The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B5. Beijing 2008
new algorithms is the integration of photogrammetric equations
with other devices commonly used in the Computer Vision
application field such as rectification (Zhang et al., 2003; Du, et
al., 2004) and tensorial notations (Roncella et al., 2005; Hartley,
et al., 2000). The practicality of these instruments is essentially
that they simplify and speed up the photogrammetric process or
provide traditional techniques with information obtained using
new techniques such as the shape from shading (Fassold et al.,
2004) where the texture is not sufficient for matching
techniques.
Thanks to these algorithms, over the last decade some 3D
modelling software packages have been created: these new
instruments can be divided into semi-automated and automated
software.
Other papers (El-Hakim et al., 2007) have already described the
performances of human interactive (semi-automated) software
and demonstrated the great potential of these techniques.
The main aim of this paper is to describe the tests that have
been carried at the Politecnico di Torino on the new Menci
Software ZScan, one of the first fully automated multi-image
matching commercial software programmes. This software was
first devised to allow inexperienced users to generate very
dense point clouds from image triplets; a customized calibrated
bar (0,90 m length) is used for image acquisition. Recently, a
new version of ZScan has been created to elaborate triplets of
oriented aerial images and generate DSM.
Some tests are described in the following sections with the aim
of demonstrating the advantages of using a multi-image
matching approach compared with traditional stereopair
management, using ZScan software both for terrestrial and
aerial applications.
The principal goal of these tests was to understand whether a
multi-image approach could produce a point cloud with the
same precision and density as LIDAR point clouds.
3. POINT CLOUDS MANAGEMENT
Once a point cloud has been acquired, many automated and
manual interventions have to be applied in order to segment,
classify and model the surveyed points. The main topic is the
extraction of the breaklines.
Different ways of solving this problem have already been
proposed in scientific literature. In photogrammetry one way of
solving this problem is represented by the extraction of edges
from images and then their matching in the space using
different algorithms (Hauel, et al., 2001; Zhang, 2005).
A segmentation in LIDAR applications has instead only been
performed using the point cloud information as curvature
(Beinat et al., 2007) or static moment (Roggero, 2002).
However practical experience has shown that the use of images
can be of help for breakline extraction, so all LIDAR surveys
are usually integrated by digital images recorded during LIDAR
acquisition. If the images are oriented in the LIDAR coordinate
system, it is possible to integrate LIDAR point clouds and
derive automatic segmentation algorithms to find the breaklines.
If point clouds are generated by means of a photogrammetric
approach the breaklines can be directly plotted from
stereomodels.
A possible research project is described in the last section with
the aim of transferring useful information about breakline
location from images to point clouds in order to derive the
segmentation algorithms automatically.
4. ZSCAN SYSTEM
The ZScan system was originally thought up to allow
inexperienced users to easily generate point clouds from image
triplets, especially in architectural and cultural heritage surveys.
In order to reach this goal, ZScan Software needs to have three
images acquired using a 0.9 m long calibrated bar. The
maximum baseline between adjacent images is 0.45 m.
The image acquisition is performed acquiring each image and
then translating the camera to the following position (figure 1).
Finally, the acquired images are processed by the ZScan
Software, just defining the area of interest on the images and
the dense matching step. The computational times are usually
moderate as ZScan takes about 30-45 minutes per image triplet
with a matching step of 3 pixels.
Figure 1. The ZScan calibrated bar
In our tests, a Canon EOS 5D camera was used as it guarantees
good image definition. The technical characteristics of this
camera are reported in table 1.
The ZScan software has been modified to also manage aerial
images and, in a more general approach, each kind of
normalized triplet of images. The absence of a y-parallaxes
constitutes a constraint for the correct use of the software. It is
well known that the relative orientation and the normalization
of a set of images is an autonomous procedure in digital
photogrammetry.
Camera
f
[mm]
Pixel dim.
[mm|
Image dim.
[pixel x pixel]
Canon EOS 5D
25,10
0,0082
4368 x2912
Table 1. Technical characteristics of the camera