See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/228962857

Visual Tracking and Control of a Quadcopter Using a Stereo Camera System and
Inertial Sensors

Article · August 2009

DOI: 10.1109/ICMA.2009.5246421

CITATIONS READS
96 5,748

4 authors, including:

Markus Wilhelm Achtelik Kolja Kühnlenz

ETH Zurich University of Applied Sciences Coburg
34 PUBLICATIONS   3,016 CITATIONS    118 PUBLICATIONS   1,810 CITATIONS   

SEE PROFILE SEE PROFILE

Martin Buss
Technische Universität München
575 PUBLICATIONS   10,543 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Autonomous Outdoor Shopping Assistant View project

Emotion Display with Dynamic Intuitive Expressions (EDDIE) View project

All content following this page was uploaded by Markus Wilhelm Achtelik on 21 May 2014.

The user has requested enhancement of the downloaded file.

 Visual Tracking and Control of a Quadcopter
Using a Stereo Camera System and Inertial Sensors∗
Markus Achtelik, Tianguang Zhang, Kolja Kühnlenz and Martin Buss
Institute of Automatic Control Engineering (LSR)
Technische Universität München
D-80290 Munich, Germany
[email protected], {tg.zhang, kolja.kuehnlenz, m.buss}@ieee.org

Abstract— In this paper a complete system is designed and

implemented, in which the motion of a quadcopter is sta-
bly controlled based on visual feedback and measurements
of inertial sensors. We focus on developing a cost effective
and easy-to-setup vision system. Active markers were finely
designed to improve visibility under various perspectives as
well as robustness towards disturbances in the image-based pose
estimation. Moreover, position- and heading controllers for the
quadcopter were implemented to show the system’s capabilities.
The performance of the controllers was further improved by the
use of inertial sensors of the quadcopter. A closed-loop control
system is successfully conducted. Fig. 1: The quadcopter with illuminated markers

Index Terms— Vision system, motion estimation, multiple

sensing system, UAV/MAV control. novel correction scheme compensating accumulated errors
was integrated. However, the real time capability of them is
I. I NTRODUCTION
not sufficient for quadcopter motion control. In [2] a position
In recent years, unmanned aerial vehicle (UAV) and micro controller for a quadcopter was implemented using an optical
aerial vehicle (MAV) are major focuses of active researches, tracking system by VICON in the “Holodeck” lab at MIT.
since they can extend our capability in a variety of areas, This intelligent space is nevertheless very expensive and not
especially for applications like search-and-rescue and inspec- transportable. 3D-pose estimation and control of a quadcopter
tion. A significant challenge in developing UAVs is to extract using only visual feedback had been done in [5] based on
and fuse the useful information in a robust manner and to one camera fixed on the ground an additional second camera
provide stable flight and an accurate navigation. To exploit mounted on the quadcopter facing downwards. The flight
various flight behaviors and control strategies of UAVs, we volume in this work is very limited and the feature detection
use a quadcopter [1] [2] as testbed. applied is sensitive to the light condition and therefore not re-
The control of quadcopter during autonomous flights re- liable enough. We see our system not competing with systems
lies on knowledge of variables like position, velocity and developed in [6] [7] where on-board sensors such as IMU,
orientation, which can be partly calculated using information camera and sonars or laser rangefinders are used for pose
provided by on-board inertial sensors. However, the drift of estimation and obstacle avoidance for an autonomous flight.
inertial sensors leads to errors during time-discrete integra- In contrast, it could be an extremely helpful addition during
tion, making a steadily accurate estimation of the absolute the development of onboard pose estimation systems men-
pose nearly impossible. Therefore, an absolute reference is tioned above by providing absolute position information and
needed, e.g. a GPS system for outdoor applications. Indoors, ground truth. These algorithms usually need a well structured
where no GPS is available, a visual tracking system could environment that might be totally different from laboratories
be applied. In [3], an efficient iterative algorithm is proposed where tracking systems like VICON are installed, for which
in order to achieve a globally convergent pose estimation reason we like to emphasize the transportability and easy
of an object with known 3D geometry. Another accurate setup of our system.
pose estimation of moving rigid object using a calibrated To deal the aforementioned problems, a complete system
stereoscopic vision setup is described in [4], in which a is designed and implemented, in which the motion of a
∗ This work is partially supported by the DFG excellence initiative
quadcopter is stably controlled based on visual feedback and
research cluster Cognition for Technical Systems – CoTeSys and the Bernstein measurements of inertial sensors. We focus on developing a
Center for Computational Neuroscience Munich. cost effective and easy-to-setup vision system. We use an off-
board stereo camera setup consisting of two web cameras to for each axis of rotation and three acceleration sensors for
provide accurate position information. Active markers were each translation axis. The IMU estimates the absolute roll-
finely designed with high-power LEDs of different colors. and pitch angles by fusing the acceleration vector and angular
This design allows visibility under various perspectives and speeds measured by the gyroscopes. A relative yaw-angle is
improves robustness towards disturbances. Finally, the image- estimated by integrating the angular speed measured by the
based pose estimation was experimentally evaluated using yaw-gyro. The IMU is equipped with two processors - one
a positioning table. We also show the implementation of “black box” performing the low level attitude and heading
a position- and heading controller as well as an approach control and another high-level processor for user specific
on how to incorporate IMU data. Hovering and autonomous applications. Communication to a ground station is realized
flight are excellently implemented as shown in video. over a serial port and wifi modules forwarding serial data
This paper is organized as follows: firstly, in Section II, we transparently, like XBeePro modules from Digi2 .
describe the hardware configuration in our application as well As the quadcopter itself is unstable during flight, all three
as their functional principle. In Section III, marker detection axes of rotation (roll, pitch, yaw) are already stabilized by the
and pose estimation of the quadcopter are introduced. Pro- IMU with three independent PD controllers on each angle at
cessing times and accuracy are evaluated in Section IV. The a control loop frequency of 1 kHz [2]. Using these inertial
implementation of the controllers and experimental results sensors, it is only possible to determine the roll and pitch
are shown in Section V. Conclusions are given in Section DOFs absolutely without drift or integration. In all other
VI. DOFs, only angular velocities or translational accelerations
can be measured directly. Although the errors due to time
II. HARDWARE CONFIGURATION discrete integration are small at the rate of 1 kHz, an external
A. Quadcopter reference is needed to compensate for the drift and to control
all six DOFs of the quadcopter.
The 6 DOF of the quadrotor are controlled by four inputs
(roll, pitch, yaw, thrust) by varying the lift forces and the
balance of reaction torques through changing the rotating B. Cameras
speed of the rotors. Two pairs of rotors are spinning clock- As one of the goals of this work was to keep costs small,
wise and counterclockwise respectively, so that the sum of only webcameras were used and were connected to the PC
their reaction torques is zero during hovering. Unlike normal via USB. Quickcam9000 Pro cameras from Logitech were
helicopters, the propellers of the quadcopter applied in this used for this work. The cameras acquire images in VGA
work have fixed pitch angles. This minimalistic hardware resolution at a framerate of up to 30 FPS. The field of
design (no rotor linkages etc) makes the quadcopter robust view of 60◦ should be a good trade-off between accuracy
such that it can survive crashes during experiments without and maximum possible as well as equally spaced flight
getting seriously damaged. Fig. 2 shows the basic design of volume. The maximum distance, at which the markers could
the quadcopter. The z-axis points towards the ground. be recognized properly was determined to be approximately
5 m in the experiments.
Accuracy is reduced when the cameras are aligned that
their optical axes are parallel, which can be seen in Fig. 3:
there is always an unused field of view of α − α0 . Therefore,
an alignment of the cameras with a distance of 1.4 m and a
converging angle of 15◦ for each camera was determined to
provide best results.

Fig. 2: Basic design of the quadcopter

The quadrotor applied in this work is a ”Hummingbird”

with an ”AutoPilot” controllerboard from Ascending Tech-
nologies1 . The configuration of the quadcopter is basically Fig. 3: Comparison between fields of view of converging and
the same as described in [2]. The main controller board (in parallel cameras
the remainder refered to as the inertial measurement unit –
IMU) contains inertial sensors consisting of three gyroscopes
1 Ascending Technologies multi-rotor air-vehicles: http://www.asctec.de 2 Digi networking solutions: http://www.digi.com
C. Markers itself did not. The markers are bullets, so a circle in the
To be more independent from light conditions, active binary image gained from the thresholds above is expected.
markers were mounted at the end of each boom of the To reconstruct the real center of the marker, the foreground
quadcopter. The basic idea was to illuminate the markers of the binary image is corrected by morphological closing
as bright as possible and to reduce the exposure time of the to fill gaps in the foreground but keep the basic structure
cameras then. This method reduces motion blur as well as approximately and edge filtering is done afterwards. Now,
the influence of ambient light. a circle can be fitted into the edges found in the step
To compute all six DOFs of the quadcopter from the before. The deviations of the center of the markers could
tracked marker-positions, three markers would be sufficient. be reduced significantly using this method, especially on
For redundancy reasons and to avoid potential failure due to partially occluded markers. Fig. 5 shows step-by-step the
self-occlusion, four markers were applied. Two high power procedure of recovering the circle.
LEDs each were aligend inside a table-tennis ball dispersing
the light of the LEDs diffusely. As a result, a homogeneously
lit sphere could be recognized by the cameras as a circle from
all directions of view. Fig. 4a and 4b shows the design and
a lightened view of a marker.

Fig. 5: Step-by-step fitting of the circle

B. 3-D reconstruction of the marker’s positions

After the 2-D coordinates of the center of the markers
have been successfully determined in each image, their 3-D
(a) (b) coordinates can be recovered. The coordinates from markers
of the same color now serve as a corresponding pair of
Fig. 4: Design of the markers coordinates for the reconstruction. A 3-D point in space
is projected into the camera’s image planes up to a scalar
factor λ by the following equations [8]. All coordinates are
III. TRACKING AND POSE RECONSTRUCTION
represented as homogeneous coordinates.
A. Recognition of the markers  
λx1 = K1 0 · X 0 = Π1 · X 0 (1)
For each marker, the area and a center of gravity (COG)  
is calculated which is then used for computing the stereo λx2 = K2 R K2 T · X 0 = Π2 · X 0 (2)
correspondences needed for the 3-D reconstruction after- xi ∈ R3×1 denotes the projection of X 0 ∈ R4×1 into
wards. The following steps are performed for each of the the image plane of the left and right cameras through the
four markers in the image of each camera. projection matrices Πi ∈ R3×4 which consist of the intrinsic
First of all, the image is transformed from RGB into parameter matrices Ki as well as of the extrinsic parameters
the YUV-colorspace so that the search for colors can be R and T . The frame of the left camera is taken as reference
performed independently from lighting conditions. Each pixel frame. Rewriting of (1, 2) leads to four constraints for the
that falls within predifined ranges is called hit. Before com- point X 0 we wish to recover from both views:
puting the COG, a robust filter is applied: only hits whose
x1 π 3T 1T
y1 π 3T 2T



coordinates in both x and y are close to the median of the 1 − π1 X0 = 0 1 − π1 X 0 = 0 (3)
x2 π 3T 1T 3T 2T



coordinates will be considered in the calculation of the center 2 − π 2 X 0 = 0 y π
2 2 − π 2 X 0 = 0 (4)
of gravity. Compared to a median filter on the intensity T
Where xi = xi yi 1 . π jT

values of a neighborhood of each pixel, the procedure of are the j th row vectors of the
 1i T
sorting the coordinates in order to calculate the median is not projection matrices Πi = π i π 2i π 3i . (3) and (4) now
expensive: due to the line-by-line search, the y-coordinates form a 4 × 4 homogeneous system of equations that has to
are already sorted and the x-coordinates are presorted. Once be solved: M X 0 = 0. In practice, this system does not have
a valid center of gravity is found, a region of interest (ROI) a solution due to small errors in the correspondences or in
is determined around this position and search in the next the camera calibration. The best solution in the sense of least
timestep only takes place in this region. squares for X 0 is the eigenvector of M T M corresponding to
In case a marker was partially occluded by parts of the its smallest eigenvalue. To complete the reconstruction, X 0
quadcopter, the COG of a marker shifts although the marker has to be normalized so that its last coordinate is equal to 1.
C. Reconstruction of the quadcopter’s pose a “helper point” normal to the plane spanned by the markers,
Having reconstructed the 3-D positions of the markers, the was constructed for each the initial (Y i ) and current (X i )
position and orientation of the quadcopter can be determined. sets of points, which is shown in Fig. 6. To ensure that this
In [9], Umeyama proposes an approach to determine the helper point is always located on the same side of the plane,
rotation R and the translation T from two given sets of the crossproduct of the vectors connecting the markers is
points X i and Y i where the mean squared errors e(R, T ) always calculated in the same sense of rotation. This is done
are minimized. by the following steps, before computing R and T :
Y i were chosen as initial positions of the quadcopter’s 1) Span the vectors l1 and l2 between the markers
markers. As the positions of the markers were reconstructed (X f ront , X right , X back , X lef t ):
w.r.t. the left camera’s frame, this frame was also taken If all four markers were recognized:
as reference frame for the initial positions. These initial l1 = X f ront − X back ; l2 = X right − X lef t (11)
positions have the coordinates of the markers mounted at the
end of the quadcopter’s booms and lie in the x-y plane. The If only three markers were recognized: sort the markers
transformation between the initial positions and the current in the order front-right-back-left
positions of the markers can be described as follows: ⇒ X 1, X 2, X 3
Xi = R · Yi + T (5) l1 = X 1 − X 2 ; l2 = X 1 − X 3 (12)

To gain R and T uniquely in 3-D space, at least three 2) Calculate the vector normal to the plane:
correspondences are necessary. Firstly, the rotation will be n = l1 × l2 (13)
calculated. The sets of points are translated by their centroids
µx and µy to the origin and the covariance matrix Σxy of the 3) Calculate the mean vector µ as in (6)
translated sets of points is computed. n denotes the number 4) Calculate the helper point:
of corresponding pairs of points. X helper = µ + n (14)
n n
1X 1X
µx = Xi µy = Yi (6)
n i=1 n i=1
n
1X
Σxy = (Y i − µy )(X i − µx )T (7)
n i=1
The best solution for R in the sense of minimizing the
squared errors is now achieved by computing the singular
value decomposition Σxy = U SV T and setting the singular
Fig. 6: Construction of the helper points
values equal to 1. If rank (Σxy ) = 2, which is always the
case for only three corresponding pairs of points or more
than three points that are coplanar, the transformation matrix The covariance matrix Σxy calculated from X i , X helper ,
itself is exactly determined, but not uniquely if it is a rotation Y i and Y helper by (6), (7) has full rank and R will always
or a reflection. This can be checked by calculating det (R). be a valid rotation matrix. T can still be computed similar to
Then, if necessary, S has to be corrected: (10).
Thus far, the rotation and translation of the quadcopter are
R = U S̃V T (8) still computed w.r.t. the left camera’s frame. In most cases,
the camera rig will be placed “somewhere” in a room, so
(
diag (1, 1, 1) for det (R) = 1
with S̃ = (9) the pose of the camera would not be a desirable reference
diag (1, 1, −1) for det (R) = −1 frame for further applications such as position control. Thus,
Having calculated the rotation matrix, the translation vector another transformation from the camera to world coordinates
can now be gained by simply rewriting (5) and using the is required. The idea is to determine a new reference frame
mean vectors µx and µy of the two sets of points: by the frame spanned by the quadcopter lying e.g. at its start
position. The desired rotation/translation of the quadcopter
T = µy − Rµx (10) w.r.t. its start position is given by:
Although the distinction between rotation and reflection was T
R0 = RQC,start · RQC,camera (15)
made through calculating the determinant of R, this decision T T
T0 = RQC,start · TQC,camera − RQC,start · TQC,start (16)
caused problems in practice. For example, the roll angle
calculated from the rotation matrix (see below) jumped RQC,camera denotes the rotation w.r.t. the camera frame,
between 0◦ and −180◦ in certain cases. To avoid this effect, RQC,start the start position (i.e. the new reference frame) and
R0 the desired rotation w.r.t. the startposition, T respectively. absolute deviations to the trajectory ∆ were computed as a
When the camera rig’s position changes, simply R0 and T0 quality criterion which are listed in Table II.
need to be stored at the quadcopter’s start position which are
derived by the algorithms above.
IV. PERFORMANCE OF THE SYSTEM
After the implementation of the tracking of the quadcopter,
its performance was evaluated before attempting position
control.
A. Processing times and delays
To measure the total delay between a change of the
quadcopter and a control reaction, processing times of certain
algorithms were logged. In addition, the time between a
change and the moment that the image is delivered by the
camera driver was measured with a digital oscilloscope. As
it can be seen in Table I, the largest part of the delay was
caused by the cameras. This delay included an interframe
delay of 33 ms on average or 66 ms at maximum, assuming
15 FPS. Compared to the communication- and camera delays,
the other processing times were negligible. In the worst case, Fig. 7: measured values (blue) vs. desired trajectory (red)
it takes 121 ms + 96/2 ms = 169 ms and in the mean 71 ms
from a change in position to a response at the quadcopter.
TABLE II: Evaluation of absolute position accuracy
TABLE I: Processing times and delays alignment ∆x [mm] ∆y [mm] ∆yaw [◦ ]
t [ms] std(t)[ms] tmax [ms] parallel, D = 2 m 43 102 1.1
load image from driver 0.5 0.4 1.9 parallel, D = 4 m 98 137 2.8
LED-recogn., all LEDs 0.7 0.2 2.1 converging, D = 2 m 18 87 1.1
LED-recogn., three LEDs 5.8 0.6 7.3 converging, D = 4 m 42 120 2.0
LED-recogn., no LED 20.1 1.2 28.4
3D-reconstruction < 0.1 - - To measure how accurately small changes could be recog-
position- and orientation < 0.1 - - nized by the system, the position was increased in only one
communication, roundtrip 55.1 18.8 96 axis by steps of 50 mm. This was repeated for each of the
camera delay 86.9 20.6 121 x- and y-axis and the camera alignments listed above. This
time, the mean value δ of absolute deviations to the desired
The Software was executed on a laptop with a 2.4 GHz step size of 50 mm was calculated over all steps, shown in
Core2Duo processor and 3 GB RAM. CPU load with only Table III.
3 recognized LEDs at 15 FPS was about 25-30 % includ-
TABLE III: Evaluation of relative position accuracy
ing 15 % needed by the cameradrivers to capture images.
Memory usage was about 50 MB. alignment δ x [mm] δ x [%] δ y [mm] δ y [%]
parallel, D = 2 m 4.3 8.7 1.8 3.5
B. Pose accuracy parallel, D = 4 m 26.0 52.1 3.7 7.3
converging, D = 2 m 4.2 8.3 0.9 1.7
Absolute as well as relative accuracy was measured by converging, D = 4 m 6.6 13.1 1.5 3.1
mounting the quadcopter on a programmable 2D-positioning
table with a traverse path of 600 mm in x- and 1200 mm in The convergent camera-alignment should be clearly pre-
y- direction.The quadcopter could also be rotated around its ferred especially regarding the results of relative accuracy
z-axis. Trajectories were programmed and taken as reference. for larger distances. Overall, the position resolution is suf-
The experiments were repeated with the cameras’ optical axis ficient to stabilize the quadcopter on a desired trajectory. A
aligned parallel as well as converging at certain distances D resonable flight volume ranges from 1.5 m distance from the
to the positioning table. To measure the absolute accuracy, a cameras to 4 m and up to 1.5 in height.
rectangular trajectory that used the whole traverse paths was
programmed. V. POSITION- AND HEADINGCONTROLLER
Fig. 7 shows the measured values (blue) by the tracking We show an exemplary design of a position controller
system compared to the desired trajectory (red). The mean to demonstrate the capabilities of the overall system. The
controllers are implemented on the high-level processor of
the IMU (II-A) using a Matlab/Simulink SDK available for
it.

A. Basic controller design and settings

Four DOF (x, y, z, yaw) must be controlled with the help of
an absolute reference to achieve stable hovering. The IMU’s
high-level processor receives filtered and fused sensor data
from the low level processor and can send control commands
both at a rate of 1 kHz. All four controllers are implemented
independently as PID (x, y, z) or PI (yaw) controllers. The
interface of the low-level processor offers the following input
signals: Fig. 8: system structure
Roll: roll-angle ⇒ acceleration in y-direction
Pitch: pitch-angle ⇒ acceleration in x-direction
Thrust:momentum of motors ⇒ acceleration in z-direction
Yaw: angular velocity around z-axis The system is now able to respond quickly based on the
fast onboard velocity estimates whereas drift resulting from
As no dynamic or aerobatic maneuvers should be per-
integrating noisy signals is compensated by the much slower,
formed, the control inputs were assumed to be linear to
but absolute vision signal. Other filtering techniques like the
the resulting accelerations when the quadcopter is hovering
family of Kalman filters to combine vision and inertial data
(small angle approximation). The controller’s differential
would also be suitable and could further improve results, but
components are computed from vision- data since we have
are not implemented yet.
no sensors to directly measure the speed. The integral com-
ponent is non-linear, i.e. errors are not summed up when Important is the choice of the reference frame in which
the quadcopter already moves towards its setpoint, so that the quadcopter is controlled. We use the world frame as
overshooting is reduced. reference but omitting yaw, denoted by a subscript 0̄. This
For the yaw controller, only a P-term is needed to com- is possible, since yaw is already controlled by the low-
pensate for drift because the angular velocity was already level processor on the IMU that uses an estimated heading
controlled by the IMU. A small I-term was added to avoid from the yaw gyroscope. With this approach, control outputs
constant errors. The P-terms of the x-, y- and z- controllers gained from faster IMU data can be mapped directly into
had to be chosen small until oscillations occurred due to the appropriate motor commands. In contrast, control in the
reduced bandwidth of the controller caused by delays induced “pure” world reference frame (subscript 0) would require
by the cameras and data transmission. For the height or z- IMU data available at much higher rates to be transformed
controller, an I-term is needed to keep the height constant at w.r.t. the heading (yaw) determined by the vision system.
changing battery voltages or different payloads. Control outputs in the x- and y- axis would then rely on
the heading provided from vision that might be up to 50-
B. Improvements by using onboard inertial sensors 100 IMU-timesteps old. Our controllers can now damp the
We were able to achieve stable hovering with the quad- vehicle’s dynamics and can quickly respond to disturbances.
copter with the controllers described above. Inertial measure- In the slower vision loop, the absolute position errors gained
ments are provided by the IMU’s low level processor at a from slower vision and setpoint commands are then rotated
rate of 1 kHz (section II-A). We show a simple approach to into the quadcopter frame to compensate for the remaining
improve the controller’s performance by incorporating this drift.
additional information. In order to achieve better damping The final resulting controller structure can be seen in
of the fast dynamics of the vehicle and fast responses to Fig. 8. The acceleration vector ẍB is measured w.r.t the
disturbances, especially the speed-signal has to be improved. quadcopter’s current attitude. Before integrating accelera-
Therefore, we used the acceleration measurements as well tions, this vector has first to be rotated from body to 0̄ -
as the differential terms from vision data as inputs for a coordinates by R0̄B . This rotation matrix is determined from
complementary filter. The filter and controller is implemented the pitch- and roll-angles estimated by the IMU. R0̄0 rotates
on the onboard high-level processor running at a rate of 1 the quadcopter’s pose w.r.t. the world reference frame by
kHz. This avoids delays induced by wireless transmission to the yaw-angle to 0̄. The “inversion” block transforms the
a ground station and back. Just pose information and setpoint acceleration commands into appropriate control commands
commands have to be transmitted to the quadcopter at much for the quadcopter. In the current implementation, it is a linear
slower rates from the ground station evaluating vision data. block for small angle approximation.
C. Results
In the following graphs, the behavior of the system in
certain situations is shown. Fig. 9 shows stationary hovering.
Only slow oscillation with an amplitude of ≈ 15 cm occur.
Fig. 10 shows how disturbances at t = 12.5 s and t = 3.5 s in
the x- and y-axis were handled. The large disturbance of the
y-axis for example could be compensated within 5 s. Fig. 11
shows the behavior when following a trajectory. The graph
of the yaw-angle is a lot more smother because the angle
was controlled by the IMU at 1 kHz, so that our controller
only needs to compensate for drifts (see also II-A). A video
of the results can be seen at the author’s website3 .
Fig. 11: following trajectories for x and yaw
x, y, z vs . time
0.6
pos ition [m]

0.4 x y z
0.2 ACKNOWLEDGMENT
0
-0.2 This work was supported in part within the DFG excellence
0 5 10 15 20 25 30 initiative research cluster Cognition for Technical Systems –
time [s ] CoTeSys, see also www.cotesys.org and the Bernstein
Center for Computational Neuroscience Munich, see also
Fig. 9: stationary hovering
www.bccn-munich.de. The authors also gratefully ac-
knowledge Daniel Gurdan, Jan Stumpf, Michael Achtelik and
Klaus-Michael Doth from Ascending Technologies GmbH
for their technical support with the quadcopter.
R EFERENCES
[1] P. Pounds, R. Mahony, J. Gresham, P. Corke, and J. Roberts, “To-
wards dynamically-favourable quad-rotor aerial robots,” in Australian
Conference on Robotics and Automation,, 2004.
[2] D. Gurdan, J. Stumpf, M. Achtelik, K.-M. Doth, G. Hirzinger, and
D. Rus, “Energy-efficient autonomous four-rotor flying robot controlled
at 1 khz,” in IEEE International Conference on Robotics and Automa-
tion, Roma, Italy, Apr. 2007, pp. 361 – 366.
Fig. 10: disturbances of x- and y-position [3] C. Lu, G. D. Hager, and E. Mjolsness, “Fast and globally convergent
pose estimation from video images,” IEEE Transactions on Pattern
Analysis and Machine Intelligence, vol. 22, pp. 610–622, 2000.
[4] R. Laganiere, S. Gilbert, and G. Roth, “Robust object pose estimation
from feature-based stereo,” IEEE Transactions on Instrumentation and
VI. CONCLUSIONS Measurement. Volume 55, Number 4. August 2006., vol. 55, pp. 1270–
1280, 2006.
We presented an effective and reliable method for tracking [5] E. Altug, J. P. Ostrowski, and C. J. Taylor, “Control of a quadrotor he-
licopter using dual cameravisual feedback,” The International Journal
a quadcopter or general objects with active markers. The of Robotics Research, vol. 24, pp. 329–341, 2005.
tracking system is highly transportable, easy to set up and [6] S. Bouabdallah, “Design and control of quadrotors with application to
the can be executed on a laptop. We also showed that autonomous flying,” Ph.D. dissertation, Ecole polytechnique fdrale de
Lausanne (EPFL), 2007.
negative influence of the time delays on the performance [7] R. He, S. Prentice, and N. Roy, “Planning in information space for
of the controllers could be compensated by evaluating the a quadrotor helicopter in a gps-denied environments,” in Proc. ICRA,
onboard acceleration sensors of the quadcopter. In future, Los Angeles, CA, 2008, pp. 1814–1820.
[8] Y. Ma, S. Soatto, J. Kosecká, and S. S. Sastry, An Invitation to 3-D
we plan to apply more than two cameras to gain a larger Vision. Springer Verlag New York, 2005.
flight volume. Position control and sensor data fusion could [9] S. Umeyama, “Least-squares estimation of transformation parameters
be further enhanced by using techniques like the family of between two point patterns,” in IEEE Transactions on Pattern Analysis
and Machine Intelligence, vol. 13, Apr. 1991, pp. 376–380.
Kalman filters. [10] G. Hoffmann, H. Huang, S. L. Waslander, and C. J. Tomlin, “Quadrotor
helicopter flight dynamics and control: Theory and experiment,” in Pro-
3 http://www.lsr.ei.tum.de/research/videos/ ceedings of the AIAA Guidance, Navigation, and Control Conference,
robotics/qc-tracking-high 2007.

View publication stats