'anbul 2004 
eter check 
ılts the sin- 
he specific 
ameter can 
Due to the 
he demand 
the camera 
e angle (or 
"Wise prior 
| the weak 
guarantees 
the camera 
to impose 
f gross ob- 
able. Fur- 
e measure- 
— 18 advan- 
> recorded. 
/ from Sin- 
iished Dis- 
Y 
5. 
d Structure 
t. Interna- 
cial Intelli- 
| View Ge- 
rsity Press, 
P., „2002. 
ance Envi- 
ice Evalua- 
ypenhagen, 
st Squares. 
ty Press of 
. and Flan- 
Cambridge 
. Tea 
-Calibrated 
‘onference. 
ebraic Pro- 
ik. 
tion to the 
Department 
at Chapel 
SEGMENTATION OF LIDAR DATA USING THE TENSOR VOTING FRAMEWORK 
Hanns-F. Schuster 
Institute of Photogrammetry, University of Bonn 
Nussallee 15, D-53115 Bonn, Germany 
Schuster@ipb.uni-bonn.de 
KEY WORDS: LIDAR, Segmentation, Algorithm, Automation, Modelling, Point Cloud 
ABSTRACT 
We present an investigation on the use of Tensor Voting for categorizing LIDAR data into outliers, line elements (e.g. 
high-voltage power lines), surface patches (e.g. roofs) and volumetric elements (e.g. vegetation). 
The Reconstruction of man-made objects is a main task of photogrammetry. With the increasing quality and availability 
of LIDAR sensors, range data is becoming more and more important. With LIDAR sensors it is possible to quickly aquire 
huge amounts of data. But in contrast to classical systems, where the measurement points are chosen by an operator, the 
data points do not explicitly correspond to meaningful points of the object, i.e. edges, corners, junctions. To extract these 
features it is necessary to segment the data into homogeneous regions wich can be processed afterwards. 
Our approach consists of a two step segmentation. The first one uses the Tensor Voting algorithm. It encodes every data 
point as a particle which sends out a vector field. This can be used to categorize the pointness, edgeness and surfaceness 
of the data points. After the categorization of the given LIDAR data points also the regions between the data points are 
rated. Meaningful regions like edges and junctions, given by the inherent structure of the data, are extracted. 
In a second step the so labeled points are merged due to a similarity constraint. This similarity constraint is based on a 
minimum description length principle, encoding and comparing different geometrical models. 
The output of this segmentation consists of non overlapping geometric objects in three dimensional space. 
The aproach is evaluated with some examples of Lidar data. 
1 INTRODUCTION 
With the increasing quality and availability and falling costs 
of LIDAR-data there is a growing need for automatic de- 
tection and reconstruction of the objects contained in the 
data. A human can easily read the content of a point cloud 
because our brain is highly trained in such context-based 
segmentation tasks, but for automatic reconstruction we 
need to have the location of meaningful features like cor- 
ners, edges or junctions. 
The Problem with LIDAR-data is, that the measured points 
do not have any context information and the grid in which 
they are measured is not oriented on these features. Nor- 
mally the wanted features are only indirectly observable 
e.g. by segmenting two planes and intersecting them. 
In this paper we show the extraction of features like curves, 
surfaces and junctions from a point cloud. therefore we 
present a two-step procedure that uses the tensor voting 
framework as a first step to categorize the input points into 
three types of appearance. In a second step we use a seg- 
mentation to merge the categorized points into curves and 
surfaces. 
The tensor voting framework (Tang et al., 2000) can not 
only be used for handling 2D or 3D (Tang and Medioni, 
1999) data but also to process motion fields (Nicolescu and 
Medioni, 2003) or stereo data (Lee and Medioni, 1998). 
In most cases the input data is of small scale (Tang and 
Medioni, 1998) in contrast to LIDAR-data and the output 
is only used for visualisation in pixel or voxel representa- 
tion (G. Guy, 1997). 
In section two we will have a look on the tensor voting 
framework. In section three we show how the output of 
the tensor voting can be segmented. The results of the 
1073 
approach are presented in section four. In section five a 
conclusion is presented followed by an outlook. 
2 TENSOR VOTING 
The goal of the tensor voting is to extract the structure in- 
herently given in the point cloud. 
The results of the tensor voting process are three coutinu- 
ous vector fields, represented by discrete grid points. The 
scalar part of these fields represent the likelihood of the lo- 
cation in space to be a point, part of a curve, a surface. The 
vector part represents the orientation of the occurence. 
These three fields can be searched through to find maxima 
which represent the most likely location of a wanted fea- 
ture. 
2.1 TensorVoting in physical analogy 
To explain the concept of Tensor Voting with an analogy 
to physics, we can compare the Tensorfield with a physical 
field of force, e.g. a magnetic field. We can imagine that 
the object which is represented in the point cloud has a 
magnetic field. It propagates its field into the space around 
the object. 
If we put iron particles into this field, these are affected by 
the field so that they act as little magnetic dipoles which 
align their field along the field lines of the object. If we 
add enough particles we can infer the form of the field of 
the object and thus the form of the object by interpolating 
the little parts of the field send out by the particles. 
In the case of the tensor voting we walk this path back- 
wards: First we have the particles in space which are our