CMRT09: Object Extraction for 3D City Models, Road Databases and Traffic Monitoring - Concepts, Algorithms, and Evaluation
54
the theoretical work on fully automated object extraction and
practical applications of the same (Mayer 2008). Success in
automatic object extraction will also help in determining
changes that occur between noise surveys (5 years) by
comparing the extracted objects at the different epochs and
should speed up the updating process of the G1S database.
1.2 Related Work
LiDAR has been extensively used for the generation of both
DSMs (Digital Surface Model) and DTMs (Digital Terrain
Model). Different classification methods have been used for the
classification of terrain and off terrain points (Sithole &
Vosselman, 2003). Different approaches have been used for the
detection and reconstruction of buildings from LiDAR data
(Brunn & Weidner, 1997 and Clode et al., 2004). Haitao et al.
(2007) used aerial images and LiDAR data for land cover
classification based on SVM (Scalable Vector Machine). Haala
& Brener (1999) also used the combination of multispectral
imagery and LiDAR data for the extraction of buildings, trees
and grass covered areas. Trees and grass covered areas were
classified easily from the multispectral imagery but were found
difficult to separate. Similarly, trees and buildings were
separated using height differences between DSM and DTM.
Both data sources were combined in order to identify the three
classification types. Rottensteiner et al. (2004) classified land
cover into four different classes namely, buildings, trees, grass
lands, and bare soil. This was achieved by combining LiDAR
data and multispectral images. Prior to performing building
detection by data fusion based on the theory of Dempster-
Shafer, the LiDAR data was pre-processed to generate a DTM.
For the extraction of roads different information sources such
as multispectral images from airborne and space borne sensors
were used. Clode et al. (2007) used only LiDAR for road
extraction. Despite encouraging results, there are still many
fundamental questions to be answered for road extraction in
urban areas (Mayer et al., 2008).
2. METHOD
The method under investigation is based on a workflow that
identifies and classifies buildings, trees and other objects by
fusing the information from LiDAR and aerial image data.
This information includes the normalised digital surface model
(NDSM) and multiple echoes from the LiDAR data together
with Normalized Difference Vegetation Index (NDVI) data
generated from the airborne imagery. The method is depicted
in Figure 1. Three major task groups may be identified, namely
the image group, LiDAR group and object extraction group.
3. WORKFLOW
Within the image group of tasks, the first step is to produce
orthophotos for each spectral channel of the ADS40 sensor, i.e.
Red (R), Green (G), Blue (B) & Near Infrared (NIR). For these
orthophotos, the required DSM can be created relatively
automatically using the panchromatic forward and backward
image data captured by the ADS40 sensor. The effect of DSM
quality on orthophoto generation is shown in Figure 2. The
upper part of the figure shows a rectified building using a DSM
generated by aerial images and the lower part shows the same
building rectified using a DSM from LiDAR data.
Figure 1 : Method Workflow
Figure 2: Effect of DSM on Building Rectification
In case of the building illustrated in Figure 2, a DSM created
from LiDAR data with a resolution of 0.5 m (the lower
example) provided sharper edges compared to that generated
from the image DSM and was used in the generation of the
orthophotos. As a prerequisite to this step the quality of the
registration between the airborne imagery and the LiDAR data
must be verified. The Nearest Neighbourhood method was
used as a sampling method for orthophoto generation. Separate
