In: Wagner W., Szekely, B. (eds.): ISPRS TC VII Symposium - 100 Years ISPRS, Vienna, Austria, July 5-7, 2010, IAPRS, Vol. XXXVIII, Part 7B
300
2. DATA FUSION AND ITS APPLICATION
IN PHOTOGRAMMETRY
In photogrammetry, data fusion is a technique of image
processing for combining two or multi images which
were acquired from an object in different times or from
different locations for achieving accurate output. Data
fusion generally is carried out for change detection,
updating existing maps, or ortho image production. There
are generally three groups of data fusion: image to map,
image to image, and image to database registration.
Morgado and Dowman (1997) carried out an image to
map registration for automatic absolute orientation.
Derenyi (1996) implemented map to image and image to
map registration for investigation on change detection
and Bouziani et al (2010) carried out image to map
registration for change detection. Khoshelham et al
(2010) utilised image to image registration for change
detection, and Suveg and Vosselman (2004) combined
aerial photographs with GIS database for extracting
building form aerial images.
Integrating an image with a laser scanning data is another
technique in photogrammetry for producing ortho-image
or ortho-rectify-image. Laser scanners are able to provide
a fair accurate topographic data along intensity values
from surface of objects that the data can be used for DTM
generation and 3D model. There are various approaches
and techniques for integrating images and laser scanning
data. For example, Iwashita et al (2007) integrated a grey
scale image on a 3D model which was obtained from a
laser scanner using fast marching algorithm, or Zhao and
Popescu (2009) assessed map leaf area index by
investigation the integration of Quickbird image and
Lidar data, or Mizowaki et al (2002) registered a CT
image on MRS image for developing a method of
treatment for prostate cancer. Obviously, there are
numerous studies on registration of an image on a scanner
data for enhancing the image for extracting an object or
producing a map precisely and accurately.
In conventional method of image registration despite of
which mathematical model would be utilised, an image
which was a two dimensional plane would be finally
transferred to another two dimensional plane. The main
aspiration of image registration is to convert an image to
a map or it is better to say to convert a perspective
projection to an orthographic projection. Since the
emergence of digital image in photogrammetry,
enormous approaches on image transformation have been
developed and demand on ortho-image has been
significantly raised. The benefit of using digital image
over conventional film base photos is the flexibility of
digital image. Digital images can be easily stretched,
squeezed, rotated, radiometric and geometric editing, etc.
But it has to be always remember that the output of an
image processing is a two dimensional image. There are
not any defined approaches for omitting or reducing the
distortion from final output. The existing approaches
transfer whole of an image according to a mathematical
modelling which its components are a matrix of rotation
(R), translation vector (V), and scale factor (s). It requires
at least four control points for obtaining the elements of
R, V, s, if R is a 3x3 matrix and V has three components.
The proposal for this project is based on developing a
novel approach in order to register aerial images on laser
scanning data without pre-knowledge about interior
parameters and providing a robust and reliable output
which is free from any distortion. Then the approach has
been extended for registering terrestrial image on a 3D
model. The approach can be easily expanded to register
any images to any data such as DTM, DMS, Digital
Topographic data, and GIS data. According to the
proposal, images have been initially divided to sub area
according to geometry of object and topography of
terrain. Then each sub area will transferred to the host or
source data pixel by pixel. During transformation, pixels
are converted to points which their elements includes
geometric coordinates and intensity values.
Usually a correlation method has been implemented for
matching between two or multi images which were
acquired by similar sensor from an object. Hence, the
camera sensor and laser scanning sensor are different a
new approach has been developed for matching between
the image and laser scanning data and 3D model. The
laser scanners are acquiring data from surface of object
and provided those data in point clouds format; however,
a digital image is two dimensional format consists of
pixels and rasters which included intensity values.
The existing correlation matching for sequence images
has been developed based on comparison between the
gradients of intensity values of a pixel with its
neighbouring pixel on images. Correlation matching has
been recognised as template matching, cross-correlation,
and convolution. In contrast, scanning matching
approaches for matching laser scanning data have been
developed based on presentation of location of points in
the 2D or 3D space and included point to point matching
e.g. Iterative Closest Point (ICP), or feature based
matching, or point to feature matching.
In this study the matching between an image and data
scanner data has been proposed as following: