CLASSIFICATION METHODS FOR 3D OBJECTS IN LASERSCANNING DATA |
D. Tóvári, T. Vógtle |
University of Karlsruhe, Institute of Photogrammetry and Remote Sensing, 76128 Karlsruhe, Englerstr. 7. |
vocetie(éipf.uni-karlsruhe.
Commis
KEY WORDS: Laser scanning, Classification, Detection, Fuzzy logic, Quality |
ABSTRACT:
The object classification can play an important role in a lot of applications of airborne laserscanning data. The filtering process and |
the subsequent DTM generation using airborne laserscanning
objects (e.g. vegetation, buildings etc.). On the other hand classification can be also the first step of object-specific modelling, like |
vegetation or building reconstruction for 3D city models, design of telecommunication networks, urban planning or disaster
management.
A pixel-wise classification — especially when using laserscannning data - is limited in terms of reliability of its results. Therefore, the |
first step of this approach will be a segmentation of 3D objects.
etc.) are extracted and used for subsequent classification process. In this phase the method is based on raster data.
For segmentation a normalised DSM (nDSM) is generated by subtracting the original laser data (DSM) from a rough DTM (created
by a strong filtering of the DSM). Now 3D objects can be segmented by means of specific a region growing algorithm on this
nDSM.
Different kind of object-oriented features are calculated for each segment, like height texture, border gradients, first/last pulse height
differences, shape parameters or laser intensities. For classification two methods have been applied, on one hand a fuzzy logic
classification, on the other hand a statistical method (maximum likelihood). The fuzzy logic approach resulted in an overall |
classification rate of about 95% for test site ‘Salem’ (hilly t
confusion matrix for ‘Salem’ show that buildings were erroneous classified as trees (5%) resp. trees as buildings (4%). The most
errors can be observed at terrain objects which are confused mainly with trees (7%). Investigations concerning the statistical |
approach are currently done. Results and a comparison with fuzzy logic approach will be presented in this paper.
1. INTRODUCTION
During the last years airborne laserscanning has become one of
the standard data acquisition methods in the field of surveying.
Starting from the extraction of digital surface and terrain
models (DSM, DTM) a great variety of applications has been
Internai
de, tovari(dipf.uni-karlsruhe.de
sion VI, 111/4 |
data can be significantly improved by classification of non-terrain
For each segment object-specific features (e.g. height texture, shape
errain) and about 90% for test site ‘Karlsruhe’ (flat terrain). The
Inside these segments object-specific features will be extracted
which are used in the subsequent classification process. Two
classification methods has been used, fuzzy logie and a
stochastic approach (maximum likelihood). The influences and
dependences on different feature combinations as well as a
comparison of the results of different classification
schemes/schemata is the main topic of these investigations.
developed, like creation of 3D city models, determination of
tree parameters in forestry or control of power lines, e.g. in
Lohr (1999). At our institute we use laserscanning data in two
2. DATA
different projects. On one hand detection and modelling of
buildings is based on these data to recognize and classify rough
damages after strong earthquakes. On the other hand a high
resolution terrain models including the determination of
vegetation areas (position, size, density and height of trees etc.)
has to be extracted from airborne laserscanning data to model
hydrologic processes, e.g. runoff models to simulate floods.
For these purposes it is necessary to classify all 3D objects on
the surface of the earth, i.e. mainly buildings and trees/bushes,
in some cases also terrain objects like rough rocks which may
be additionally included in the detected objects. Such a
classification is the precondition for a class-specific modelling
of buildings as well as vegetation objects. On the other hand the
knowledge about the object type can be used for a significant
improvement of the extraction of terrain models by a class
dependent filtering of the original laser point cloud.
The first step of this approach is a segmentation of the
laserscanning data for detecting 3D objects on the terrain.
At this state of our approach all features are derived exclusively
from laserscanning data itself without additional information 3.1 De
like spectral images or GIS data. This is caused by specific
restrictions in context of disaster management - as mentioned As mer
above - where data acquisition has to be carried out also during Salem a
night time and poor weather conditions. On the other hand the Wa dO:
potential as well as the limitations of analysing airborne lasersca
laserscanning data should be investigated. efe m
were in
For this approach data of TopoSys II sensor in raster format Karlsrul
(grid size=1.0m) for two different test areas are used, Karlsruhe and veg
(urban environment, flat terrain, size: approx. 2km x 2km) and occurs ¢
Salem - near Lake Constance (rural environment, hilly terrain,
size: approx. 2km x Ikm). Both areas were captured in first and 32 Seg
last pulse mode while for test site Salem additionally laser Althous
intensity was registered. Figure 1 and 2 show an subset of these used :
test sites. Salem data set was used by kindly permission ol RN
TopoSys (Germany). other w
408
