The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B5. Beijing 2008 
625 
6. POINT CLOUD PHOTOGRAMMETRIC SURVEY: 
AERIAL CASE 
The ZScan System was originally realized for terrestrial 
applications, using a calibrated bar. The first test results using 
aerial images are presented in this paper. 
An image triplet acquired over the city of Turin (Italy) was used. 
These images were acquired using a Leica RC 30 camera at a 
1:8000 nominal scale in 2004 and they were later scanned at a 
1200 dpi resolution. The achieved ground sample distance 
(GSD) was about 20 cm. The image overlap was 80% and the 
external orientation parameters were known from previous 
works: in some ways, these images can be considered as having 
been acquired by an extremely long calibrated bar. 
The adopted matching step was of 5 pixel, that is, 1 pt/m 2 of 
resolution. 
In order to define the achieved precision, the generated point 
cloud was compared with a DSM whose precision is known (± 
10 cm). This DSM considers roofs as flat surfaces and does not 
describe any tree or vegetation, so an error in correspondence to 
the roof tops and trees can be expected. 
An overview of the ZScan generated point cloud is presented in 
figure 6 and a zoom of the Olympic Stadium in Turin is shown 
in the top right panel. The point cloud seems to be complete and 
blank areas are limited to hidden regions. 
Figure 6. Point cloud from aerial images 
Furthermore, a DSM was generated using two images a time in 
LPS software in order to estimate the improvement due to 
multi-image algorithms. 
6.1 Results 
In order to compare the generated point cloud and the reference 
DSM, the RSI ENVI commercial software was used. The 
coordinate system adopted in the comparison was the UTM 
WGS84 reference system. 
The comparison results are shown in figure 7. As can be seen 
the ZScan systems offer a good precision over flat areas, as the 
differences are usually less than 1 m (green area). Furthermore, 
differences of between 1 m to 5 m (yellow area) are 
concentrated on roofs and on wooden areas but, as known, the 
reference DSM considers roofs as being flat and does not 
describe trees or vegetation. 
As an example, the mean differences in the violet area are 0.31 
m and standard deviation is 0.49 m: in other words, residuals 
have the same order of magnitude as GSD and they are 
comparable to the reference DSM precision. 
However there are still prominent differences around buildings 
(red area), in correspondence to breaklines where the 
differences are usually above 5 m and holes in the point cloud 
are concentrated. The blank areas can only partially be 
explained by the shadows around buildings. The black panel 
(figure 6) shows residuals along blue lines; from this panel it is 
possible to clearly see the largest negative differences over 
building comers and small residuals in correspondence to 
pitched roofs. Negative differences mean that, in the same 
planimetric position, the point cloud defines points at the roof 
heights while, according to the reference DSM, they should be 
on the ground. In other words, the ZScan System has a sort of 
systematic error in correspondence to the breaklines (border of 
roofs): this problem is probably due to the lack of an efficient 
algorithm for the segmentation of the images. 
Figure 7. Comparison between the ZScan and DSM data 
Figure 8. Source image, ZScan DSM and 2-image generated 
DSM 
The DSM generated using only two images gave approximately 
the same results, in terms of geometrical precision. As in the 
ZScan point cloud building, the borders are hard to detect, but 
flat areas and roofs are correctly surveyed. 
In spite of this, the number of detected points roughly decrease 
when using only a stereopair: there were 1138783 detected 
points in the ZScan triplet while, for the same area, the 
stereopair DSM detected only 432599 points. This difference