- 150 - 
acquisition has an important role: direct sunlight creates 
strong shadows, diffuse light is preferable. 
• Orthogonal viewpoint: the quality of texture mapping 
increases if the pictures are taken from a viewpoint 
approximately orthogonal to the object surface. A shallow 
angle will result in a stretching of the texture, which will 
be visible on the final model. 
• Camera viewpoint close to laser acquisition point: It is 
not a requirement that the acquisition points for the laser 
and the camera are the same, but it makes life easier for the 
algorithms and for the user if they are close to each other. 
One problem with different acquisition points is that we 
have different shadows in the 2D and 3D data, so that it 
gets even more complicated to obtain a complete model. 
4. DATA PROCESSING 
This chapter describes the different processing steps that are 
necessary to convert the raw point cloud produced by the laser 
scanner into information that can easily be exploited by the 
user. We divide the data processing in the following steps: pre 
processing, registration, integration, triangulation and texture 
processing (Sequeira, 99). Figure 1 illustrates the workflow 
during data acquisition and processing. 
Figure 1: Workflow for creating a texture-mapped triangular 3D 
model (data acquisition and processing). 
4.1 Pre-processing 
Pre-processing is performed immediately after data capture and 
includes a set of partly unrelated processes, as for example: 
• Normal computation: Using the neighbourhood 
information, which is given through the 2D grid of the 
scan, it is possible to compute the local surface normal for 
each point measured by the scanner. The normal direction 
is very valuable information and can be used later in the 
process for tasks like visualisation (e.g.'backface culling), 
confidence computation, data integration and data 
segmentation. 
• Confidence computation: For various processing steps 
(e.g. registration, integration) it is important to have 
information about the reliability for each 'single 
measurement point. This confidence value depends on 
various factors, as for example the scanner used for 
acquisition, the angle between the local surface normal and 
the scanning direction, the distance to the object, the 
intensity of the returned signal and the variance of the 
measured distance (if the distance is measured as the mean 
of several measurements). A simple way to compute a 
confidence value from the different factors is through a 
weighted average. 
• Edge detection: Edges provide very useful information 
about the scanned object, in particular in architectural 
environments. They also can be used during subsequent 
processing steps, in particular to improve model 
triangulation and model segmentation. Edges can be 
extracted from different source images (range image, 
reflectance image, orientation image, rgb image) using 
standard 2D image processing techniques (Canny, 86). 
Special care has to be taken to achieve edges with sub 
pixel accuracy. One of the problems of laser scanning is 
that the laser will usually not hit an edge directly, but 
slightly next to it thus producing a jagged edge. Applying 
an algorithm that uses sub-pixel accuracy can reduce this 
problem. 
• Noise reduction: The presence of noise in the range data 
is a problem for all laser range scanners. Although the 
absolute noise level varies between scanners, the relative 
accuracy (noise level to maximum range) remains fairly 
constant between scanners. A number of image processing 
algorithms exist that smooth the range data on the 2D grid. 
Simple linear algorithms (e.g. Gaussian smoothing) have 
the disadvantage that they do not distinguish between 
noise and high-frequency information and thus smooth 
edges and details too much. Therefore a non-linear filter 
that respects edges and details is more adequate (Lingua, 
01 ). 
«jcuâ 
• Data clustering: Handling the huge amount of data 
produced by a laser scanner is one of the main problems of 
Figure 2: Two snapshots of the Hamburg data set, which 
contains ca. 30 millions points. The top image shows 
an overview of the data, the bottom image shows a 
zoom into a detail.