door environments, and the use of appearance-based
mapping requires large numbers of distinct features
for accurate odometry and loop closure. RTAB-Map
does not provide any correction on the odometry es-
timation between two successive frames, only large
loop closures are subject to optimization once a map
has been created. In the agricultural context the great-
est challenge is traversal across farmers’ fields, where
very few nearby visual references are available for
odometry and loop closure. Figure 5 shows a naviga-
tional map obtained from a farm field with a fenced
boundary, in which good visual odometry is main-
tained as long as distinct objects around the bound-
ary are present. This represents a relatively easy map
to obtain at a speed of approximately 0.5 m/s due to
the diversity of features available and navigation is
straightforward as long as the boundary is in view of
the rover. A more difficult situation is shown in 6,
where the farm field boundary is grassy and lacking in
distinct features. A slower forward speed of 0.2 m/s or
lower is necessary to perform mapping while travel-
ling along the boundary of this field, and visual odom-
etry is less consistent. It is still possible to maintain
good odometry when travelling along the perimeter
of this field as long as some features are tracked well.
However, as soon as the boundary is out of the rover’s
view when entering the centre of the field, odometry
is easily lost and the quality of the map is compro-
mised.
To evaluate the quality of odometry in an outdoor
environment, a test was run outdoors in which five
points were chosen in a 100 squared meter area and
their relative position from a starting point was mea-
sured. The rover was manually driven through the five
points, and the odometry estimation was recorded.
Table 1 reports the average error between real position
and recorded odometry under different configurations
of the extended Kalman Filter node. The left column
highlights the sensor information that has been fused
in the EKF node, and the right column contains the
average error in meters. The smallest error recorded
in this test was 1.77 meters.
Our experiment confirmed that while the laser
scanner was the most reliable sensor indoors, provides
little reliable information in a large outdoor area. The
irregular shapes of the plants does not only offer a
challenge to Hector-Slam node in the reconstruction
of a 2-D laser map, but the uneven terrain adds noisy
shifts in the laser readings. As 3D reconstruction at-
tempts proved that the error was too large to produce
a full 3-D map of the area under test, emphasis in the
mapping process is now being shifted to producing a
reliable 2-D obstacle map that can be overlaid on a
topographic elevation map.
4 LOCAL PLANNER
In our original design we planned to use the
global planners and local planners available for the
move base node of ROS, which provides a general
infrastructure to integrate different pairs of global and
local planners. The node requires a 2-D occupancy
map and a set of goal locations to compute a global
plan and output velocity commands in real time to
the drive motors until the destination is reached. Ini-
tially, the basic Navfn global planner in ROS was
tested on the rover since it provides valid planning
around mapped obstacles. However, it was found that
nearly all ROS-based planners made critical assump-
tions about the underlying mobility hardware, in par-
ticular that commands and odometry feedback would
be executed immediately with negligible time lag and
that small incremental movements would be possible.
As the rover platform uses four-wheel steering and
takes time to accelerate and turn, the planning algo-
rithms would repeatedly issue commands for small,
incremental angular adjustments (or in the case of
the ROS teb local planner, back-and-forth steering
movements) and when this failed due to either actua-
tor lag or lack of precise movement, execute recovery
behaviours. To remedy this problem, we developed
a new and simplified planner that relaxed the con-
straints on timing and precision in favour of a best-
effort approach.
For trial testing, we created a very simple al-
gorithm: at the start the rover rotates on the spot
and aligns itself along the direction of the global
path, then it tries to follow the path by steering
when needed either when the direction of the curves
changes too much or the rover moves too far away
from the plan. Once the destination is reached a sig-
nal is sent to the navigator nodes which may provide
another goal or halt the rover. The planner uses five
states of operation with appropriate commands sent
to the drive motors for the duration of each state:
Start, Forward, Backward, Pause, and Rotation. A
flowchart of the process the planner uses when mov-
ing from the current movement state to the next move-
ment state is given in Figure 7.
Making use of the high-quality odometry informa-
tion available by using Hector SLAM and the scan-
ning laser rangefinder in the lab environment, this
planner showed much better reliability on the rover
platform. In the trial for the planner, seven different
points were set spaced far apart on the lab floor and
the rover was directed to visit them autonomously in
order. Table 2 reports the distance between the set
goals and the point the rover stopped at, and an aver-
age error of 15cm was recorded at the points the rover
stopped. Figure 8 shows the 2-D Hector SLAM and
Autonomous Navigation with ROS for a Mobile Robot in Agricultural Fields
83