Autonomous Landing of a
Multirotor Micro Air Vehicle
on a High Velocity Ground Vehicle ⋆
Alexandre Borowczyk ∗Duc-Tien Nguyen ∗
Andr´e Phu-Van Nguyen ∗Dang Quang Nguyen ∗
David Saussi´e ∗Jerome Le Ny ∗
∗Mobile Robotics and Autonomous Systems Laboratory, Polytechnique
Montreal and GERAD, Montreal, Canada (e-mail:
{alexandre.borowczyk, duc-tien.nguyen, andre-phu-van.nguyen,
dang-quang.nguyen, d.saussie, jerome.le-ny}@polymtl.ca).
Abstract: While autonomous multirotor micro aerial vehicles (MAVs) are uniquely well suited
for certain types of missions benefiting from stationary flight capabilities, their more widespread
usage still faces many hurdles, due in particular to their limited range and the difficulty of fully
automating the deployment and retrieval. In this paper we address these issues by solving the
problem of the automated landing of a quadcopter on a ground vehicle moving at relatively high
speed. We present our system architecture, including the structure of our Kalman filter for the
estimation of the relative position and velocity between the quadcopter and the landing pad,
as well as our controller design for the full rendezvous and landing maneuvers. The system is
experimentally validated by successfully landing in multiple trials a commercial quadcopter on
the roof of a car moving at speeds of up to 50 km/h.
Keywords: Kalman filters, Autonomous vehicles, Mobile robots, Guidance systems, Computer
vision, Aerospace control
1. INTRODUCTION
The ability of multirotor micro aerial vehicles (MAVs) to
perform stationary hover flight makes them particularly
interesting for a wide variety of applications, e.g., site
surveillance, parcel delivery, or search and rescue opera-
tions. At the same time however, they are challenging to
use on their own because of their relatively short battery
life and short range. Deploying and recovering MAVs from
mobile Ground Vehicles (GVs) could alleviate this issue
and allow more efficient deployment and recovery in the
field. For example, delivery trucks, public buses or marine
carriers could be used to transport MAVs between loca-
tions of interest and allow them to recharge periodically
(Garone et al., 2014; Mathew et al., 2015). For search and
rescue operations, the synergy between ground and air
vehicles could help save precious mission time and would
pave the way for the efficient deployment of large fleets of
autonomous MAVs.
The idea of better integrating GVs and MAVs has indeed
already attracted the attention of multiple car and MAV
manufacturers (Kolodny, 2016; Lardinois, 2016). Research
groups have previously considered the problem of landing
a MAV on a mobile platform, but most of the existing
work is concerned with landing on a marine platform or
precision landing on a static or slowly moving ground
target. Lange et al. (2009) provide an early example,
⋆This work was partially supported by CFI JELF Award 32848 and
a hardware donation from DJI.
where a custom visual marker made of concentric rings was
created to allow for relative pose estimation, and control
was performed using optical flow and velocity commands.
More recently, Yang et al. (2015) used the ArUco library
from Garrido-Jurado et al. (2014) as a visual fiducial and
IMU measurements fused in a Square Root Unscented
Kalman Filter for relative pose estimation. The system
however still relies on optical flow for accurate velocity
estimation. This becomes problematic as soon as the MAV
aligns itself with a moving ground platform, at which point
the optical flow camera suddenly measures the velocity of
the platform relative to the MAV instead of the velocity
of the MAV in the ground frame.
Muskardin et al. (2016) developed a system to land a
fixed wing MAV on top of a moving GV. However, their
approach requires that the GV cooperates with the MAV
during the landing maneuver and makes use of expensive
RTK-GPS units. Kim et al. (2014) show that it is pos-
sible to land on a moving target using simple color blob
detection and a non-linear Kalman filter, but test their
solution only for speeds of less than 1 m/s. Most notably,
Ling (2014) shows that it is possible to use low cost sensors
combined with an AprilTag fiducial marker (Olson, 2011)
to land on a small ground robot. He further demonstrates
different methods to help accelerating the AprilTag detec-
tion. He notes in particular that as a quadcopter pitches
forward to follow the ground platform, the bottom facing
camera frequently loses track of the visual target, which
stresses the importance of a model-based estimator such
as a Kalman filter to compensate.
arXiv:1611.07329v1  [cs.RO]  22 Nov 2016
The references above address the terminal landing phase
of the MAV on a moving platform, but a complete system
must also include a strategy to guide the MAV towards the
GV during its approach phase. Proportional Navigation
(PN) (Kabamba and Girard, 2014) is most commonly
known as a guidance law for ballistic missiles, but can
also been used for UAV guidance. Holt and Beard (2010)
develop a form of PN specifically for road following by
a fixed-wing vehicle and show that it is suitable for use
with visual feedback coming from a gimbaled camera.
Gautam et al. (2015) compare pure pursuit, line-of-sight
and PN guidance laws to show that out of the three,
PN is the most efficient in terms of the total required
acceleration and the time required to reach the target.
On the other hand, within close range of the target PN
becomes inefficient. To alleviate this problem, Tan and
Kumar (2014) propose a switching strategy to move from
PN to a PD controller. Finally, to maximize the likelihood
of visual target acquisition for a smooth transition from
PN to PD, it is possible to follow the strategy of Lin and
Yang (2014) to point a gimbaled camera towards a target.
Contributions and organization of the paper. We describe
a complete system allowing a multirotor MAV to land au-
tonomously on a moving ground platform at relatively high
speed, using only commercially available and relatively
low-cost sensors. The system architecture is described in
Section 2. Our algorithms combine a Kalman filter for
relative position and velocity estimation, described in Sec-
tion 3, with a PN-based guidance law for the approach
phase and a PID controller for the terminal phase. Both
controllers are implemented using only acceleration and
attitude controls, as discussed in Section 4. Our design
was tested both in simulations and through extensive ex-
periments with a commercially available MAV, as Section
5 illustrates. To the best of our knowledge, we demonstrate
experimentally automatic landing of a multirotor MAV on
a moving GV traveling at the highest speed to date, with
successful tests carried up to a speed of 50 km/h.
2. SYSTEM ARCHITECTURE
In this section we describe the basic elements of our system
architecture, both for the GV and the MAV. Specific
details for the hardware used in our experiments are given
in Section 5.
The GV is equipped with a landing pad, on which we place
a 30 × 30 cm visual fiducial named AprilTag designed by
Olson (2011), see Fig. 4. This allows us to visually measure
the 6 Degrees of Freedom (DOF) pose of the landing pad
using onboard cameras. In addition, we use position and
acceleration measurements for the GV. In practice, low
quality sensors are enough for this purpose, and in our
experiments we simply place a mobile phone on the landing
pad, which can transmit its GPS data to the MAV at 1 Hz
and its Inertial Measurement Unit (IMU) data at 25 Hz at
most. We can also integrate the rough heading and velocity
estimates typically returned by basic GPS units, based
simply on successive position measurements.
The MAV is equipped with an Inertial Navigation System
(INS), an orientable 3-axis gimbaled camera (with separate
IMU) for target tracking purposes, as well as a camera with
a wide angle lens pointing down, which allows us to keep
track of the AprilTag even at close range during the very
last instants of the landing maneuver. The approach phase
can also benefit from having an additional velocity sensor
on board. Many commercial MAVs are equipped with
velocity sensors relying on optical flow methods, which
visually estimate velocity by computing the movement of
features in successive images, see, e.g., (Zhou et al., 2015).
Four main coordinate frames are defined and illustrated
in Fig. 1. The global North-East-Down (NED) frame,
denoted {N}, is located at the first point detected by the
MAV. The MAV body frame {B} is chosen according to
the cross “×” configuration, i.e., its forward x-axis points
between two of the arms and its y-axis points to the
right. The gimbaled camera frame {G} is attached to the
lens center of the moving camera. Its forward z-axis is
perpendicular to the image plane and its x-axis points to
the right of the gimbal frame. Finally, the bottom facing
rigid camera frame {C} is obtained from the MAV body
frame by a 90◦rotation around the zB axis and its origin
is located at the optical center of camera.
{N}
xN
yN
zN
zB
Horizontal plane
zG
xG
yG
AprilTag
{G}
Gimbaled camera
{C}
zR
Rigid Camera
xR
xB
Global NED
{B}
M1
M2
M3
M4
yB
Image plane
Quadrotor
yR
Fig. 1. Frames of reference used
3. KALMAN FILTER
Estimation of the relative position, velocity and accelera-
tion between the MAV and the landing pad, as required
by our guidance and control system, is performed by a
Kalman filter running at 100 Hz. The architecture of this
filter is shown in Fig. 2 and described in the following
paragraphs.
3.1 Process model
In order to land on a moving target, the system estimates
the three dimensional position p(t), linear velocity v(t) and
Kalman filter
GPSa
Transformation
Gimbal 
Camera
Gimbal 
IMU
Transformation
Bottom 
Camera
MAV  
INS
Mobile 
phone IMU
Transformation
Mobile 
phone GPS
Prediction: 100Hz
𝑝𝑎𝐶 
𝑅𝐶
𝑁 
𝑝𝑎𝑁−𝑝𝑚𝑁 
𝑝𝑎𝑁−𝑝𝑚𝑁 
𝑝𝑎𝑁, 𝑣𝑎𝑁 
𝑅𝐺
𝑁 
𝑝𝑎𝐺 
𝑎𝑎𝑁 
𝑝 𝑚𝑁, 𝑣 𝑚𝑁, 𝑎 𝑚
𝑁 
𝑝 𝑎𝑁, 𝑣 𝑎𝑁, 𝑎 𝑎𝑁 
20Hz
30Hz
1Hz
25Hz
𝑝𝑚
𝑁, 𝑣𝑚
𝑁, 𝑎𝑚
𝑁 100Hz
Fig. 2. Kalman filter architecture
acceleration a(t) of the MAV and of the GV. Thus, the
state variables can be written as
x = [x⊤
m x⊤
a ]⊤∈R18
(1)
where xm =

p⊤
m v⊤
m a⊤
m
⊤and xa =

p⊤
a
v⊤
a
a⊤
a
⊤are
respectively the state vectors for the MAV and AprilTag,
expressed in the NED frame. The superscript ⊤denotes
the matrix transpose operation.
We use a simple second-order kinematic model for the
MAV and the GV dynamics
a(t) = ¨p(t) = w(t)
(2)
where w(t) is a white noise process with power spectral
density (PSD) qw. The corresponding discrete time model
using zero-order hold (ZOH) sampling is given by
xk+1 = Fxk + wk
(3)
F =
"
Fm 0
0
Fa
#
, Fm = Fa =


1 Ts
T 2
s
2
0 1
Ts
0 0
1

⊗I3
where Ts is the sampling period, ⊗denotes the Kronecker
product and I3 is the 3 × 3 identity matrix. The process
noise wk is assumed to be a Gaussian white noise with a
covariance matrix Q given by
Q =

qwm
0
0
qwa

⊗Q0, Q0 =


T 5
s
20
T 4
s
8
T 3
s
6
T 4
s
8
T 3
s
3
T 2
s
2
T 3
s
6
T 2
s
2
Ts


⊗I3
where qwm and qwa are the PSD of the MAV and GV
acceleration, respectively. These parameters are generally
chosen a priori as part of an empirical tuning process.
3.2 Measurement Model
In general, the measurement vector at time k is given by
zk = Hxk + vk,
(4)
where H is a known matrix, vk is a Gaussian white noise
with covariance matrices Vk and not correlated with the
process noise wk. The following subsections describe the
rows of the matrix H for the various kinds of sensor
measurements, which we simply call H for simplicity of
notation.
MAV position, velocity and acceleration from INS
The
INS of our MAV combines IMU, GPS and visual measure-
ments to provide us with position, velocity and gravity
compensated acceleration data directly expressed in the
global NED frame
zk =

p⊤
m v⊤
m a⊤
m
⊤, H = [I9 09×9]
(5)
As mentioned previously, the velocity measurement relying
on optical flow methods is not correct when the MAV
flies above a moving platform. Therefore, we increase the
standard deviation of the velocity noise from 0.1 m/s in
the approach phase to 10 m/s in the landing phase.
Target’s GPS measurements
The GPS unit of the mobile
phone on the GV provides position, speed and heading
measurements. This information is sent to the MAV on-
board computer (OBC) via a wireless link, which gives
access to the landing pad’s position in the global NED
frame {N}
xN
a ≈(la −la0)RE , yN
a ≈(lo −lo0) cos(la)RE
zN
a = al0 −al,
where RE = 6378137 m is the Earth radius, la, lo, al are
the latitude, longitude (both in radians) and altitude of
the landing pad respectively. The subscripts 0 corresponds
to the starting point. The above equations are valid when
the current position is not too far from the starting point
and under a spherical Earth assumption, but more precise
transformations could be used (Groves, 2013).
The GPS heading ψa and speed Ua are also used to
calculate the current velocity in the global NED frame
˙xN
a = Ua cos(ψa),
˙yN
a = Ua sin(ψa).
So the measurement model (4) is expressed as follows
zk =

xN
a yN
a zN
a ˙xN
a ˙yN
a
⊤, H = [05×9 I5 05×4]
However, because the GPS heading measurements have
poor accuracy at low speed, we discard them if Ua <
2.5 m/s, in which case
zk =

xN
a yN
a zN
a
⊤, H = [05×9 I3 05×6]
The measurement noise covariance matrix Vk is provided
by the GPS device itself.
Our GPS receiver is a low-cost device with output rate of
about 1 Hz. This source of data is only used to approach
the target but is insufficient for landing on the moving GV.
For the landing phase, it is necessary to use the AprilTag
detection with the gimbaled and bottom facing cameras.
Gimbaled Camera measurements
This camera provides
measurements of the relative position between the MAV
and the landing pad with centimeter accuracy, at range up
to 5 m. This information is converted into the global NED
frame {N} by
pN
m −pN
a = RN
GpG
m/a
where RN
G is the rotation matrix from {N} to {G} returned
by the gimbal IMU. Therefore, the observation model (4)
corresponds to
zk =

xN
m −xN
a
yN
m −yN
a
zN
m −zN
a
⊤
H = [I3 03×6 −I3 03×6] .
Here the standard deviation of the measurement noise is
empirically set to 0.2 m. To reduce the chances of target
loss, the gimbaled camera centers the image onto the
AprilTag as soon as visual detection is achieved. When the
AprilTag cannot be detected, we follow the control scheme
proposed by Lin and Yang (2014) to point the camera
towards the landing pad using the estimated line-of-sight
(LOS) information obtained from the Kalman filter.
Bottom camera measurements
The bottom facing cam-
era is used to assist the last moments of the landing, when
the MAV is close to the landing pad, yet too far to cut
offthe motors. At that moment, the gimbaled camera can
not perceive the whole AprilTag but a wide angle camera
like the mvBlueFOX can still provide measurements.
This camera measures the target position in the frame
{C}, as illustrated in Fig. 1. Hence, the observation model
is the same as (3.2.3), except for the transformation to the
global NED frame
pN
m −pN
a = RN
C pC
m/a
where RN
C denotes the rotation matrix from {N} to {C}.
Landing pad acceleration using the mobile phone’s IMU
Finally, since most mobile phones also contain an IMU, we
leverage this sensor to estimate the GV’s acceleration.
zk =

¨xN
a ¨yN
a ¨zN
a
⊤, H = [03×15 I3]
Vk = diag(0.62, 0.62, 0.62) m/s2.
The Kalman filter algorithm follows the standard two
steps, with the prediction step running at 100 Hz and
the measurement update step executed as soon as new
measurements become available. The output of this filter
is the input to the guidance and control system described
in the next section, which is used by the MAV to approach
and land safely on the moving platform.
4. GUIDANCE AND CONTROL SYSTEM
For GV tracking by the MAV, we use a guidance strategy
switching between a Proportional Navigation (PN) law
(Kabamba and Girard, 2014) for the approach phase and
a PID for the landing phase, which is similar in spirit to
the approach in (Tan and Kumar, 2014).
The approach phase is characterized by a large distance
between the MAV and the GV and the absence of visual
localization data. Hence, in this phase, the MAV has
to rely on the data transmitted by the GV’s GPS and
IMU. The goal of the controller for this phase is to follow
an efficient “pursuit” trajectory, which is achieved here
by a PN controller augmented with a closing velocity
controller. In contrast, the landing phase is characterized
by a relatively close proximity between the MAV and the
GV and the availability of visual feedback to determine
the target’s position. This phase requires a higher level
of accuracy and faster response time from the controller,
and a PID controller can be more easily tuned to meet
these requirements than a PN controller. In addition, the
system should transition from one controller to the other
seamlessly, avoiding discontinuity in the commands sent
to the MAV.
Proportional Navigation Guidance The well-known PN
guidance law uses the fact that two vehicles are on a
collision course if their LOS remains at a constant angle
in order to steer the MAV toward the GV. It works by
keeping rotation of the velocity vector proportional to the
rotation of the LOS vector. Our PN controller provides an
acceleration command that is normal to the instantaneous
LOS
a⊥= −λ| ˙u| u
|u| × Ω, with Ω= u × ˙u
u · u ,
(6)
where λ is a gain parameter, u = pN
a −pN
m and ˙u = vN
a −
vN
m are obtained from the Kalman filter and represent
the LOS vector and its derivative expressed in the NED
frame, and Ωis the rotation vector of the LOS. We then
supplement the PN guidance law (6) with an approach
velocity controller determining the acceleration a∥along
the LOS direction, which in particular allows us to specify
a high enough velocity required to properly overtake the
target. This acceleration component is computed using the
following PD structure
a∥= Kp∥u + Kd∥˙u
where Kp∥and Ki∥are constant gains. The total accelera-
tion command is obtained by combining both components
a = a⊥+ a∥.
As only the horizontal control is of interest, the accelera-
tion along z-axis is disregarded. The desired acceleration
then needs to be converted to attitude control inputs that
are more compatible with the MAV input format. In frame
{N}, the quadrotor dynamic equations of translation read
as follows
mam =
" 0
0
mg
#
+ RN
B
" 0
0
−T
#
+ FD
where RN
B denotes the rotation matrix from {N} to {B},
RN
B =
"cθcψ sφsθcψ −cφsψ cφsθcψ + sφsψ
cθsψ sφsθsψ + cφcψ cφsθsψ −sφcψ
−sθ
sφcθ
cφcθ
#
,
T the total thrust created by rotors, FD the drag force,
m the MAV mass and g the standard gravity. The force
equations simplify as
m
"¨xm
¨ym
¨zm
#
=
" 0
0
mg
#
−
"cφsθcψ + sφsψ
cφsθsψ −sφcψ
cφcθ
#
T +
"−kd ˙xm| ˙xm|
−kd ˙ym| ˙ym|
−kd ˙zm| ˙zm|
#
where the drag is roughly modeled as a force proportional
to the signed quadratic velocity in each direction and kd
is constant, which we estimated by recording the terminal
velocity for a range of attitude controls at level flight and
performing a least squares regression on the data. For
constant flight altitude, T = mg/cφcθ and assuming ψ = 0,
it yields
m

¨xm
¨ym

= mg

−tan θ
tan φ/ cos θ

−

kd ˙xm| ˙xm|
kd ˙ym| ˙ym|

The following relations can then be obtained:
θ = −arctan ((m¨xm + kd ˙xm| ˙xm|)/mg)
φ = arctan (cos θ (m¨ym + kd ˙ym| ˙ym|) /mg) ,
where θ and φ are the desired pitch and roll angles for
specific acceleration demands.
PID controller The landing phase is handled by a PID
controller, with the desired acceleration computed as
a = Kpu + Ki
Z
u + Kd ˙u,
where Kp, Ki and Kd are constant gains. The tuning
for the PID controller was selected to provide aggressive
dynamic path following, promoting a quick disturbance
rejection. The controller was first tuned in simulation and
then the settings were manually adjusted in flight.
Controller switching The controller switching scheme cho-
sen is a simple fixed distance switching condition with
a slight hysteresis. The switching distance selected was
6 m to allow for any perturbation due to the switching to
dissipate before reaching the landing platform.
Vertical control The entire approach phase is done at
a constant altitude, which is handled by the internal
vertical position controller of the MAV. The descent is
initiated once the quadrotor has stabilized over the landing
platform. A constant vertical velocity command is then
issued to the MAV and maintained until it reaches a height
of 0.2 m above the landing platform at which point the
motors are disarmed.
5. EXPERIMENTAL VALIDATION
5.1 System Description
We implemented our system on a commercial off-the-shelf
DJI Matrice 100 (M100) quadcopter shown in Fig. 3. All
computations are performed on the standard OBC for this
platform (DJI Manifold), which contains an Nvidia Tegra
K1 SoC. The 3-axis gimbaled camera is a Zenmuse X3,
from which we receive 720p YUV color images at 30 Hz. To
reduce computations, we drop the U and V channels and
downsample the images to obtain 640 × 360 monochrome
images. We modified the M100 to rigidly attach a down-
ward facing Matrix Vision mvBlueFOX camera, equipped
with an ultra-wide angle Sunex DSL224D lens with a
diagonal field of view of 176 degrees. The M100 is also
equipped with the DJI Guidance module, an array of up
to five stereo cameras, which are meant to help develop-
ers create mapping and obstacle avoidance solutions for
robotics applications. This module seamlessly integrates
with the INS to provide us with position, velocity and
acceleration measurements for the M100 using a fusion
of on-board sensors described in Zhou et al. (2015). This
information is used in our Kalman filter in equation (5).
Our algorithms were implemented in C++ using ROS
(Quigley et al., 2009). Using an open source implemen-
tation of the AprilTag library based on OpenCV and with
accelerations provided by OpenCV4Tegra, we can run the
tag detection at a full 30 fps using the X3 camera and 20
fps on the BlueFOX camera.
As mentioned in Section 4, we implemented our control
system using pure attitude control in the xy axes and
velocity control in the z axis. The reason for this is that the
internal velocity estimator of the M100 fuses optical flow
measurements from the Guidance system. These measure-
ments become extremely inaccurate once the quadcopter
flies over the landing platform. Although optical flow could
be used to measure the relative velocity of the car, it
would be difficult to pinpoint the moment where flow
measurements transition from being with respect to the
ground to being with respect to the moving car.
Fig. 3. The M100 quadcopter. Note that all side facing
cameras of the Guidance module were removed and
the down facing BlueFox camera sits behind the
bottom guidance sensor.
5.2 Experimental Results
Fig. 4. Experimental setup showing the required equip-
ment on the car. In practice the mobile phone could
also be held inside the car as long as GPS reception
is still good.
A first experiment was done using only data from the
mobile phone (no visual feedback) to prove the validity
of our Proportional Navigation controller for the long
range approach. Figure 5 shows the output of our EKF
estimating the MAV’s and the target’s positions. With the
target and the MAV starting at about 30 meters from each
other, we see the MAV fly a smooth rendezvous trajectory
eventually ending on top of the target.
The close range system was experimentally validated with
the GV moving at speeds as low as a human jogging and on
a private race track at 30, 40 and 50 km/h with successful
landings in each case. Videos of our experiments can be
found at https://youtu.be/ILQqD2xQ4tg.
Figure 6 shows the landing sequence where the quadrotor
takes off, tracks and lands on the landing pad. The curves
gain in altitude as the trajectory progresses because of the
elevation profile of the race track. The effect is seen more
clearly in Fig. 7, where we can also see the filtered AprilTag
altitude rise, thanks to visual data and the M100’s internal
altitude estimator, even before the phone’s GPS data
indicates a change in altitude.
Furthermore, we can see in Fig. 7 how the M100 closely
matches the velocity of the AprilTag to perform the
landing maneuver. The two peaks at the 24 and 27 second
marks are strongly correlated to the visual loss of the
tag by the BlueFOX camera which we can observe in
Fig. 8. The descent starts at the 24 second mark slightly
before the car hits its designated velocity of 14 m/s or
50.4 km/h. We can also see in Fig. 7 how the M100’s
velocity estimation from the INS becomes incorrect when
it is on top of the car. Which explains why we decided
to dynamically increase the standard deviation of the
measurement as described in Section 3.2. Figure 9 shows
the quadcopter’s attitude during the flight. Notice how
the roll is stable close to 0◦while the pitch stays between
−10◦and −25◦for consistent forward flight. Finally, the
yaw changes drastically after 10 seconds at the moment
the car starts moving.
40
20
South-North (meters)
0
0
40
-20
30
West-East (meters)
20
10
-40
0
Proportional navigation trajectory in ENU
5
Up (meters)
10
MAV
target
Fig. 5. The PN controller efficiently catches up with the
target at long range even when the only source of
information is the mobile phone’s GPS and IMU.
M100
AprilTag
0
20
2
200
Up (meters)
0
4
150
3D Landing trajectory in ENU
West-East (meters)
South-North (meters)
100
6
-20
50
-40
0
Platform starts moving
Descent
Motors Cutoff
Fig. 6. Landing trajectory at 50km/h using the PID
controller.
6. CONCLUSION
The problem of the automatic landing of a MAV on a
moving vehicle was solved with experimental tests going
up to 50 km/h. A Proportional Navigation controller was
0
5
10
15
20
25
30
time (s)
0
5
10
15
speed (m/s)
Speed
M100_INS
M100_KF
AprilTag_KF
0
5
10
15
20
25
30
time (s)
0
1
2
3
4
5
6
altitude (m)
Altitude
M100_INS
M100_KF
AprilTag_GPS
AprilTag_KF
Fig. 7. Motions of M100 and AprilTag.
0
10
20
30
time (s)
-1
0
1
2
x (m)
Estimated LOS X
0
10
20
30
time (s)
-1
-0.5
0
0.5
y (m)
Estimated LOS Y
0
10
20
30
time (s)
0
2
4
6
z (m)
Estimated LOS Z
0
10
20
30
time (s)
0
2
4
6
Distance (m)
Distance
KF
X3
Bluefox
Fig. 8. LOS and distance (coordinates in NED frame).
0
10
20
30
time (s)
-40
-20
0
20
degree
M100 roll
0
10
20
30
time (s)
-100
-50
0
50
degree
M100 pitch
0
10
20
30
time (s)
-100
-50
0
50
degree
M100 yaw
Fig. 9. M100 attitude.
used for the long range approach, which subsequently tran-
sitioned into a PID controller at close range. A Kalman
filter was used to estimate the position of the MAV relative
to the landing pad by fusing together measurements from
the MAV’s onboard INS, a visual fiducial marker and a
mobile phone. Furthermore, it was shown that this system
can be implemented using only commercial off-the-shelf
components.
Future work may include using a better multiscale visual
fiducial on the landing pad to allow visual target tracking
at longer and closer ranges using a single camera or simpli-
fying the system by removing the requirement for a mobile
phone. Performance improvements could also be achieved
by adding a model of the ground vehicle’s turbulence or
adding wind estimation in the control system.
REFERENCES
Garone, E., Determe, J.F., and Naldi, R. (2014).
Gen-
eralized traveling salesman problem for carrier-vehicle
systems. AIAA Journal of Guidance, Control and Dy-
namics, 37(3), 766–774.
Garrido-Jurado, S., noz Salinas, R.M., Madrid-Cuevas, F.,
and Mar´ın-Jim´enez, M. (2014). Automatic generation
and detection of highly reliable fiducial markers under
occlusion. Pattern Recognition, 47(6), 2280 – 2292.
Gautam, A., Sujit, P.B., and Saripalli, S. (2015). Appli-
cation of guidance laws to quadrotor landing. In Un-
manned Aircraft Systems (ICUAS), 2015 International
Conference on, 372–379.
Groves, P.D. (2013).
Principles of GNSS, inertial,
and multisensor integrated navigation systems. Artech
House, 2nd edition.
Holt, R. and Beard, R. (2010). Vision-based road-following
using proportional navigation. Journal of Intelligent and
Robotic Systems, 57(1-4), 193 – 216.
Kabamba, P.T. and Girard, A.R. (2014). Fundamentals
of Aerospace Navigation and Guidance.
Cambridge
University Press.
Kim, J., Jung, Y., Lee, D., and Shim, D.H. (2014). Out-
door autonomous landing on a moving platform for
quadrotors using an omnidirectional camera.
In Un-
manned Aircraft Systems (ICUAS), 2014 International
Conference on, 1243–1252.
Kolodny, L. (2016). Mercedes-benz and Matternet unveil
vans that launch delivery drones.
http://tcrn.ch/
2c48his. (Accessed on 09/09/2016).
Lange, S., Sunderhauf, N., and Protzel, P. (2009).
A
vision based onboard approach for landing and position
control of an autonomous multirotor uav in gps-denied
environments. In Advanced Robotics, 2009. ICAR 2009.
International Conference on, 1–6.
Lardinois, F. (2016). Ford and DJI launch $100,000 devel-
oper challenge to improve drone-to-vehicle communica-
tions. tcrn.ch/1O7uOFF. (Accessed on 09/28/2016).
Lin, C.E. and Yang, S.K. (2014). Camera gimbal tracking
from uav flight control. In Automatic Control Confer-
ence (CACS), 2014 CACS International, 319–322.
Ling, K. (2014). Precision Landing of a Quadrotor UAV
on a Moving Target Using Low-cost Sensors. Master’s
thesis, University of Waterloo.
Mathew, N., Smith, S.L., and Waslander, S.L. (2015).
Planning paths for package delivery in heterogeneous
multirobot teams. IEEE Transactions on Automation
Science and Engineering, 12(4), 1298–1308.
Muskardin, T., Balmer, G., Wlach, S., Kondak, K., La-
iacker, M., and Ollero, A. (2016). Landing of a fixed-
wing uav on a mobile ground vehicle.
In 2016 IEEE
International Conference on Robotics and Automation
(ICRA), 1237–1242.
Olson, E. (2011). Apriltag: A robust and flexible visual
fiducial system. In Robotics and Automation (ICRA),
2011 IEEE International Conference on, 3400–3407.
Quigley, M., Conley, K., Gerkey, B.P., Faust, J., Foote, T.,
Leibs, J., Wheeler, R., and Ng, A.Y. (2009). ROS: an
open-source robot operating system. In ICRA Workshop
on Open Source Software.
Tan, R. and Kumar, M. (2014).
Tracking of ground
mobile targets by quadrotor unmanned aerial vehicles.
Unmanned Systems, 2(02), 157–173.
Yang, S., Ying, J., Lu, Y., and Li, Z. (2015).
Precise
quadrotor autonomous landing with SRUKF vision per-
ception.
In 2015 IEEE International Conference on
Robotics and Automation (ICRA), 2196–2201.
Zhou, G., Fang, L., Tang, K., Zhang, H., Wang, K., and
Yang, K. (2015). Guidance: A visual sensing platform
for robotic applications. In 2015 IEEE Conference on
Computer Vision and Pattern Recognition Workshops
(CVPRW), 9–14.