Recognising structure in laser scanner point clouds

Abstract

Both airborne and terrestrial laser scanners are used to capture large point clouds of the objects under study. Although for some applications, direct measurements in the point clouds may already suffice, most applications require an automatic processing of the point clouds to extract information on the shape of the recorded objects. This processing often involves the recognition of specific geometric shapes or more general smooth surfaces. This paper reviews several techniques that can be used to recognise such structures in point clouds. Applications in industry, urban planning, water management and forestry document the usefulness of these techniques.

Full text

RECOGNISING STRUCTURE IN LASER SCANNER POINT CLOUDS
1
G. Vosselman
a
, B.G.H. Gorte
b
, G. Sithole
b
, T. Rabbani
b
a
International Institute of Geo-Information Science and Earth Observation (ITC)
P.O. Box 6, 7500 AA Enschede, The Netherlands
vosselman@itc.nl
b
Delft University of Technology, Faculty of Aerospace Engineering
Department of Earth Observation and Space Systems, P.O. Box 5058, 2600 GB Delft, The Netherlands
{b.g.h.gorte, g.sithole, t.rabbani}@lr.tudelft.nl
KEY WORDS: Laser scanning, LIDAR, Surface, Extraction, Point Cloud, Segmentation
ABSTRACT:
Both airborne and terrestrial laser scanners are used to capture large point clouds of the objects under study. Although for some
applications, direct measurements in the point clouds may already suffice, most applications require an automatic processing of the
point clouds to extract information on the shape of the recorded objects. This processing often involves the recognition of specific
geometric shapes or more general smooth surfaces. This paper reviews several techniques that can be used to recognise such
structures in point clouds. Applications in industry, urban planning, water management and forestry document the usefulness of these
techniques.
1
This paper was written while the first author was with the Delft University of Technology.
1. INTRODUCTION
The recognition of object surfaces in point clouds often is the
first step to extract information from point clouds.
Applications like the extraction of the bare Earth surface
from airborne laser data, reverse engineering of industrial
sites, or the production of 3D city models depend on the
success of this first step. Methods for the extraction of
surfaces can roughly be divided into two categories: those
that segment a point cloud based on criteria like proximity of
points and/or similarity of locally estimated surface normals
and those that directly estimate surface parameters by
clustering and locating maxima in a parameter space. The
latter type of methods is more robust, but can only be used
for shapes like planes and cylinders that can be described
with a few parameters.
This paper gives an overview over different techniques for
the extraction of surfaces from point clouds. Section 2
describes various approaches for the segmentation of point
clouds into smooth surfaces. In section 3 some variation of
these methods are described that result in the extraction of
planar surfaces. Section 4 describes clustering methods that
can be used for the recognition of specific (parameterised)
shapes in point clouds. Section 5 shows that these kind of
surface extraction methods can be used in a wide range of
applications.
2. EXTRACTION OF SMOOTH SURFACES
Smooth surfaces are often extracted by grouping nearby
points that share some property, like the direction of a locally
estimated surface normal. In this way a point cloud is
segmented into multiple groups of points that represent
surfaces. The recognition of surfaces can therefore be
considered a point cloud segmentation problem. This
problem bears large similarities to the image segmentation
problem. Because the point clouds used in early research
were captured in a regular grid and stored in raster images,
this problem is often also referred to as range image
segmentation, which makes the similarity to (grey value)
image segmentation even stronger. All point cloud
segmentation approaches discussed in this section have their
equivalent in the domain of image processing.
2.1 Scan line segmentation
Scan line segmentation first splits each scan line (or row of a
range image) into pieces and than merges these scan line
segments with segments of adjacent scan lines based on some
similarity criterion. Thus it is comparable in strategy to the
split-and-merge methods in image segmentation. Jiang and
Bunke (1994) describe a scan line segmentation method for
polyhedral objects, which assumes that points on a scan line
that belong to a planar surface form a straight 3D line
segment. Scan lines are recursively divided into straight line
segments until the perpendicular distance of points to their
corresponding line segment is below some threshold. The
merging of the scan line segments is performed in a region
growing fashion. A seed region is defined as a triple of line
segments on three adjacent scan lines that satisfy conditions
with respect to a minimum length, a minimum overlap, and a
maximum distance between the neighbouring points on two
adjacent scan lines. A surface description is estimated using
the points of the seed region. Scan line segments of further
adjacent scan lines are merged with the region if the
perpendicular distances between the two end points of the
additional segment and the estimated surface are below some

The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences
threshold. Finally a minimum region size is required for a
successful detection of a surface.
Several variations can be made to the above principle. Sithole
and Vosselman (2003) describe a scan line segmentation
method that groups points on scan lines based on proximity
in 3D. These groups do not need to correspond to a sequence
of points in the scan line. In this way, points on either side of
an outlier to a surface are still grouped together. The scan
line segmentation is repeated for scan lines with different
orientations. Artificial scan lines are created by splitting the
data set into thin parallel slices of a user specified
orientation. Several scan lines sets (3 or 4) with different
orientations are segmented (Figure 1). Because all points are
present in each scan line set, many points will be part of
multiple scan line segments with different orientations. This
property is used to merge the scan line segments to regions:
scan line segments of different orientations are merged if they
share one or more points.
Figure 1: Segmentation of a scene with a building part.
Shaded view (top left), segmented scan lines with
two different orientations (top right and bottom
left) and the result of merging the scan line
segments (bottom right).
2.2 Surface growing
Surface growing in point clouds is the equivalent of region
growing in images. To apply surface growing, one needs a
method to identify seed surfaces and criteria for extending
these surfaces to adjacent points. Several variants of this
method are described by Hoover et al. (1996).
A brute-force method for seed selection is to fit many planes
and analyse residuals. For each point a plane is fit to the
points within some distance that point. The points in the
plane with the lowest square sum of residuals compose a seed
surface if the square sum is below some threshold. This
method assumes that there is a part in the dataset where all
points within some distance belong to the same surface.
Outliers to that surface would lead to a high residual square
sum and thus to a failure to detect a seed for that surface. It
depends on the application domain whether this smoothness
assumption holds. Robust least squares adjustment of planes
or Hough transform-like detection of planes (Section 3.1) are
more robust methods that would also detect seed surfaces in
the presence of outliers.
The growing of surfaces can be based on one or more of the
following criteria:
• Proximity of points. Only points that are near one of the
surface points can be added to the surface. For 2.5 D
datasets this proximity can be implemented by checking
whether a candidate point is connected to a point of the
surface by a (short) edge of a Delaunay triangulation.
This condition may, however, be too strict if some outlier
points are present. In that case, also other points within
some distance of the surface points need to be
considered. Otherwise the TIN edge condition may lead
to fragmented surfaces.
• Locally planar. For this criterion a plane equation is
determined for each surface point by fitting a plane
through all surface points within some radius around that
point. A candidate point is only accepted if the
orthogonal distance of the candidate point to the plane
associated with the nearest surface point is below some
threshold. This threshold and the radius of the
neighbourhood used in the estimation of the plane
equation determine the smoothness of the resulting
surface.
• Smooth normal vector field. Another criterion to enforce
a smooth surface is to estimate a local surface normal for
each point in the point cloud and only accept a candidate
point if the angle between its surface normal and the
normal at the nearest point of the surface to be grown is
below some threshold.
2.3 Connected components in voxel space
It is not uncommon to perform certain steps in processing
airborne laser point clouds in the 2-dimensional grid (image)
domain. Quite naturally, a point cloud represents a 2.5D
surface. When converted to a 2D grid, grid positions are
defined by the (x,y) coordinates of the points, and their z-
coordinates determine the pixel values. To the resulting
regular-grid DSM, operations can be applied that are known
from image processing to perform certain analysis functions.
Examples are thresholding to distinguish between terrain and
e.g. buildings in flat areas, mathematical morphology to filter
out vegetation and buildings also in hilly terrain, textural
feature extraction to distinguish between trees and buildings,
(Oude Elberink and Maas 2000), and region growing to
identify planar surfaces (Geibel and Stilla, 2000).
Point clouds obtained by terrestrial laser scanning, however,
are truly 3D, especially when recordings from several
positions are combined. There is not a single surface that can
be modelled by z=f(x,y) as there is in the 2.5D case, and
converting such point clouds to a 2D grid would cause a
great loss of information.
Recently we adopted the alternative approach of converting a
3D point cloud into the 3-dimensional grid domain. The cells
in a 3D grid are small cubes called voxels (volume elements,
as opposed to pixels or picture elements in the 2D case). The
size of the grid cells determines the resolution of the 3D grid.
Usually, the vast majority of voxels (grid positions) will
contain no laser points and get value 0, whereas the others
are assigned the value 1, thus creating a binary grid with
object and background voxels. A slightly more advanced
scheme is to count the number of points that falls into each
grid cell and to assign this number as a voxel value.

The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences
Similar operations as in 2D image processing can be applied
to the 3D voxel spaces. An important class of 2D image
processing operators is formed by neighbourhood operators,
including filters (convolution, rank order) and morphologic
dilation, erosion, opening and closing. In the 3D case, 3D
neighbourhoods have to be taken into account, which means
that filter kernels and structuring elements become 3
dimensional as well. Roughly speaking, mathematical
morphology makes most sense for binary voxel spaces,
whereas filters are more useful for ‘grey scale’ cases, such as
density images. The former can be used to shrink and enlarge
objects, suppress ‘binary’ noise, remove small objects, fill
holes and gaps in/between larger objects, etc. etc. Application
of convolutions and other filters also has a lot of potential,
for example for the detection of 3-dimensional linear
structures and boundaries between objects.
At a slightly higher level, 2D operations like connected
component labelling, distance transform and skeletonisation
can be defined, implemented and fruitfully applied in 3D
(Palagyi and Kuba 1999, Gorte and Pfeifer 2004).
In all cases, the benefit of voxel spaces, compared to the
original point cloud, lies in the implicit notion of adjacency
in the former. Note that each voxel has 26 neighbours.
Regarding voxels as cubes that fill the space, 6 of the
neighbours share a face, 12 share an edge and 8 share a
corner with the voxel under consideration.
3. ITERATIVE EXTRACTION OF PLANAR
SURFACES
For many applications, the objects under study are known to
be polyhedral. In that case, the segmentation algorithms
should determine planar surfaces. This can be considered a
specific case of smooth surface extraction. Several small
variations can be made to the methods described in the
previous section which enable the extraction of planar
surfaces.
3.1 Plane growing
Both the scan line segmentation (Section 2.1) and the surface
growing (Section 2.2) algorithm contain a merging step. This
step can easily be adapted to ensure the extraction of planar
surfaces.
The scan line method as described by Jiang and Bunke
(1994) in fact was originally designed for the extraction of
planar surfaces. In the process of grouping the scan line
segments (that are linear in 3D space) they demand that the
resulting surface is planar. A new scan line segment is only
added if the end points of that segment are within some
distance of the plane. The plane equation is updated after
adding a new scan line segment.
The surface growing algorithm is transformed into a plane
growing algorithm if the criterion to enforce local planarity is
modified to a global planarity. This is achieved by using all
points of the surface in the estimation of the plane equation.
3.2 Merging TIN meshes
In most algorithms, the surfaces are defined as a group of
points. Gorte (2002), however, describes a variation in which
the triangular meshes of a TIN are the units that compose a
surface. Planar surfaces are extracted by merging two sur-
faces if their plane equations are similar. At the start of the
merging process, a planar surface is created for each TIN
mesh. Similarity measures are computed for each pair of
neighbouring surfaces. Those two surfaces that are most simi-
lar are merged and the plane equation is updated. This proc-
ess continues until there are no more similar adjacent sur-
faces.
4. DIRECT EXTRACTION OF PARAMETERISED
SHAPES
Many man-made objects can be described by shapes like
planes, cylinders and spheres. These shapes can be described
by only a few parameters. This property allows the extraction
of such shapes with robust non-iterative methods that detect
clusters in a parameter space.
4.1 Planes
A plane is the most frequent surface shape in man-made
objects. In the ideal case of a noiseless point cloud of a plane,
all locally estimated surface normals should point in the same
direction. However, if the data is noisy or if there is a certain
amount of surface roughness (e.g. roof tiles on a roof face),
the surface normals may not be of use. The next two
paragraphs describes the extraction of planes without and
with the usage of surface normals respectively.
4.1.1 3D Hough transform
The 3D Hough transform is an extension of the well-known
(2D) Hough transform used for the recognition of lines in
imagery (Hough, 1962). Every non-vertical plane can be
described by the equation
dYsXsZ
yx
++=
(1)
in which s
x
and s
y
represent the slope of the plane along the
X- and Y-axis respectively and d is the height of the plane at
the origin (0, 0). These three plane parameters define the
parameter space. Every point (s
x
, s
y
, d) in this parameter
space corresponds to a plane in the object space. Because of
the duality of these two spaces, every point (X, Y, Z),
according to Equation 1, also defines a plane in the parameter
space (Maas and Vosselman 1999, Vosselman and Dijkman,
2001).
The detection of planar surfaces in a point cloud can be
performed by mapping all these object points to planes in the
parameter space. The parameters of the plane equation in the
point cloud defined by the position in the parameter space
where most planes intersect. The number of planes that
intersect at a point in the parameter space actually equals the
number of points in the object space that are located on the
plane represented by that point in the parameter space.
For implementation in a computer algorithm, the parameter
space, or Hough space, needs to be discreet. The counters of
all bins of this 3D parameter space initially are set to zero.
Each point of the point cloud is mapped to a plane in the
parameter space. For each plane that intersects with a bin, the
counter of this bin is increased by one. Thus, after all planes
have been mapped to the parameter space, a counter
represents the number of planes that intersected the bin. The

The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences
coordinates of the bin with the highest counter define the
parameters of the object plane with the largest amount of
points of the point cloud.
These points, however, do not necessarily belong to the one
and the same object surface. They may belong to multiple
coplanar object surfaces and some points may be even part of
object surfaces that only intersect the determined plane. To
extract surfaces that correspond to planar object faces, one
therefore needs group the points based on proximity. Only
groups that exceed some minimum size should be accepted as
possible planar object surface.
For the optimal bin size of the parameter space, a balance
needs to be determined between the accuracy of the
determined parameters on the one hand and the reliability of
the maximum detection on the other hand. The smaller the
bin size, the more accurate the determination of the plane
parameters will be. However, with very small bin sizes, all
planes in the parameter size will not intersect with the same
bin due to noise in the point cloud. Therefore, the maximum
in the Hough space will become less distinct and may not be
detectable any longer.
4.1.2 3D Hough transform using normal vectors
If normal vectors can be computed accurately, they can be
used to speed up the Hough transformation and to increase
the reliability. The position of a point in object space,
together with the normal vector, completely defines a plane
in object space. Therefore, the parameters of this plane can
directly be mapped to a single point in the parameter space.
In this way, there is no need to calculate the intersection of a
plane in the parameter space with the bins of that space. Only
the counter of a single bin needs to be incremented. The
availability of the normal vector information will also reduce
the risk of detecting spurious object planes.
To reduce the dimension of the parameter space and thereby
memory requirements, it is also possible to split the plane
detection in to steps: the detection of the plane normal and
the detection of the distance of the plane to the origin. In the
first step the normal vectors are mapped onto a Gaussian
(half) sphere. Because all normal vectors belonging to the
same plane should point into the same direction, they should
all be mapped to the same position on the Gaussian sphere.
This sphere is used as a two-dimensional parameter space.
The maximum on the Gaussian sphere defines the most likely
direction of the normal vector. This normal vector defines the
slopes s
x
and s
y
of Equation 1. This equation can be used to
calculate the remaining parameter d for all points with a
normal vector similar to the maximum on the Gaussian
sphere. These values d can be mapped to a one-dimensional
parameter space. The maximum in this space determines the
most likely height of a plane above the origin.
4.2 Cylinders
Cylinders are often encountered in industrial scenes. A
cylinder is described by five parameters. Although one could
define a five-dimensional parameter space, the number of
bins in such a space make the detection of cylinders very time
and memory consuming and unreliable. To reduce the
dimension of the parameter space, the cylinder detection can
also be split into two parts: the detection of the cylinder axis
direction (2 parameters) and the detection of a circle in a
plane (3 parameters). For this procedure the availability of
normal vectors is required.
In the first step, the normal vectors again plotted on the
Gaussian sphere. Because all normal vectors on the surface of
a cylinder point to the cylinder axis, the Gaussian sphere will
show maxima on a big circle. The normal of that circle is the
direction of the cylinder axis. Figure 2 shows an industrial
scene with a few cylinders. The Gaussian half sphere of this
scene shows maxima on several big circles that correspond to
the different axis directions of the cylinders. By extracting
big circles with high counters along a larger part of the circle,
hypothesis for cylinder axis directions are generated.
Figure 2: Industrial scene with points colour coded by their
surface normal direction (left). Gaussian half
sphere with circles corresponding to the dominant
cylinder axis directions (right).
All points that belong to a selected big circle are now
projected onto a plane perpendicular to the hypothesised
cylinder axis. In this plane, the points of a cylinder should be
located on a circle. The detection of a circle in a two-
dimensional space is a well-known variant to the Hough
transformation for line detection (Kimme et al. 1975). The
three-dimensional parameter space consists of the two
coordinates of the circle centre and the circle radius. Each
point is mapped to a cone in the parameter space. I.e. for each
radius, each point is mapped to a circle in the parameter
space. The circle centre is known to lie on this circle. If usage
is made of the normal vector information, the number of bins
that need to be incremented can again be reduced. In this case
each point can be mapped to two lines in the parameter space
(assuming that the sign of the normal vector is unknown). I.e.
for each radius and a known normal vector, there are only
two possible locations of the circle centre.
4.3 Spheres
A sphere can be detected in a four dimensional parameter
space consisting of the three coordinates of the sphere centre
and the radius. For each radius, each point in the point cloud
defines a sphere in the parameter space on which the sphere
centre should be located. Making use of the normal vector
information, each point can be mapped to two lines in this
four dimensional space. I.e. for each radius and a known
normal vector, there are two possible location for the sphere
centre.
Alternatively, one can first locate the sphere centre and then
determine the radius. Each point and a normal vector define a
line along which the sphere centre should be located. In this
case the parameter space consist of the coordinates of the
circle centre and each point with a normal vector is mapped
to a line in the parameter space. Ideally, the lines belonging
to the points of the same sphere should intersect in the sphere
centre. Once, the sphere centre is located, the determination
of the radius is left as a one-dimensional clustering problem,

The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences
like the determination of the distance of a plane to the origin
as discussed in paragraph 4.1.
5. EXAMPLES
In various research projects at the Delft University of
Technology, we have been using and developing several of
the above point cloud segmentation techniques. Results are
presented here where segmentations have been used to model
industrial installations, city landscapes, digital elevation
models and trees.
5.1 Industrial installations
Three-dimensional CAD models of industrial installations are
required for revamping, maintenance information systems,
access planning and safety analysis. Industrial installation in
general contain a large percentage of relatively simple
shapes. Recorded with terrestrial laser scanners the high point
density point clouds accurately describe these shapes. Point
clouds of such scenes can therefore be processed well with
clustering methods as described in section 4.
However, scenes with a large number of different objects may
result in parameter spaces that are difficult to interpret.
Therefore, large datasets (20 million points) have been
segmented using the surface growing technique as described
in paragraph 2.2. Using a very local analysis of the surface
smoothness, this segmentation will group all points on a
series of connected cylinders to one segment. Such segments
usually only contain a few cylinders and planes. The point
cloud of each segment is then further analysed with the
methods for direct recognition of planes and cylinders.
Figure 3: Cylinders and planes in an industrial scene.
Figure 3 shows the result of the automatic modelling of the
point cloud shown in Figure 2. A large percentage of the
cylinders is recognised automatically.
5.2 City landscapes
Three-dimensional city models are used for urban planning,
telecommunications planning, and analysis of noise and air
pollution. A comparative study on different algorithms for
the extraction of city models from airborne laser scanning
data and/or aerial photographs is currently conducted by
EuroSDR. Figure 4 shows the results of modelling a part of
the city centre of Helsinki from a point cloud with a point
density of a few points per square meter.
A large part of the roof planes was detected using the 3D
Hough transform (paragraph 4.1.1). For other parts of the
roof landscape the operator interactively decided on the
shape of the roof to be fitted to the point cloud of a building.
Figure 4: Model of the city centre of Helsinki
The terrain surface was extracted by segmentation of the
point cloud into smooth surfaces with the surface growing
method (paragraph 2.2). Most terrain points are grouped in
one segment. Around the cathedral, the stairs and the plat-
form were detected as separate segments. By a few mouse
clicks, the operator can specify which segments belong to the
terrain. In particular in complex urban landscapes, some in-
teraction is often required to determine the terrain surface.
5.3 Digital elevation models
Digital elevation models are widely used for water
management and large infrastructure construction projects.
The extraction of digital elevation models from airborne laser
scanning data is known as filtering. A large variety of
filtering algorithms has been developed in the past years.
Sithole and Vosselman (2004) present and experimental
comparison of these algorithms. Typical filtering algorithms
assume that the point cloud contains one low smooth surface
and locally try to determine which points belong to that
surface. Most often, they do not segment the point cloud.
Segmentation can, however, also be used for the purpose of
the extraction of the terrain surface. Segmenting the point
cloud has the advantage that large building structures can be
removed completely (something which is difficult for
algorithms based on mathematical morphology).
Segmentation also offers the possibility to further analyse
point clouds and detect specific structures.
The definition of a digital elevation model (DEM) is often
dependent on the application. Even within the domain of
water management, some tasks require bridges to be removed
from the DEM, whereas they should be part of the DEM for
other tasks. Bridges, as well as fly-overs and entries of
subways or tunnels, are difficult to handle for many filter
algorithms. On some sides these objects smoothly connect to
the terrain. On other sides, however, they are clearly above or
below the surrounding terrain. Often, these objects are
partially removed.
Sithole and Vosselman (2003) use the scan line segmentation
algorithm with artificial scan lines sets with multiple
orientations (paragraph 2.1) to extract objects. Because
bridges are connected to the terrain, they are extracted as part
of the bare Earth surface. In a second step bridges are
extracted from this surface by analysing the height
differences at the ends of all scan line segments. Scan lines
that cross the bridge will have a segment on the bridge

The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences
surface that is higher than the two surrounding segments. By
connecting those scan line segments that are raised on both
sides, objects like bridges and fly-overs are detected. A
minimum size condition is applied to avoid the detection of
small objects.
Figure 5 (left) shows a scene with a bridge surrounded by
dense vegetation and some buildings. The right hand side
depicts the extracted bare Earth points and the bridge as a
separately recognised object.
Figure 5: Point cloud with a bridge, dense vegetation and
buildings (left). Extracted digital elevation model
and bridge (right).
5.4 Trees
Conversion and subsequent
processing of terrestrial laser
points to a 3D voxel space has
been applied recently during a
cooperation between Delft
University of Technology and the
Institute for Forest Growth
(IWW) in Freiburg. The purpose
was 3D model reconstruction of
trees, and to estimate parameters
that are relevant to estimate the
ecological state, but also the
economical value of a
(production) forest, such as wood
volume and length and
straightness of the stem and the
branches.
A crucial phase in the
reconstruction process is
segmentation of the laser points
according to the different
branches of the tree. As a result,
to each point a label is assigned
that is unique for each branch, whereas points are removed
(labelled 0) that do not belong to the stem or a significant
branch (leafs, twigs, noise).
In terms of voxel spaces, the problem resembles that of
recognizing linear structures in 2-dimensional imagery, such
as roads in aerial photography. Therefore, we decided to
tackle it in a similar manner, i.e. by transferring a variety of
image processing operators to the 3D domain, such as
erosion, dilation, overlay, skeletonisation, distance transform
and connected component labelling. Also Dijkstra’s shortest-
route algorithm plays an important role during the
segmentation (Gorte and Pfeifer, 2004).
The examples of this section demonstrate that segmentation
of point clouds is an important step in the modelling of
objects. Various segmentation algorithms were discussed and
applied. The most suitable segmentation algorithm may
depend on the kind of application.
REFERENCES
IAPRS – International Archives of Photogrammetry, Remote
Sensing and Spatial Information Sciences.
Geibel R. and U. Stilla, 2000. Segmentation of laser-altimeter
data for building reconstruction: Comparison of different
procedures. IAPRS, vol. 33, part B3, Amsterdam, 326-334.
Gorte, B., 2002. Segmentation of TIN-Structured Surface
Models. In: Proceedings Joint International Symposium on
Geospatial Theory, Processing and Applications, on CD-
ROM, Ottawa, Canada, 5 p.
Gorte, B.G.H. and N. Pfeifer (2004). Structuring laser-
scanned trees using 3D mathematical morphology, IAPRS,
vol. 35, part B5, pp. 929-933.
Hoover, A., Jean-Baptiste, G., Jiang, X., Flynn, P.J., Bunke,
H., Goldgof, D.B., Bowyer, K., Eggert, D.W., Fitzgibbon, A.,
Fisher, R.B., 1996. An Experimental Comparison of Range
Image Segmentation Algorithms. IEEE Transactions on
Pattern Analysis and Machine Intelligence 18 (7): 673-689.
Hough, P.V.C., 1962. Method and means for recognizing
complex patterns. U. S. Patent 3,069,654.
Jiang, X.Y., Bunke, H., 1994. Fast Segmentation of Range
Images into Planar Regions by Scan Line Grouping. Machine
Vision and Applications 7 (2): 115-122.
Kimme, C., Ballard, D.H., and Sklansky, J., 1975. Finding
Circles by an Array of Accumulators. Communications of the
ACM 18 (2):120-122.
Maas, H.-G., Vosselman, G., 1999. Two Algorithms for
Extracting Building Models from Raw Laser Altimetry Data
ISPRS Journal of Photogrammetry and Remote Sensing 54
(2-3): 153-163.
Oude Elberink, S. and H. Maas, 2000. The use of anisotropic
height texture measures for the segmentation of airborne laser
scanner data. IAPRS, vol. 33, part B3, Amsterdam, pp. 678-
684.
Palagyi, K. and A. Kuba, 1999. Directional 3D thinning
using 8 subiterations, LNCS 1568, Springer Verlag, pp. 325-
336.
Sithole, G., Vosselman, G., 2003. Automatic Structure
Detection in a Point Cloud of an Urban Landscape. In:
Proceedings 2
nd
GRSS/ISPRS Joint Workshop on Remote
Sensing and Data Fusion over Urban Areas, URBAN2003,
Berlin, Germany, pp. 67-71.
Sithole, G., Vosselman, G., 2004. Experimental Comparison
of Filter Algorithms for Bare Earth Extraction from Airborne
Laser Scanning Point Clouds. ISPRS Journal of
Photogrammetry and Remote Sensing 59 (1-2): 85-101.
Vosselman, G., Dijkman, S., 2001. 3D Building Model
Reconstruction from Point Clouds and Ground Plans. IAPRS,
vol. 34, part 3/W4, Annapolis, pp. 37-43.
Figure 6:
Segmented
tree