623
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B5. Beijing 2008
5. POINT CLOUD PHOTOGRAMMETRY SURVEY:
TERRESTRIAL CASE
It is well known that image matching gives different results,
depending on the texture of the analysed object. For this reason,
a first series of tests was performed analysing building facades
characterised by different textures, in order to evaluate the
strength of point cloud generation even on poorly textured
facades.
Four different facades, made of different materials, were
analysed in particular. In order to make an easy comparison
between the texture of the facades possible, a coefficient was
calculated for each facade. This coefficient was achieved
considering the grey levels image and computing the standard
deviation of 9X9 neighbouring pixels around each pixel, as
shown in the following formula:
where N = dimension of the template: 9 pixels
a x>y = standard deviation of the central template pixel:
N/2+0.5, N/2+0.5
Xjj= grey level value of the pixel
x= mean grey level value of the 9X9 template
In this way a standard deviation value has been computed for
each image pixel. The texture coefficient was finally obtained
computing the mean of these values on the image. This value
was considered representative of the minimum template
dimensions in an Area Based Matching (ABM) approach.
The texture coefficient values of the tested facades are reported
in the following table.
Table 2. Texture coefficients and image examples
According to scientific literature (Kraus, 1993), the base-to-
distance ratio greatly influences the precision of the generated
points: in particular, considering the used camera and the
maximum base achievable by the ZScan System, it is possible
to define the precision that can be achieved at a certain distance.
In order to test the point cloud geometric accuracy, several tests
were performed varying this ratio from 1/4 to 1/18.
The main advantage of the ZScan System is given by the use of
three images in DSM generation at the same time instead of two,
as in other commercial software. In order to quantify the
improvement of multi-image techniques, using the same images,
a comparison was made between a ZScan point cloud and a
DSM generated by LPS (Leica Photogrammetry Suite), using
only two images (e.g. a traditional photogrammetric approach).
Another kind of test considered different rotations between the
image plane (defined by the acquisition bar) and the facades.
Finally, different matching steps were considered in order to
compare the geometrical precision that could be achieved
changing this parameter.
In order to define the precision of the ZScan System each point
cloud generated during the tests was compared with reference
surfaces acquired using a traditional laser scanner. In particular,
a Riegl LMS-Z420 laser scanner was used whose precision is of
about ± 5 mm in range measures. This comparison was
performed using a best fitting approach.
5.1 Results
Texture traditionally represents one of the most difficult issues
in image-matching technologies. Furthermore, the use of more
than two images has not appreciably improved the results: less-
textured areas are difficult to model and are affected by noise.
The previously defined effect is clearly related to the texture
coefficient: image regions characterized by a low texture
coefficient value show large noisy areas, and vice-versa.
Figure 2. Noisy area on a painted wall
If the image triplet has a medium texture coefficient (that is to
say from 6 to 10) the quality of the point cloud is already good.
A generated point cloud is shown in figure 3.
The image triplet was acquired at a distance of 12 meters. The
modelled façade has a texture coefficient of only about 8, but
the geometrical details are correctly represented.
In order to define the change and the loss of geometrical
precision jn point cloud generation, several tests were carried
out over this façade, increasing the taking distance by a metre