IEEE Robotics & Automation Magazine - September 2014 - 37

(m)

whole map. Figure 18 shows the pose
graph of the final map. The black lines
between the cameras of different submaps show the detected loop closures.
The global bundle adjustment is able to
35
remove the drift in the individual sub30
maps; and thus, the resulting global map
10
25
is drift free and in the correct absolute
8
20
6
(metric) scale.
15
4
(m)
A 3-D occupancy map was built, as
2
10
0
described in the "3-D Mapping" section.
5
0
5
10
15
0
20
Out of the 3-D occupancy grid, a height
25
30
35
40
45
map was generated (Figure 19) and fed
(m)
to the CAO algorithm to compute the
optimal-coverage poses. The produced Figure 19. The height map of the Zürich firefighters' training area.
map covers a 42 m # 32 m area with a
maximum height of 8.3 m. The final poses for the optimal sure-sensor height measurements. Second, recent work on
surveillance coverage of the area by the three MAVs are visual-inertial sensor fusion proposes online calibration of the
shown in Figure 20.
time offset between the two sensors [44], [45]. While such
Figure 10 shows a textured visualization of the 3-D envi- approaches have high theoretical value, in our experiments, we
ronment map of the firefighter area created from three MAVs. did not see noticeable differences when increasing or decreasing this offset of maximum 5 ms. This change is significantly
Lessons Learned
larger than the accuracy of common time synchronization
protocols like NTP, including jitter on universal serial bus
Visual-Inertial Sensor Fusion
(USB) connections. We estimated once a fixed delay in USB
The flight of more than 350 m outdoors in an unprepared transmissions but did not adapt this estimate during flights or
environment (Figure 15) revealed important insights about the between missions. Third, in the beginning of the project, we
system running under real-world conditions. First, the observ- experienced significant issues of the visual pipeline [original
ability analysis of the system, described in the "Local Naviga- parallel tracking and mapping algorithm (PTAM)] in self-simtion" section, shows that the system requires excitation to ren- ilar outdoor scenes. Map failures occurred often and marked
der all states, and, in particular, the visual scale factor, the end of the mission. Our improvements, described in detail
observable. Our tests showed that, under real conditions, this in [59], were key to ensuring continuous operation of the
requirement is generally fulfilled. We observed that initializing MAV. The most important adaptations include modifying
the visual scale factor correctly (up to about 10% of the true PTAM to a visual odometry framework with constant compuvalue) is crucial for proper state convergence. In our experi- tational complexity as well as improved feature handling, drasments, we initialized the scale factor either by GPS or by pres- tically reducing false positives in the map-building process and

(a)

(b)
Figure 18. The (a) top and (b) front views of the pose graphs of the
three flight trajectories in Figure 17 after map merging and global
bundle adjustment. The black lines show the loop closures between
the three submaps.

Figure 20. The final configuration of a robot team performing
surveillance coverage: the red squares represent the final positions
of the MAVs, while the red areas represent the invisible part of
the map.

september 2014

IEEE ROBOTICS & AUTOMATION MAGAZINE

Table of Contents for the Digital Edition of IEEE Robotics & Automation Magazine - September 2014