Graphics Reference
In-Depth Information
Fig. 4.16
Flying under tree canopy:
a
top
to
bottom
: left rectified stereo camera view; disparity
map; C-space map with expansion size of 60 cm;
b
top-down
view of flown trajectory (
red
)with
overlaid accumulated C-space point clouds (
green
) that are generated from subsampled C-space
disparity maps; only points are shown that are more than 2.5 m above the ground and not farther
away from the vehicle than 2 m. The
inset
shows a view from behind the vehicle in an upright view,
illustrating flight below a tree branch
used 9 %, and motion planning used 65 % (RRT planner 60 %, trajectory server 5 %),
generating motion plans at 2 Hz. On average, 6.51 % of the time used by the RRT
planner was spent on C-space expansion, and 8.19 % on collision checking, which
demonstrates the effectiveness of our approach. Additionally, C-space expansion and
collision checking required less than 1.5 MB of memory for processing 376
×
240
disparity images from the stereo system. In total, the navigation frame work used
133 % of the available 200 % processing power of our two core CPU.
Figure
4.16
a gives an example of the viewing volume that is used for motion
planning when approaching a tree. Figure
4.16
b illustrates the flight trajectory of a
representative experiment in a top-down view. The vehicle started at the left of the
image and was commanded to fly to the right, avoiding four different trees on its
traverse.
At the beginning of each experiments, we started generating motion plans at an
altitude of 4 m and restricted the planning volume to sample 3D points between 2
and 5 m altitude. The vehicle was able to avoid obstacles autonomously and plan
its motion around tree branches. This included a portion of the trajectory where the
vehicle flew correctly below an overhead branch (tree 2, Fig.
4.16
b).
