC method is based on the
analysis of the gray level pattern around the detected point of
interest and on the search for the most similar pattern in the
successive image (Giachetti, 2000). Each comparison yields a
score, a measure of similarity. The match is assigned to the
corner with highest matching score (Smith ef a/, 1998).
By selecting a suitable patch size (correlation window) and
threshold for the matching process reduces the number of
detection of false correspondence pairs. However, in our case,
the number of mismatches (referred to as outliers) may be quite
large (as can be observed in figure 5). This occurs in particular
when some corners cannot be matched. Also, there are likely to
be several candidates matches for some corners which are very
similar (Smith ez a/, 1998). These correspondences are refined
using a robust search procedure such as the RANdom SAmple
Consensus (RANSAC) algorithm (Capel and Zisserman, 2003;
Fischler and R. C. Bolles, 1981). This algorithm allows the
user to define in advance the number of potential outliers
through the selection of a threshold. The best solution is that
which maximizes the number of points whose residuals are
below a given threshold. Details can be found in Fischler and
R. C. Bolles (1981). Once outliers are removed, the set of
points identified as inlers may be combined to give the final
solution (RANSAC inliers) and the result is shown in figure 6.
These inlying correspondences are used in the model-based
image fusion.
4. MODEL-BASED IMAGE FUSION
In this context, model-based fusion is the process of
establishing a link between each pixel in the 2D intensity
image data to its corresponding sampled 3D point on the object
surface. The task is to determine the relationship the coordinate
systems of the image and the object by photogrammetric
process of exterior orientation. The exterior orientation process
is achieved in two steps. For the first step, we relate each
matched pixel of the extracted feature in the SCI data to its
corresponding 3D point from the point cloud data using
interpolation constants. That is, the automatic link between the
object coordinate system and the image coordinate system has
been established. This means that the image coordinate, object