on
1e method is
odeling. The
process the
main idea is
nt to feature
e number of
system. This
letection can
eeds robust
algorithms
d. The edge
atures are
to identify.
| is implied.
atch job and
ems, we are
e amount of
m (Random
features on
m sampling
eding each
uge storage
astic Hough
> lines accu-
re precisely
rresponding
red. The low
t automated
esent a fully
y automatic
to automate
1sed here as
pick up all
| some con-
D-lines, per-
e matching
The 3D-feature reconstruction is based on methods
presented by Mulawa and Mikhail", namely linear
feature based photogrammetry. The approach aims to
use liner features as primitive, so the point to point
correspondence is substituted with feature to feature
correspondence. That means we do not have to
measure image points of a 3D-point, but image points
belonging to a 3D-feature. The feature, in our case line
feature, is identified in multiple frames and bundle
adjustment is performed to solve the camera orien-
tations as well as the feature line parameters. The
feature lines are presented as parametric lines. This
formulation has been used earlier in presenting
geometrical modeling systems in CAD/CAM app-
lications. The photogrammetric presentation binds one
single pixel observation to 3D-feature estimation, so
no line form fitting is needed in 2D space. The line
features are not the only feature type which have this
kind of formulation to present. Also circles, ellipses,
and other conic section curves, as well as splines, have
the photogrammetric relation between image obser-
vations and feature parameters. The lines are though
the most robust features to identify and to extract.
Actually, lines are not able to solve the feature
triangulation alone, also other feature types like
circles have to be included, unless some constraints of
line intersection are determined. In this case,
triangulation is possible with line features alone.
The feature triangulation constructs a stable frame
work for additional measurements. Other features can
be measured from image sequence by directing the
measurements and identifying the feature types by
the operator. After that, the automatic feature
extraction as well as feature matching will do the rest;
find and extract the feature from subsequent images.
This presented system can be adapted when
measuring facades of buildings as well as other objects
which have to be measured precisely. Also e.g. in car
collision tests, the system can be applied in a little
modified form.
2. IMAGE OBSERVATIONS
The three dimensional form fitting can be done by
using pixel coordinates as observations, as well as
points measured with subpixel accuracy. Using rough
pixel coordinates means that we have larger variance
of the observations but if the estimates are unbiased,
results should be the same as when using subpixel
coordinate values. The reliability is based on how
accurately the edge detector can find the real edge and
how invariant it is against noise of the image. The
result of the LSQ-estimation is depending on the
"goodness" of the observations. That means all gross
errors have to be excluded out of the estimation with
some robust way. One way of doing it is to use Hough
transformation which is a very robust method to find
out gross errors and to use it for feature classification.
In this research we have chosen the Random Hough
221
Transformation because of its low computing
consumption and high accuracy. More details of RHT
are given in Chapter 3.1.
2.1 Edge detection
For edge detection traditional gradient operators
(Roberts, Prewitt, Sobel etc.) are adequate if the
images are free of noise. In case of video images and
outdoor circumstances, noise unfortunately is part of
the game. Those gradient operators which indicate the
local gradients, produce a large response for a large
grayscale gradient, where the “Gradient-sum”,
proposed by Rosenfeld', is more immune to large edge
spikes due to the smoothing effect of the summation.
The standard procedure in practice is that before edge
detection some smoothing will be performed for the
image. The Canny operator’ is based on linear
gradient of the input signal with Gaussian smoothing
as an integral part of the operator. This operator is
appropriate for video images which usually need
smoothing before edge detection. With Canny operator
the level of smoothing is determined by the o of the
Gaussian function. The Canny operator is related to
the Laplacian of Gaussian (LoG) operator, but it uses
the first derivate of the Gaussian function when LoG
uses the second derivate. The direction of the gradient
can also be calculated, which might be helpful in the
feature matching stage although the estimates of the
direction are not quite accurate.
Edge strengthening is especially worthy when using
noisy images. This helps finding the final edges by
detecting the maximum of gradients. The automatic
thresholding is often based on the maximum value and
variance of gradients. The thresholding is applied for
extracting all prominent edges and ignoring weak,
noisy edges. Thresholding can be done locally or
globally. Local thresholding means that in a smaller
region the maximum gradient and the variance are
calculated and in this area the threshold value is
based on those indicators.
3. EDGE GROUPING
After finding the prominent edges, edges have to be
grouped together with some criteria. That might be
e.g. the common gradient direction. One way is to use
a line following algorithm. As we are trying to use
linear features to depict the object structure, Hough
transformation is appropriate for the task. Also combi-
nation of these is possible, here we have used edge
linking algorithm and applied Hough transformation
afterwards.
If we consider the 2D projection of three dimensional
features, a space line which is also a line in 2D, a
circle which is an ellipse in 2D and an ellipse which
projection is an ellipse are best features to use respect
to automatic feature classification with Hough trans-
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B5. Vienna 1996
