Istanbul 2004
iatic building
RS J. 50 (4),
EXTRACTION OF SPATIAL OBJECTS FROM LASER-SCANNING DATA USING A
CLUSTERING TECHNIQUE
Bin Jiang
Division of Geomatics, Dept. of Technology and Built Environment, University of Gävle, SE-801 76 Gävle, Sweden -
bin.jiang(g)hig.se
Commission III, WG III/3
KEY WORDS: DEM/DTM, LIDAR, Laser-scanning, Algorithms, Classification
ABSTRACT:
This paper explores a novel approach to the extraction of spatial objects from the laser-scanning data using an unsupervised
clustering technique. The technique, namely self-organizing maps (SOM), creates a set of neurons following a training process based
on the input point clouds with attributes of xyz coordinates and the return intensity of laser-scanning data. The set of neurons
constitutes a two dimensional planar map, with which each neuron has best match points from an input point cloud with similar
properties. Because of its high capacity in data clustering, outlier detection and visualization, SOM provides a powerful technique
for the extraction of spatial objects from laser-scanning data. The approach is validated by a case study applied to a point cloud
captured using a terrestrial laser-scanning device.
1. INTRODUCTION
Laser-scanning has been proven to be as an effective 3D data
acquisition means for extracting spatial object models such as
digital terrain models and building models and it has been
widely used nowadays in geospatial information industry
(Ackermanm 1999). However what the laser-scanning can
acquire is a digital surface model, which captures all points
from treetops, buildings and ground surface depending on the
circumstance. For many practical applications, we often expect
spatial object models such as digital terrain models (DTM) of
the bare earth surface and 3D building models. To this end,
various research efforts have been made over the past years in
an attempt to processing the original captured datasets for the
derivation of various spatial objects. For instance, in order to
get a DTM of the bare earth surface, we have to remove those
non-terrain and undesired points.
This process of deriving spatial objects involves a range of
operations such as filtering, interpolation, segmentation,
classification, modelling and possible human interaction if no
complete automatic way is reached (Tao and Hu 2001). Most
filtering algorithms are targeted to the derivation of DTM, thus
assumptions on the spatial distributions of points or geometric
characteristics of a point relative to its neighbourhood in a
terrain surface are used to construct various filtering strategies.
For instance, through using TopScan, Petzold et al. (1999) used
the lowest points found in a mowing window to create a rough
terrain model. And it is used to filtering out those points higher
than a given threshold. Then repeat the procedure several times
with smaller sizes of moving window, and finally lead to a
DTM. Kraus and Pfeifer (1998) used an averaging surface
between terrain points and vegetation points to derive residuals
of individual points, and then use the residuals to determine the
weights of individuals to be selected or eliminated. Maas and
Vosselman (1999) adopted approaches based on the ideas of
moment invariants and the intersection of planar faces in
triangulated points for extracting building models. The slope-
based filter (Vosselman 2000) considers an observed fact that a
large height difference between two nearby points is unlikely to
be caused by a steep slope in the terrain. These algorithms
together with many others as reviewed in a recent comparison
study (Sithole and Vosselman 2003) are proven to be effective
and efficient in the studies, noting that they sometimes require
interpolation of a point cloud into regular grid format in order
to carry out the post-processing. When it is too complex to
distinguish different object points, additional information is
needed for classification or to achieve better results (e.g. Haala
and Brenner 1999, McIntosh and Krupnik 2002). However,
these filtering algorithms are all based one way or another on
supervised classifications with prior knowledge or assumptions
about different spatial objects. The supervised classification
solutions show various constraints in the sense of efficiency,
e.g. sensitivity to varying point densities, limited applicability
for certain kinds of spatial objects or under a certain
circumstance, and difficulty in dealing with stripe etc.
The supervised classification solutions rely much on the human
understanding or prior knowledge of the point geometric
characteristics of spatial objects. However, it is very difficult in
reality to get a true understanding, in particular when many
objects are involved in a point cloud. It also depends on our
specific task: e.g. to derive one single object or all objects with
one point cloud. It is probably an easy task to derive one object
rather than to distinguish all objects from an input point cloud.
For the case of single object, we can investigate the point cloud
and try to figure out the characteristics of its point distribution
and further design an appropriate algorithm. Furthermore, many
assumptions about the point characteristics do not always hold
true, and they depend on the circumstances of laser-scanning.
When come to the situation where all objects should be derived,
we believe unsupervised clustering seems a more appropriate
way.
One of the major reasons why unsupervised methods are so
important in the post-processing is that it is very difficult to
assume some characteristics of a certain object. Instead of
figuring out the assumption, unsupervised methods put these
characteristics aside and adopt a simple assumption, i.e. same
objects should have the same similarities in terms of their xyz