- 151 - 
laser scanning. For example, current PCs cannot render 
more than circa 1 million 3D points (the number of points 
in one average scan) in real-time. But also data processing 
itself suffers from the huge amount of data. One way to 
facilitate the viewing and processing of a number of scans 
is to cluster the 3D data in hierarchical structures (e.g. 
Octrees or Binary trees). The clustering is done off-line 
during the pre-processing and the structure is then saved to 
disk for later use (Meagher, 82). Figure 2 shows two 
snapshots of a viewer that dynamically loads the points 
required for the current view position based on an octree 
data structure. 
4.2 Registration of range data 
In most cases, it is necessary to acquire data from more than one 
viewpoint in order to cover the complete object. After 
acquisition, range data obtained from different viewpoints needs 
to be transformed into a single reference frame. This process is 
known as registration. Different methods exist to register a set 
of scans; each of them has advantages and disadvantages 
according to the given scene and the required accuracy. This 
section provides a brief overview of the different methods with 
the respective pros and cons: 
4.2.1 Registration using targets 
Targets (planar or spherical) are the standard way to register 
two scans. Each pair of scans needs to contain at least three 
targets, which can be recognised by the processing software. 
Their 3D position is computed in the respective local coordinate 
system and the matching positions are used to compute the 
transformation between the two scans. The accuracy of the 
resulting transformation depends on the precision with which 
the targets are scanned and localised in the scans. Some 
scanners support automatic detection and high-resolution 
scanning of the targets. 
The main drawback of using targets is that they need physically 
to be placed in the scene. Especially for scanners with a small 
field of view the number of required targets can be quite high. 
Often it is difficult or even impossible to place the targets; 
registering scans that have been acquired at different points of 
time (e.g. for monitoring) can be difficult, as the targets might 
have moved. An advantage is that the targets can be surveyed 
with traditional technologies, so that the scans can be reference 
to an external network. 
4.2.2 Feature-based registration 
Another way for registering two scans is to extract parametric 
surfaces (e.g. planes, cylinders, spheres), which are present in 
the scene. A number of objects need to be extracted from each 
scan and matched between the scans. The respective positions in 
the local coordinate frames can be used to compute the 
transformation between the scans. 
Feature-based registration can be useful in industrial 
environments, which contain many standard shapes, i.e. pipes 
and walls. In other environments it is usually not possible to 
reliably find enough features. 
The accuracy of the transformation depends on the accuracy 
with which the features have been extracted from the scans and 
therefore on the precision of the scan data, on the quality of the 
fitting algorithm used and on how well the real object fits the 
parametric surface. A sufficient quality cannot always be 
guaranteed. 
4.2.3 Point-based registration 
A standard method for aligning two scans based on the point 
cloud is the Iterative Closest Point (ICP) algorithm (Besl, 92). It 
requires that an estimation of the relative position between two 
scans is known a-priori, which is usually provided interactively 
by the user. Based on this estimation the algorithm searches for 
each point (or a subset of control points) in one scan the closest 
point in the other scan and uses the corresponding point pairs to 
compute a new relative position between the scans. This process 
is repeated iteratively until the relative position of the scans 
converges. 
The ICP algorithm works well if the scans contain relatively 
large areas of continuous surfaces and have sufficient overlap. It 
faces problems in strongly clustered areas (e.g. industrial plants) 
and when many scans are concatenated (e.g. along a road or 
tunnel) as the error propagates from one scan to the next and 
externally surveyed points are not integrated into the standard 
ICP algorithm. 
4.2.4 Summary 
Different methods exist for scan registration, each suitable for 
particular situations. The main drawbacks of currently 
commercially available systems are that: 
• The registration works on a per scan basis, no global 
optimisation is done when working with a large set of 
scans. Global optimisation (comparable to bundle 
adjustment in traditional techniques) is particularly 
important when using a scanner with a small field of view 
or when scanning an object from outside where it is not 
possible to acquire a single reference scan that covers most 
of the object. 
• The different methods are not integrated. Ideally, a 
registration system should use all available information 
(targets, prisms, externally measured points, common 
features and overlapping areas) to compute a globally 
optimised registration. 
4.3 Integration 
During scan integration the processing software has to remove 
the redundancy that is introduced through overlapping scans. 
Integration can be done on different levels: 
• On a point basis immediately after the registration, i.e. the 
software has to test for each measurement point whether a 
corresponding point exists in a different scan and if so 
whether it should discard the point for further processing. 
• On an object level, i.e. the data is modelled on a scan basis 
(e.g. triangulation, surface fitting, intersection, edge 
detection). Subsequently corresponding objects from 
different scans are merge into single coherent objects. 
• The data modelling is directly done on the combined 3D 
point cloud and thus immediately produces a single 
coherent object from the data. Modelling on the 3D point 
cloud without using the 2D information of the single scans 
usually makes the modelling more difficult, however it 
avoids the problem of merging the objects afterwards. 
4.4 Meshing 
Meshing converts the set of raw 3D points into a continuous 
surface and thus produces a visually more intuitive 
representation (especially when mapped with reflectance or 
texture data) while at the same time reducing the amount of 
data. The triangulated mesh can be used for subsequent