CMRT09: Object Extraction for 3D City Models, Road Databases and Traffic Monitoring - Concepts, Algorithms, and Evaluation
188
Preparatory work regarding the existing database is required in
order to exploit the on-line segmentation results in a pilot
assistance system. A set of planar surfaces characterizing the
urban terrain is needed for the later comparison (e.g. facades,
rooftops). In our experiments, even this information originated
from previously collected ALS data, which were recorded under
optimal DGPS conditions, but it might as well be derived from
any other existing city model.
Within our approach, we assume that INS navigation is
continuously available and that we have an initial guess of the
sensor position (e.g. from non-differential GPS). If we have to
navigate through GPS dropouts, the positioning accuracy will
degrade because of INS drift effects, but we can assume that the
measured ALS data are still roughly aligned to the stored
information. In addition to their position in 3D space, features
like size and normal direction are assigned to all segmented
planar patches, thus it is comparatively easy to find
corresponding objects in the database. We use this information
to achieve precise alignment between the measured ALS data
and the existing city model, which finally enables us to correct
the presumed sensor position.
1.3 Related work
In recent years, airborne laser scanning systems have been
explored by various scientists from different points of view. The
complexity of ALS data acquisition leads to a number of
potential error sources. Schenk (2001) and Filin (2003) address
this problem and categorize different influences that should be
considered. In addition to varying exactness of the navigational
information sources, several systematic effects can lead to
reduced point positioning accuracy. Exemplary limiting factors
are scanning precision and range resolution of the specific laser
scanning device. Other negative effects can be introduced by
inaccurate synchronization of the system components.
Considerable discrepancies are caused by mounting errors or
disregarded lever arms (displacements between laser scanner,
INS, and GPS antenna). Skaloud and Lichti (2006) approached
this problem with a rigorous method to estimate the system
calibration parameters such that 3D points representing a plane
are conditioned to show best possible planarity. In order to use
ALS within the scope of aircraft navigation, we presuppose that
the sensor system has been calibrated beforehand.
Some procedures described in this paper are concerned with the
segmentation of point clouds into planar surfaces. Many
different methods regarding this topic can be found in literature.
Some authors are interested in detecting even more kinds of
objects like spheres, cylinders, or cones. Rabbani et al. (2007)
describe two methods for registration of point clouds, in which
they fit models to the data by analyzing least squares quality
measures. Vosselman et al. (2004) use a 3D Hough transform to
recognize structures in point clouds. Filin & Pfeifer (2006)
propose a segmentation method that is based on cluster analysis
in a feature space. Among all available approaches, the
RANSAC algorithm (Fischler & Bolles, 1981) has several
advantages to utilize in the shape extraction problem (Schnabel
et al., 2006). We apply a RANSAC-based robust estimation
technique to fit straight line segments to the scan line data.
Moreover, an extension of this method is used to identify
locally planar patches in the model data. The amount of outliers
lets us distinguish between buildings and irregularly shaped
objects like trees. Fundamental ideas on fast segmentation of
range data into planar regions based on scan line analysis have
been published by Jiang and Bunke (1994). Their algorithm
divides each row of a range image into straight line segments
which are combined in a region growing process. Despite the
fact that we are considering continuously recorded scan lines
instead of range images, we basically follow this approach
during the on-line data analysis.
Several existing concepts of terrain-based navigation for aerial
vehicles can be found, e.g. image based navigation (IBN),
terrain-following radar (TFR), or terrain contour matching
(TERCOM). Other than these methods, laser scanning is a
comparatively new technique. Toth et al. (2008) propose the use
of LiDAR for terrain navigation, as it provides distinct 3D
measurements that can easily be used for exact comparison to
previously recorded data. In their concept, the iterative-closest-
point algorithm (Besl & McKay, 1992) is chosen for surface
matching. Instead of an 1CP approach, we identify matching
planar objects with regard to several geometric features (i.e.
position, size, normal direction). Similar methods have
demonstrated high performance for markerless TLS registration
(Brenner et al., 2008). The problem of determining the
transformation parameters is transferred to a system of linear
equations that can be solved immediately.
2. EXPERIMENTAL SETUP
Data used for this study were collected during a field campaign
in 2008, using the sensor equipment that is briefly described in
this section.
2.1 Sensor carrier
The sensors described below have been attached to a helicopter
of type BellUH-lD (Figure 1). Laser scanner and IMU are
mounted on a common sensor platform at the side of the
helicopter, which can be tilted to allow different perspectives,
i.e. nadir or oblique view. In an operational system, the pilot
must be able to react to upcoming dangers, e.g. during degraded
visibility conditions. Therefore, an obliquely forward-looking
sensor configuration was used in our experiments. The lever
arms of the components in the system are known, and the
correct bore-sight angles have been determined. Calibration of
these parameters is not topic of this paper, suitable methods can
be found in (Skaloud & Lichti, 2006).
2.2 Laser Scanner
The RIEGL LMS-Q560 laser scanner makes use of the time-of-
flight distance measurement principle with a pulse repetition
rate of 100 kHz. Opto-mechanical beam scanning provides
single scan lines, where each measured distance can be
georeferenced according to position and orientation of the