Graphics Reference
In-Depth Information
Fig. 4.6
Asctec Pelican quadrotor flying through forest
only provides information when the aircraft is moving, and poor estimation of time
to collision near the focus of expansion limits the ability to perceive obstacles in
the direction of motion. Stereo vision can provide 3D perception around the focus
of expansion whether the aircraft is moving or not, but so far with relatively short
look-ahead distance [
18
]. Stereo and optical flow have been used together in purely
reactive obstacle avoidance based on sparse perception with point features [
23
].
The most common obstacle representations for MAV are image space data
products that serve reactive obstacle avoidance [
14
,
25
,
50
] and Cartesian voxel
data structures that serve deliberative planning [
5
,
18
,
56
]. Reactive obstacle avoid-
ance with image space data structures use very little memory and computation, but
have limited ability to reason about 3D structure and vehicle dynamics. Cartesian
voxel data structures enable much greater 3D reasoning, have mature temporal fusion
algorithms for error reduction, and have been used to plan high speed, aggressive
maneuvers [
48
]; however, they use much more memory and computing time. Uniform
voxel sizes are also problematic for representing both very near and very far objects,
which can lead to more complex, multiresolution data structures. Polar representa-
tions parameterized by azimuth and range have been used in a few efforts because
they naturally capture range-dependent variations in angular and range resolution
[
7
,
68
]; however, these efforts were only tested in simulation and [
68
] only repre-
sented sparse, discrete obstacles. Some stereo vision-based navigation systems for
ground vehicles have had characteristics that are interesting for MAVs. A simple ver-
sion of collision testing and path planning with the stereo disparity image was done
in [
44
]. A 2D polar grid-based representation in the ground plane was used in [
6
],
where the radial axis was parameterized as inverse range; this matched the angular
and range resolution characteristics of stereo and gave a compact representation of
all space from a minimum range to infinity. This was used to represent and reason
about distant obstacles, while a 2D Cartesian map was used for nearby obstacles.
Inverse range is equivalent to nearness fields that have been used for reactive MAV
obstacle avoidance with optical flow [
25
].
