The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part BI. Beijing 2008
1238
configuration, in comparison, requires dynamic heading
alignment and delivers heading measurements that suffer from
drift and the rate of drift is heavily dependent on vehicle
dynamics. GAMS uses a carrier phase differential GPS
algorithm to measure the relative position vector between the
two antennas. These carrier phase measurements from five or
more satellites are used to estimate, and eventually, to identify a
set of integer phase ambiguities for each satellite being tracked
by both receivers. For the ambiguity resolution algorithm to
work, both receivers must track at least five common satellites.
Once tracking has been obtained, GAMS will continue to
operate with as few as four satellites. The GAMS heading
system will not provide measurements when fewer than 4 GPS
satellites are available. During GPS outages, POS LV will
continue to provide accurate heading measurements drifting at
the rate of about 1 arc min/min. Accurate heading is critical for
robotic vehicle navigation especially when intermittent or non
existent GPS conditions occur over extended periods of time.
The distance measurement instrument (DM1) is another
essential piece of the POS LV hardware which outputs pulses
representing fractional revolutions of the instrumented wheel.
These pulses are converted by the POS LV into measurements
of incremental distance travelled by the vehicle when no GPS is
available. In the 2005 DARPA Grand Challenge both Red
Team vehicles Highlander and Sandstorm utilized DMI data
not only to bridge GPS outages and provide POS LV with
incremental distance estimation, but as an input into the velocity
controller for detection of when the vehicle may be stuck.
Wheel slippage is monitored by comparing the DMI output to
the velocity reported by the POS LV system. When the system
reports speeds over 5m/sec and a velocity of 0 m/sec., the
vehicles execute a set of protocols utilizing the perception
system and POS LV data to find an alternate path to the next
pre-programmed point.
3. THE 2005 DARPA GRAND CHALLENGE - A
CHALLENGE IN SENSOR FUSION
For the 2005 DARPA Grand Challenge Applanix collaborated
with Carnegie Mellon’s Red Team for the position and
orientation component of their two entries into the race,
Highlander and Sandstorm [2]. Both vehicles utilized a series
of LIDAR and radar systems to sense terrain and feed that
information into onboard computers which would modify pre
planned route information to avoid obstacles and deal with
changes in terrain. The data provided by the POS LV is
Figure 3: Sandstorm (left) and Highlander (right) were
developed to navigate at high-speed in desert terrain.
essential in governing vehicle dynamics to safely navigate the
course for real time operation. The Red Team’s approach
involved a methodical analysis of the course terrain and
modification of the RDDF (the DARPA defined route definition
file) in order to provide both vehicles the optimum path.
LIDAR data (provided through a gimbal located on the roof of
the vehicle which provided medium and long range terrain data)
and supplemental lasers (scanning the boundaries of the
navigable track) in addition to the short range radar (vital for
detecting targets in the immediate vicinity of the vehicle), were
incorporated to form a view of the world within which the
robots would sense and evaluate terrain. Position information
from the POS LV is critical in determining the direction of
rotation of the gimbal in order to sense the oncoming terrain and
georeferencing point cloud data [3].
The Red Team utilized a path-centric architecture which
provided a simple method for incorporating a pre-planned route.
The primary reason in utilizing this approach was to reduce the
search space for a planning algorithm from the square of the
path length to linear in the path length, since planning is
performed in a corridor around the pre-planned route. The path
centric approach avoided problems with arc-based arbitration
such as discontinuities in steering commands (due to
contradictory information) and jerky control (due to discrete
arc-sets) [4].
With data derived from the LIDAR and radar systems, it is
fused into a composite model of the terrain as illustrated in
figure 4. The data is processed and is assigned a value dictating
its ‘cost’. Lower elevations (shown in green) are assigned a
lower cost whereas higher elevation (shown in red) is assigned a
higher cost. The autonomous vehicle is ‘trained’ to navigate on
the low cost sections of terrain and make modifications to its
pre programmed route in the event obstacles or terrain
anomalies block its intended path. Accurate position and
orientation estimation is essential to this process. Map fusion is
critical to the robustness of the navigation process, as it enables
the system to cope with sensor failures and missing data. In
addition to this, deriving data from multiple sources compares
sensor input to account for anomalies. If a sensor is damaged
and not providing accurate data, the processing algorithms will
accord that sensor input a lower degree of confidence and adjust
its contribution to the overall weighting of the data
characterization and mapping process.
Figure 4: An example cost map showing low and high cost
terrain.
Errors in terrain characterization can, in most cases, be
attributed to errors not in the data acquired by the sensor, but by
errors in position and orientation estimation. As demonstrated
in Figure 6a which shows test data from Stanford University’s
2005 entry ‘Stanley’, inaccurate pose can cause the vehicle to
stop (Figure 6b) as all oncoming terrain will be perceived as not
being traversable. The illustrations mark red terrain as not
traversable, white is low cost terrain and grey areas are not
known. The blue corridor is the DARPA assigned route. This
pose error of less than 0.5 degrees in roll/pitch is enough to
force the vehicle off the course if ignored [5]. In tests carried
out by the Team, referenced terrain was erroneously classified