The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B5. Beijing 2008 
1104 
imagery from 22m (top view). In both cases, 20 images of 
strong geometry (converging images with different k rotation) 
were oriented using self - calibrating bundle adjustment. The 
results are summarized in Tables 1 and 2, where c is the camera 
constant, (x 0 , y 0 ) the location of the principal point and k b k 2 
the two coefficients of the symmetric radial distortion. 
3.2 Bundle adjustment 
The exterior orientation of the images was carried out with the 
photogrammetric workstations of Leica (LPS v9.2), based on 
automatic and semi-automatic techniques (tie point 
measurement) followed by stereoscopic checks. The aerial 
images of the top view were separated into four big blocks and 
oriented using bundle adjustment triangulation relying on well 
distributed control and check points. Table 3 shows the results 
of two such blocks. 
Block1 
(top view) 
Block 1 
(east wall) 
Number of images 
376 
137 
A posteriori g 0 (pixel) 
0.38 
0.41 
Control 
Points 
RMS X (m) 
0.011 
0.Q09 
RMS Y (m) 
0.008 
0.004 
RMS Z (m) 
0.013 
0.006 
Check 
Points 
RMS X (m) 
0.009 
0.004 
RMS Y (m) 
0.008 
0.006 
RMS Z (m) 
0.010 
0.007 
Table 3. Bundle adjustment results. 
Consequently, one can see that the final accuracy is equal to the 
one specified for the orthophotomosaic generation. Specifically, 
for the side views of the walls, individual well-defined points 
from the intensity maps of the laser scanner were selected and 
used as ground control in the bundle adjustment, ensuring a 
proper registration of the images against the laser data. 
3.3 Range data 
The processing of the point clouds was implemented in the 
software RealWorks (Trimble). As a first step, noise reduction 
was applied to the points. The different scans were then 
registered together and against the geodetic system (GGRS 87) 
using the coordinates of the special targets. The resolution of 
the unified point clouds was reduced, in order to agree with the 
specifications of the work (1cm for the walls and 5cm for the 
rock). The final 3D mesh was produced through a 3D 
triangulation process, while small holes, in the surface were 
corrected automatically employing a hole - filling algorithm. 
Larger gaps in the data were filled with 3D points extracted 
photogrammetrically from the images of the balloon. 
4. PRODUCTS 
The digital terrain model of the top view was generated with 
automatic terrain extraction techniques (LPS and Inpho 
software) at a resolution of 0.02m and 0.01m, for the top view 
and the walls respectively. The results were corrected manually 
using suitable collection techniques (Mavromati et al., 2003), 
regarding breaklines, improving the final quality (Figure 5). 
Figure 5. A detailed part of the 2.5D DSM of the top view. 
Concerning the walls, the surface data come exclusively from 
the laser scanner, apart from the cases where there is lack of 
points (gaps), as it was mentioned before. Additionally, the 3D 
points from the range scanner were checked through 
strereoscopic viewing, as a final check of the quality of the 
registration among the two different data sources: 
photogrammetry and laser scanning. 
An issue arises concerning the generation of the 
orthophotomosaics, where the employment of a specialized 
algorithm is a demand. In cases with strong height variations on 
the ground, conventional orthorectification software may lead 
to unexpected results such as double projections and artefacts. 
The principal aspect here is the proper visibility checking of the 
object surface in the images and the simultaneous detection of 
surface areas occluded in the initial images used. On that basis, 
orthophoto production (Figure 6) was performed with 
specialized true - orthorectification software such as Inpho 
Orthobox (Ortho Master + Ortho Vista). 
Figure 6. Detail of a top view orthophoto. 
In addition, the methodology of Karras et al. (2007) was also 
used, especially in areas where the full 3D mesh (instead of the 
2.5D DSM) could not be handled by Inpho’s programme due to 
occlusions in the direction of the ortho-projection. 
For the several orthophotos of the walls, the projection planes 
of the orthophoto subgroups, each subgroup corresponding to a 
different wall plane, were calculated by plane fitting, using the 
coordinates of the control points. Finally, the 
orthophotomosaics were radiometrically corrected in order to 
create a uniform and homogeneous result.