Computer Vision for Micro Air Vehicles - Advances in Embedded Computer Vision - page 89

Fig. 4.13 Simulated flight through virtual forest: a Initial flight trajectory from the vehicle posi-
tion at the bottom toward the goal at the top with overlaid 3D C-space point positions ( red ); the
horizontal red lines in front of the vehicle correspond to C-space points above the ground; b top-
down view during traverse with overlaid current view of the vehicle

Fig. 4.14 Flying through a door opening with added pose noise: Scene view from the starting
position with overlaid C-space point cloud ( red )

With no pose noise added, the vehicle was able to fly safely through the opening
on each run (Fig. 4.15 a). Figure 4.15 b illustrates the same flight experiment, where
limited white noise ( ) was added to the position of the collision

sensor (stereo camera system) prior to each planning cycle. When treating the pose of
the collision sensor (with noise overlay) as true vehicle pose, all node positions on
the RRT-tree become erroneous, simulating the effect of a noisy pose estimate on the
planning process. In this case, collision checking invalidates additional tree branches,
due to the added noise, forcing the vehicle to re-plan a valid trajectory when needed.
As shown in Fig. 4.15 b, additive noise affects the planning result only marginally.
The vehicle was able to execute all runs properly and pass the opening at all times,
since the added noise had zero mean (no drift).

ʷ ∈[− 15 cm , 15 cm ]