IEEE Robotics & Automation Magazine - December 2013 - 93

Processing Speed on an Embedded Vision Board
Three online experiments were conducted in different
indoor environments to analyze the computational cost of
the proposed method on our embedded vision board.
Numerous images were collected by the home cleaning

350
Conventional Pose Graph
Optimization Method
Proposed Method

300
250
Time (ms)

violates an orthogonal structure assumption existed (see the
right part of maps in Figure 10). However, the visual compass
did not suffer from this problem because the ratio of 0.7 (as a
threshold for the visual compass) used in this study was not
satisfied in the vicinity of the area (refer to the "Visual
Compass" section for more details). Figure 10 illustrates a
typical performance comparison between "before" and "after"
the proposed visual compass, consistent among a large set of
experiments. This indicates that the proposed visual compass
provides a visual SLAM with an excellent means of enhancing
its performance, as long as the environment where the visual
SLAM is applied allows proper features for the visual compass. It would be interesting to see the visual compass used as
a virtual sensor for global orientation in the future for various
other robot applications.
Second, we evaluate experimentally the computational
efficiency associated with the embedded visual SLAM with
the proposed visual compass in comparison with the conventional graph-based SLAM. More specifically, we measure the execution time for the proposed method and for
the conventional pose graph optimization method based on
an Intel Core i7-2620M CPU running at 2.7 GHz. The conventional pose graph optimization method is to correct all
poses stored in a global map using their loop closure constraints such that its computational complexity is similar to
the standard EKF SLAM.
Figure 11 shows the computational cost per image frame
in terms of the number of frames (up to 1,132 frames) to be
processed for the conventional pose graph optimization
method (in yellow) and for the proposed method (in red).
As seen in Figure 11, the computational cost of the proposed method is very low compared to the conventional
pose graph optimization method. More importantly, the
execution time of the proposed method does not increase in
terms of the number of frames to be processed. This is due
to the fact that the proposed method resorts to 1) an efficient hierarchical pose graph optimization process with the
two linear estimation problems separately applied to orientation and translation, where the proposed visual compass is
to minimize the linearization errors and 2) a localization
process for eliminating overlapping poses based on an occupancy grid map. Note that the computational burden associated with the hierarchical management of pose graph optimization turns out to be negligible since there are only 39
submaps generated at the end of the exploration. The high
level of computational efficiency as well as the complexity of
the proposed method, as demonstrated in Figure 11, clearly
verifies why the proposed method offers a computationally
feasible approach to the development of low-cost embedded
visual SLAM.

200
150
100
50
0

201

401

601
Frame

801

1,001

Figure 11. The execution time per image frame in terms of the
number of frames for the conventional pose graph optimization
method (in yellow) and for the proposed method (in red).

(a)

(b)

(c)

Figure 12. Maps constructed at the end of the exploration of
the three experimental environments. Top row: the obstacle
grid maps generated using odometry. The black lines show the
robot's movement trajectories during exploration. Bottom row:
the obstacle grid maps built using our visual SLAM. The red lines
show the robot trajectory for each submap in our hierarchical
map structure. (a) First place. (b) Second place. (c) Third place.

robot moved based on the aforementioned zigzag motionbased navigation.
When the zigzag motion-based navigation was used to
build a map, there were submaps with a few poses during
exploration. However, the submaps could not be used to obtain
a loop closure constraint. For this reason, the submaps with a
few poses (which have continuity with the time) merged into a
single submap to reduce the number of unnecessary submaps,
and they were disregarded in representing the whole map. Note
that this scheme does not influence the map accuracy.
Figure 12 illustrates the obstacle maps generated using
odometry and the proposed method at the end of the exploration of the three experimental environments. The robot trajectory stored as nodes of the graph for each submap is drawn with
red lines, and the obstacle grid maps built with the PSD sensors
are represented using blue grids. Although there was a
DECEMBER 2013

IEEE ROBOTICS & AUTOMATION MAGAZINE

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