IEEE Robotics & Automation Magazine - December 2013 - 26

been chosen, which corresponds to the smallest half-dimension of the vehicle.
Several information data are associated to each cube: the
number of inliers (3-D triangulated points) had fallen into the
cube volume, the last camera position from which an inlier has
been found, and the state of the cube. In detail, the number of
inliers represents the number of different points of view from
which the same obstacle
has been detected; the last
c ame r a p o s it i on is
When a 3-D triangulated
required in case of hovering to prevent the same
point is detected in the
image features from generating several wrong inlicube, then the value of the
ers, while it is possible that
cube inliers is incremented the same outlier is
achieved from different
points of view; finally, the
and the state is set to
state of the cube is
employed to generate the
"OCCUPIED."
force field, as will be
described in the next section, and it can be one of the following values: "FREE,"
"OCCUPIED," "OBSTACLE," "IGNORED," or "TARGET."
Initially, each cube is set to "FREE." When a 3-D triangulated point is detected in the cube, then the value of the cube
inliers is incremented and the state is set to "OCCUPIED."
When the number of inliers reaches the assigned threshold,
the state is set to "OBSTACLE." On the other hand, when a
target is identified, the relative cube state is set to "TARGET."
Moreover, since the camera can observe the target from each
position that had generated a valid target viewpoint, all the
cubes laying along these optical rays are set to "IGNORED."
In this way, the map also implicitly provides a free way to
reach the target. Figure 6 shows the reconstructed occupancy
map corresponding to the environment of Figure 5.
For wide environments, a sparse representation of the
occupancy grid map is considered together with a spatial/temporal vanishing criterion. This determines whether an
Points 3-D (466)

occupancy cube is reliable or if it has to be discarded, based on
the distance traveled by the vehicle or on the time interval
spent after its last update. In fact, due to the drift of the vehicle
pose estimation, obstacles that have been observed far from
the current position or a long time cannot be considered reliable in the current map representation, and they must be
refreshed. In this way, the reliability and scalability of the map
representation can be tuned with respect to any environment.
In Figure 7, the environment map of an outdoor unknown
space has been reconstructed by using only the onboard visual
odometry estimation system and a sparse representation of the
grid. In a particular point, three texturized objects have been
positioned in the environment: the UAV has started from a
position close to object number three and then it has been
moved away by traveling from object number one to two. 3-D
points matched at a distance more than 5 m from the current
UAV position have been discarded during map construction to
avoid noise due to the low resolution of the vision system at
greater distance. Analogously, all the points that are close to the
floor have been discarded to avoid unnecessary information.

Object 3

Object 1

Object 2

Object 3

Object 1

200
Target

z

150
100
50
0
200

Object 2
Object 2

Object 1

200
0

100

0

-100
y

-200 -200

x

Figure 6. An example of a reconstructed map of the environment
shown in Figure 5.

26

*

IEEE ROBOTICS & AUTOMATION MAGAZINE

*

DECEMBER 2013

Figure 7. An example of an occupancy grid map of an outdoor
environment that has been achieved by using only the onboard
vision-based odometry.
Table of Contents for the Digital Edition of IEEE Robotics & Automation Magazine - December 2013
