IEEE Robotics & Automation Magazine - September 2012 - 54

PHILIPP LUTZ

than that of laser odometry. Therefore, the system switches to visual
odometry.
Shortly afterward, the system is
commanded to fly outside. At 21 s,
visual odometry becomes unavailable
as indicated on the out-of-axis covariance in Figure 9. This is caused at first
by motion blur when the quadrotor
starts moving in the weak lighting con(a)
(b)
ditions in the house. The system automatically switches to laser odometry.
Figure 6. Details of avionics hardware components. (a) Modular computation stack.
During flight through the 1-m-wide
(b) An eight-port fast Ethernet switch.
window, a jump in the raw laser height
Experimental Results
measurement can be seen due to flying above the 20-cm-wide
A flight experiment was conducted to show the effectiveness wall. The jump is detected by the data fusion and a constant altiof using two sensing paradigms. Figure 7 shows the esti- tude is kept. The vertical velocity is also unaffected, so the vehicle
mated path flown inside and outside the experimental facil- passes smoothly through the window. The visual odometry is
ity (depicted in Figure 8). Figure 9 shows the reference and still unavailable as the window pane is too close to the cameras.
estimated system states as well as absolute covariances of
When the quadrotor is outside, the cameras need to
the two odometries. Time instants when the system switches adjust their exposure time, so visual odometry is again
between visual and laser odometry are marked on the time available at approximately 1 m behind the window. It is
axis. Because of practical difficulties in outdoor measure- clearly visible that the covariance of the laser odometry
ments, no ground truth is provided. A rectangular path outdoors is very large compared to indoors, due to less
around the house was chosen because it was clear of ground valid laser measurements. Therefore, only visual odometry
obstacles. The quadrotor has been yawing during the flight is used for the outdoor flight.
so that the laser scan points toward the house.
During autonomous landing, the disparity image
The quadrotor starts inside the house using laser becomes unavailable under 60 cm of altitude. This is indiodometry. Visual odometry is not available because cated by the high covariance of the visual odometry, and
the cameras are too close to the ground, and so there is the system switches to laser odometry.
not enough overlapping in the images. At 7.5 s, during
The reference velocity and position are tracked accordautonomous take-off, when the quadrotor is at 68 cm ing to the estimated values. The position control error with
altitude, the depth image becomes available, and respect to the estimated states is under 20 cm in both
the covariance of the visual odometry becomes smaller indoor and outdoor environments.
5
Laser
4
3
-z (m)

-y (m)

Laser
2
1
Visual

Laser

2
1

Visual

Laser
0

Visual

4
3

−1

1
−4

−3

−2
−1
x (m)

(a)

-y (m)

0
−1 −4

−3

-2

-1
x (m)

(b)

Figure 7. Flown path estimated in the experiment. The dashed red line shows the reference trajectory, while the solid line shows the
estimated path. The house outline is shown in gray. Locations where switching between visual and laser odometry occurs are also
indicated. (a) Estimated flown path in the horizontal (x, y) plane. (b) Three-dimensional view of the desired and estimated position
(x, y, z) of the quadrotor during the experiment.

IEEE ROBOTICS & AUTOMATION MAGAZINE

SEPTEMBER 2012

Table of Contents for the Digital Edition of IEEE Robotics & Automation Magazine - September 2012