International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XXXIX-B3, 2012 
XXII ISPRS Congress, 25 August — 01 September 2012, Melbourne, Australia 
  
Figure 2: TLS point cloud 
only 6 images before and after her. After the tie point generation 
we proceeded to the bundle adjustment of the images. The RMS 
of the bundle adjustment was 0.7 pixels. The scale can be in- 
troduced by measuring two points of known distance that can be 
seen in two images. The duration of orientation procedure on a 
Intel 2.83GHz Core2 Quad machine with 4Gb of RAM was 10 
hours. It has to be noted that the extraction of tie points is the 
most time consuming part of the APERO but since it is coded as 
a multi threading procedure it can be accelerated on multi-core 
machines. 
The next step after the orientation of the images was the recon- 
struction of the scene using MICMAC in order to produce sparse 
and dense cloud points of the stairway. We have the option within 
the software of quickly generating sparse point clouds which can 
be very useful to describe our global scene when a dense point 
cloud of the scene is not demanded and to verify that the bundle 
adjustment result is coherent with the reality. The sparse point 
cloud with the camera positions produced by IBM and the TLS 
point cloud of the stairway can be seen on Figures 2 and 3. We 
can see that IBM failed to produce a large number of points in the 
non-textured white area of the ceilings leaving the model open. 
4 RESULTS 
The two point clouds were registered using the ICP algorithm in 
order to be evaluated within the CloudCompare software (©EDF 
R& D). The CloudCompare is an open source 3D point cloud and 
mesh processing and comparing software. The software offers 
various distance measurements between two point clouds or two 
meshes or a point cloud and a mesh. In our case we are inter- 
ested in the distance between two point clouds. In this case the 
software calculates a local model for the reference point cloud in 
order to offer a more local and global precision on the distance 
calculations between the two point clouds. Various models are 
proposed for this local model calculation and we have opted for 
the approach of the height function which offers the best preci- 
sion among the different models proposed. This method initialy 
projects the points on a plane calculated with least squares. Then 
a more accurate locally modelled plane is calculated with the use 
  
Figure 3: IBM sparse point cloud 
of a height function of 
z = ax + by + ex” + dy” + exy (1) 
More information about CloudCompare and the models proposed 
by the software can be found in (Girardeau-Montaut et al., 2005) 
and (CloudCompare, 2012) 
The mean distance between the sparse point cloud and the TLS 
point cloud was 3 cm with a sigma of 7 cm. The maximum dis- 
tance observed was 26 cm (Figure 4). For the dense point cloud 
the mean computed distance was 6 cm with a sigma of 6 cm. The 
maximum calculated distance was 29 cm.(Figure 5). 
Qualitative tests have also been performed in order to better un- 
  
Figure 4: Distance comparison between the sparse IBM point 
cloud and the TLS point cloud 
derstand the problems of our model. To this purpose horizontal 
Figi 
poii 
and 
bee 
geo 
our 
the 
cou 
on | 
the 
on] 
can 
can 
MK 
of tl 
poll 
exte 
cou 
con 
acq 
  
Fig 
TL!