CMRT09: Object Extraction for 3D City Models, Road Databases and Traffic Monitoring - Concepts, Algorithms, and Evaluation
72
Few commercial software packages allow automatic terrain,
tree and building extraction from Lidar data. In TerraSCAN, a
TIN is generated and progressively densified, the extraction of
off-terrain points is performed using the angles between points
to make the TIN facets and the other parameter is the distance
to nearby facet nodes (Axelsson, 2001). In SCOP++, robust
methods operate on the original data points and allow the
simultaneous elimination of off-terrain points and terrain
surface modelling (Kraus and Pfeifer, 1998).
In summary, most approaches try to find objects using single
methods. In our strategy, this study suggests complying
different methods using all available data with the focus on
improving the results of one method by exploiting the results
from the remaining ones.
3. INPUT DATA AND PREPROCESSING
The methods presented in this paper have been tested on a
dataset of the Zurich airport. The available data for this region
are: 3D vector data of airport objects, colour and CIR (Colour
InfraRed) images, Lidar DSM/DTM data (raw and grid
interpolated). The characteristics of the input data can be seen
in Table 1.
Image Data
RGB
CIR
Provider
Swissphoto
Swissphoto
Scale
1: 10*000
1: 6*000
Scan Resolution
14.5 microns
14.5 microns
Acquisition Date
July 2002
July 2002
Ground Sampling Distance
(GSD) (cm)
14.5 cm
8.7 cm
Lidar Data
DSM
DTM
Provider
Swisstopo
Swisstopo
T yP e
Raw & grid
Raw & grid
Raw point density & Grid
1 pt / 2 sqm &
1 pt / 2 sqm &
Spacing
2m
2m
Acquisition Date
Feb. 2002
Feb.2002
Vector data
Only for validation purposes
Provider
Unique Co.
Horizontal / Vertical
Accuracy (2 sigma)
20 / 25 cm
Table 1. Input data characteristics.
The 3D vector data describe buildings (including airport
parking buildings and airport trestlework structures). It has been
produced from stereo aerial images using the semi-automatic
approach with the CC-Modeler software (Gruen and Wang,
1998). Some additional reference buildings outside the airport
perimeter were collected using CIR images with stereo
measurement by using LPS software. The images have been
firstly radiometrically preprocessed (noise reduction and
contrast enhancement), then the DSM was generated with the
software package SAT-PP, developed at the Institute of
Geodesy and Photogrammetry, ETH Zurich (Zhang, 2005). For
the selection of the optimum band for matching, we considered
the GSD, and the quality of each spectral channel based on
visual checking and histogram statistics. Finally, the NIR band
was selected for DSM generation. The final DSM was
generated with 50cm grid spacing. Using this DSM, CIR
orthoimages were produced with 12.5cm ground sampling
distance. Lidar raw data (DTM and DSM) have been acquired
with “leaves off’. The DSM point cloud includes all Lidar
points (including points on terrain, tree branches etc.). The
DTM data includes only points on the ground, so it has holes at
building positions and less density at tree positions. The height
accuracy (one standard deviation) is 0.5 m generally, and 1.5 m
at trees and buildings, the latter referring only to the DSM. The
grid DSM and DTM were interpolated from the original raw
data by Swisstopo with the Terrascan commercial software.
4. BUILDING DETECTION
Four different approaches have been applied to exploit the
information contained in the image and Lidar data, extract
different objects and finally buildings. The first method is based
on DSM/DTM comparison in combination with NDVI
(Normalised Difference Vegetation Index) analysis for building
detection. The second approach is a supervised multispectral
classification refined with height information from Lidar data
and image-based DSM. The third method uses voids in Lidar
DTM and NDVI classification. The last method is based on the
analysis of the density of the raw DSM Lidar data. The
accuracy of the building detection process was evaluated by
comparing the results with the reference data and computing the
percentage of data correctly extracted and the percentage of
reference data not extracted.
4.1 DSM/DTM and NDVI (Method 1)
By subtracting the DTM from the DSM, a so-called normalized
DSM (nDSM) is generated, which describes the above-ground
objects, including buildings and trees. As DSM, the surface
model generated by SAT-PP and as DTM the Lidar DTM grid
were used. NDVI image has been generated using the NIR and
R bands. A standard unsupervised (ISODATA) classification
was used to extract vegetation from NDVI image. The
intersection of the nDSM with NDVI should correspond to
trees. By subtracting the resulting trees from the nDSM, the
buildings are obtained. 83% of building class pixels were
correctly classified, while all of 109 buildings have been
detected but not fully, the omission error is 7% . Within the
detected buildings, some other objects, such as aircrafts and
vehicles, were included. The extracted buildings are shown in
Figure 1.
Figure 1. Building detection result from method 1. (Left: airport
buildings. Right: residential area).
4.2 Supervised classification and use of nDSM (Method 2)
The basic idea of this method is to combine the results from a
supervised classification with the height information contained
in the nDSM. Supervised classification methods are preferable
to unsupervised ones, because the target of the project is to
detect well-defined standard target classes (airport buildings,
airport corridors, bare ground, grass, trees, roads, residential
houses, shadows etc.), present at airport sites. The training areas
were selected manually using AOI (Area Of Interest) tools
within the ERDAS Imagine commercial software (Kloer, 1994).
Among the available image bands for classification (R, G and B
from colour images and NIR, R and G bands from CIR images),
only the bands from CIR images were used due to their better
resolution and the presence of NIR channel (indispensable for
