IEEE Robotics & Automation Magazine - December 2013 - 94

Table 1. Attained results and computational costs
in the three experiments.
Experiment

First

Second

Third

Length

203.3 m

332.7 m

314 m

Total poses

614

983

948

Remaining poses

323

511

462

Submaps

Average processing
speed in unoccupied
grids

1.99 Hz

1.97 Hz

1.83 Hz

Average processing
speed in occupied
grids

3.92 Hz

4.52 Hz

4.01 Hz
(a)

Time (ms)

340
31 ms

320
300
280
First Place

Second Place

Third Place

Figure 13. The average computational time for the runs of each
experimental environment.

significant amount of odometry error in the three experiments,
the proposed method successfully built accurate maps.
The meaningful data and computational costs for the proposed method in all three experiments are summarized in
Table 1. First, the remaining poses as nodes of the graph in
all submaps did not increase with the continued traversal of a
region due to the overlapping pose elimination in occupied
grids and they were much smaller than the total poses.
Additionally, the number of submaps was very small.
Therefore, the hierarchical pose graph optimization process
could be performed in a constant time. For this reason,
the proposed visual SLAM algorithm could operate at about
2 Hz for correcting the robot trajectory stored as nodes of the
graph in unoccupied grids and about 4 Hz for performing
the robot localization in occupied grids. These processing
speeds were measured in our embedded vision board with
an ARM11 processor running at 533 MHz while multiple
tasks, such as motion control, path planning, and obstacle
map building, were being concurrently performed.
Figure 13 shows the average computational time per
image frame for each experiment. The lowest computational
time is shown in the second experimental place because the
processing time to obtain a valid loop closure constraint in
the second experimental place was lower than the other
experimental places. However, the computational time difference between the three experiments is comparatively
small, as shown in Figure 13. This result verifies that the performance of the proposed method is not sensitive to a
change in the environment.
94

IEEE ROBOTICS & AUTOMATION MAGAZINE

DECEMBER 2013

(b)
Figure 14. A large-scale mapping result; the red lines show the
robot trajectories estimated by odometry and our visual SLAM,
while the blue obstacle grid maps built using the PSD sensors
are used to visualize the localization accuracy. The ground truth
(including big obstacles such as doors and walls) is shown by
the green dotted lines. (a) The odometry map. (b) Visual SLAM.

Large-Scale Map Building on
an Embedded Vision Board
An online experiment was conducted to evaluate the performance of the proposed method in a large-scale indoor building. The environment was 28 # 40 m, and the sensor data
were saved by the home cleaning robot that was driven manually along the walls. The robot's path covered about 233 m,
and 706 images were collected.
A significant amount of odometry error was accumulated
in this experiment, as shown in Figure 14(a). However, the
proposed method built an accurate map in real time.
Figure 14(b) shows the comparison result between the final
map at the end of exploration and the ground truth data
drawn with green dotted lines.
In the experiment, the processing speed of the proposed
method was almost the same as that of Table 1 because the robot
pose error was not significantly accumulated when closing loops
between local submaps and the hierarchical pose graph optimization process could be performed in a constant time.
This experimental result demonstrates that the proposed
method is a practical approach for real-time large-scale
mapping on low-cost robot platforms with limited sensory
and computational equipment.

Table of Contents for the Digital Edition of IEEE Robotics & Automation Magazine - December 2013