the reflectivity value) of the position of the points belonging 
to the same set of points. 
3.3 Data noise reduction 
One of the fundamental operations of the terrestrial laser data 
preliminary treatment is filtering. In fact the data acquired by 
laser scanner devices always has noises which are smaller 
than the tolerance of the used instruments. 
The noise is due to the divergence of the laser beam which 
causes an incorrect evaluation of the distance between the 
object and the origin of the beam. This noise can be easily 
seen if one tries to create a 3D photographic model of the 
object (see figure 4). 
  
Figure 4. Projection of the image on an original 3D model 
(left). Projection of the image on a 3D model depurated of the 
disturbance (right). 
The noisy data do not allow a correct interpretation of the 
object details. In order to obtain a “noise free" model of the 
object, it is necessary to use specific algorithms that are able 
to reduce or eliminate, as much as possible, the acquisition 
errors that can be found in the point clouds. 
The LSR 2004 filtering algorithm was developed by 
exploiting the following principle in a robust statistical 
approach. The point cloud is divided into regular meshes 
according to the horizontal and vertical angular acquisition 
step. The size of the mesh is chosen directly by the operator 
and is a function of the acquisition scan step and of the point 
density one wishes to obtain at the end of the filtering phase 
(the filtering step should usually be equal to twice the scan 
step so that there are at least 4 points in each mesh, the 
minimum possible for a reasonable noise reduction). 
Each mesh contains a set of measured points. The median of 
the distances is estimated and the deviations of the single 
values are computed from their median. 
The distances which have smaller differenced than the laser 
scanner accuracy are used for the estimation of the real 
distance using the mean; the other points are rejected (see 
Figure 3). 
This technique also allows the removal of any points which 
are not on the object of interest (e.g. trees, cars, etc.). 
  
  
  
  
  
  
  
  
  
4. dt. 3. d. dut FW e » 
gross error outliers media outliers grass error 
Figure 3. Evaluation of the outliers through the determination 
of the median value 
The proposed procedure was tested by comparing the laser 
scanner 3D model before and after filtering with a 3D model 
obtained using classical photogrammetric techniques. The 
differences of the laser scanner 3D model (before and after 
filtering) with the photogrammetric model were evaluated. 
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B5. Istanbul 2004 
516 
Figure 5 shows the obtainable results: green, yellow, red and 
black dots represent ,respectively the points where the 
differences between the two compared DEM is less than o, 
2c, 3o and 4c (o is the instrument accuracy). Using the 
original data coming from the laser scanner instrument the 
percentage of yellow points is of about a 30% of the total. 
After the filtering process that percentage rise up to 8094 of 
the total number of acquired points and no red and black 
points can be found. 
Classification E [a7] < & man 
[] smmejazi<témm RE 6mmcjaz<2mm [J 147° 2mm 
  
Figure 5. Photogrammetric DEM (left), Laser and 
photogrammetric DEM difference before filtering (centre), 
Laser and photogrammetric difference after filtering (right). 
The implemented algorithms are also able to remove any 
scattered points that do not belong to the object (vegetation, 
cars in movement, people, urban furnishings etc.). 
  
  
  
  
  
  
  
  
  
  
  
BE 
Figure 6. Example of a 3D model before filtering (on the left) 
and after filtering (on the right). 
3.4 Point clouds alignment and/or georeferencing 
In most cases (when the object has a complex shape) a single 
scan is not sufficient to record the whole object. In these 
cases a series of scans must be performed but each scan has Is 
     
  
   
   
    
     
    
   
   
    
   
       
   
  
  
  
    
   
   
   
  
   
   
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
   
  
     
   
    
   
     
    
   
   
     
Intern 
res 
own re 
the sur 
inaun 
This p 
throug 
corner: 
Once 
param 
be esti 
the cot 
referer 
If this 
an obj 
the fir: 
by at | 
The cc 
inside 
using 
on the 
found 
size O 
scanne 
alignn 
  
  
  
  
  
  
  
The ] 
using 
let | 
reflec 
overl: 
these 
and tl 
The 
right 
referc 
only 
trans|