In: Stilla U, Rottensteiner F, Paparoditis N (Eds) CMRT09. IAPRS, Voi. XXXVIII, Part 3/W4 — Paris, France, 3-4 September, 2009
189
sensor. Waveform analysis can contribute intensity and pulse-
width as additional features, but since we are mainly interested
in fast acquisition and on-line processing of range
measurements, we neglect full waveform analysis throughout
this paper. Range d under scan angle a (Figure 1) is estimated
corresponding to the first returning echo pulse as it can be
found by constant fraction discrimination. Typically, each scan
line covers a field of view of 60° with 1000 range measurements
(d,a) that can be converted to 2D Cartesian coordinates
(Figure 6). Navigational data are synchronously assigned to the
range measurements for direct georeferencing.
single point cloud. Both illustrations depict the aggregated data
within the horizontal cross-section of a building in the overlap
area. Best accuracy as shown in Figure 2a results from a global
optimization of the navigational data with the Applanix
postprocessing software. In this example, the data of six DGPS
ground stations were taken into account. Compared with this
accuracy, discrepancies of several meters can occur if the real
time navigation solution is used (Figure 2b). This situation will
even get worse in case of GPS dropouts.
3.1 Automatic generation of an adequate database
2.3 Navigational sensor system
The Applanix POS AV 410 comprises a GPS receiver and a
gyro-based inertial measurement unit (IMU), which is the core
element of the inertial navigation system (INS). The GPS data
are used for drift compensation and absolute georeferencing,
whereas the IMU determines accelerations with high precision.
These data are transferred to the position and orientation
computing system (PCS), where they are fused by a Kalman
filter, resulting in position and orientation estimates for the
sensor platform. In addition to the real-time navigation solution,
specialized software can be used for accurate postprocessing of
the recorded navigational data. Applanix POSPac MMS
incorporates the use of multiple DGPS reference stations and
the import of precise GPS ephemeris information. We consider
this corrected navigation solution while generating an optimal
database of the urban terrain.
3. USED METHODS AND DATA PROCESSING
In this chapter, we distinguish two different operating modes of
ALS data acquisition and processing. First, we assume that we
have optimal settings for creation of an adequate database: the
relevant urban area can be scanned several times from multiple
aspects with a calibrated sensor, and data can be processed and
optimized off-line. During this stage, we can resort to own
differential GPS base stations or use according information, e.g.
provided by the “Satellite Positioning Service of the German
State Survey” (SAPOS). Under these conditions, the absolute
measurement accuracy of an ALS system is typically in the
order of one decimeter (Rieger, 2008).
The intended utilization of LiDAR sensors for aircraft guidance
does not require a highly detailed GIS. We limit the creation of
a database to the extraction of planar patches in multi-aspect
ALS point clouds of the relevant urban area. As mentioned
before, these data should be collected under optimal conditions
(Figure 2a). The combined complete 3D point cloud contains
information concerning all facades and rooftops of buildings. A
workflow of off-line processing methods is used to filter points
and extract most of the planar objects. The respective flowchart
is illustrated in Figure 3.
Figure 3. Flowchart of the model creation.
Figure 2. Horizontal cross-section of a building in overlapping
point clouds: (a) after INS/DGPS postprocessing,
(b) using the real-time navigation solution.
In the second mode of ALS operation, the data is used for on
line navigation updates during helicopter missions. At this time,
we expect non-differential GPS conditions, GPS dropouts, and
loss of data points due to smoke, fog, or other negative
influences. Figure 2 shows the accuracy that can be obtained in
the different operating modes. ALS data in this example were
acquired at a skew angle of 45 degree (forward-looking). The
helicopter approached the same urban area from six different
directions, and the resulting 3D points were combined into a
Merging of several multi-aspect ALS data sets results in an
irregularly distributed 3D point cloud. We introduce a k-d tree
data structure to handle automatic processing of these data. The
search for nearest neighbors can be done very efficiently by
using the tree properties to quickly eliminate large portions of
the search space.
The subsequent segmentation method is intended to keep only
those points that are most promising to represent parts of
buildings. At first, we remove all ground points by applying a
region growing technique in combination with a local analysis
of height values. We search for sections of the point cloud in
which the histogram of height values clearly shows a