for the inclusion of the pre-calibrated interior orientation
parameters of the camera as constraints to the adjustment;
also held fixed are the reference frame co-ordinates.
Average precisions obtained over five full image sets of the
area taken at different times and in different condition are:
Ox oy oX oY oZ
(mm) (mm) (mm) (mm) (mm)
0.002 0.002 0.5 0.5 0.8
4.2 Feature Extraction
To generate a digital elevation model of the surface a dense
cloud of points needs be extracted from the images. Such
points can be found on the basis of changes in the texture of
the image. They are automatically extracted from a "target"
image by means of an interest operator after which conjugate
points on the other images of the same scene are located by
image matching.
4.2.1 Image Filtering: As some of the images may provide
a too uniformly textured surface, in terms of their grey level
contrast, they may need to be enhanced to provide greater
contrast. This is best done using a High-Pass filter algorithm
(Lim, 1984) so as to enhance the high frequency, edge
components, of the image and filter out the low frequency,
uniformly textured components. The result of the filtering
process is an output image which is visibly sharper and
contains greater localised contrast in image texture.
4.2.2 Interest Operators: In the selection of points of
interest, which represent the surface being measured, point
density must be balanced against the high demand on
computational time during the image matching process. The
number of selected points must be sufficient to accurately
represent the surface while avoiding unnecessary point
density leading to unacceptable computational times. Original
tests showed that unwanted point accumulations can occur in
local areas as a result of unsuitable interest operators. In order
to optimise this process three interest operators were
investigated.
Firstly, the Canny Interest Operator (Canny, 1986) uses an
approximation to the first derivative of the Gaussian function
to generate a convolution kernel. The Canny kernel is
convolved in both x and y-directions in the image. Due to the
nature of the Gaussian function, the Canny operator has a
smoothing effect that tends to eliminate noise and low
magnitude edges in the image.
Secondly, the Sobel Interest Operator (Haralick, 1992) is
of similar form to the Canny operator, but does not provide
any smoothing. The Sobel filter is unidirectional in two
dimensions and is convolved in both image directions.
Lastly, the method of determination of the Maximum
Gradient (van der Vlugt, 1994) makes use of the
neighbouring pixels surrounding the pixel of interest to locate
the magnitude and direction of the maximum gradient (edge
vector) at the pixel of interest. Figure 2 gives a diagrammatic
description of the algorithm used for the method of maximum
gradient location.
731
1 2 3
4
4 « »6
wv
7-[,841..9
Figure 2 - Maximum Gradient Filter
(2, -8,)
dist(iÿ)
where
gradient = i=1-4, j=9-6 (1)
The edge information derived from the convolution of the
images with the operators provides edge strength for all three
cases, where an user-selected threshold value rejects weak
edge points. For the Canny and Sobel operators, edge
directions are found by forming gradient vectors from the x
and y convolution values for each selected edge point, while
the maximum gradient operator directly provides edge
directions to the nearest 45 degrees. Selected edge points
form the centre of resampled linear pixel arrays for subpixel
edge detection based on the preservation of moments
method (Mikhail, 1984).
The results of the application of the three edge operators to
the rock surface images did not make it possible to
conclusively select one as most suitable for the surface
texture of the stope face. Local variations in the structure of
the surface responded differently to the three operators and
no systematic behaviour could be established. It was therefore
decided to apply all three operators to each of the images and
to rely on inspection and point count to select the best
resulting point cloud.
4.3 Image Matching
Once the points of interest have been detected in the "target"
image, they need to be "matched" with corresponding points
in the "search" (conjugate) images in order to calculate their
object space co-ordinates. This is the most computationally
extensive and complex of digital photogrammetric tasks and
warrants closer investigation.
Image matching has been approached in two distinct but
interlinked processes. Epipolar geometry supplies estimates
to the initial image co-ordinates of the corresponding points
in the conjugate images (Wong, 1986). This is followed up be
least squares, grey-scale matching (Gruen, 1988) to
determine the final matching image positions.
4.3.1 Epipolar Geometry: In order to find the corresponding
points in conjugate images with no a-priori knowledge of the
search patch positions, epipolar geometry is used to initialise
the search. Figure 3 below shows the basic principle of
epipolar geometry as a tool in image matching.
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B4. Vienna 1996
