1. Vision Based Autonomous Landing of UAV
Shikhar Gupta1 and Dr. Hyunchul Shin2
Abstract— This paper presents our work on quad-rotor
unmanned aerial vehicle and focuses on our vision-based
method for detection of landing platform at the destination
GPS location. The paper describes the approach for auto-
maneuvering of the UAV to the landing platform using onboard
control. The approach described uses Robot Operating System
(ROS) framework to integrate the onboard components and
generate control signals for the UAV.
I. INTRODUCTION
A lot of organizations involved in delivery systems are
promoting the use of miniature UAVs, also known as drones,
for delivering packages in a very short duration. It will also
increase the overall safety and efficiency of transportation
system. Amazon is developing its own drone delivery system,
Prime Air[1]
. Google has been testing UAVs in Australia
which aims to produce drones to deliver products sold via e-
commerce[2]
. USPS has been testing delivery systems[3]
with
HorseFly Drones.
Delivery systems used are generally GPS based and use
longitude and latitude coordinates to send off the drone to the
destination. While GPS-based control is sufficient for general
positioning (assuming the signal is not blocked), the accuracy
of GPS receivers fit for use on miniature UAVs is measured
in meters, making them unsuitable for precision tasks such
as landing. For a delivery system to function successfully, it
is mandatory for the drone to land on the target precisely.
For this paper, we use the Scale Invariant Feature Transform
(SIFT) to detect the landing site and attitude control of the
drone for its precise maneuvering.
A. The Quadcopter ”DJI Matrice 100”
The quadcopter used in this paper is Matrice 100 (Fig.1)
manufactured by DJI, a Chinese technology company[4]
. It
is propelled by 4 DC motors and can carry upto 3600g
(including the 2 LiPo 6S batteries) for 20 minutes. Apart
from the propellers, it is equipped with a GPS module, a
camera with full gimbal control and the usual accelerometers
and gyros. The GPS module enables the drone to reach the
destination and hold its position but the system has intrinsic
error sources that have to be taken into account when a
receiver reads the GPS signals from the constellation of
satellites in orbit. Hence, we have used the attitude control
mode along with the GPS module for precise landing. It gives
access to the roll and pitch control.
1Shikhar Gupta is an undergraduate with major Electronics
and Communication, Indian Institute of Technology, Guwahati
g.shikhar@iitg.ernet.in
2Dr. Hyunchul Shin is with the Faculty of Electrical
Engineering, Hanyang University, ERICA Campus, South Korea.
shin@hanyang.ac.kr
Fig. 1: DJI Matrice 100 with NVIDIA Jetson TK1 mounted
on top
B. The Onboard Processor ”NVIDIA Jetson TK1”
We have equipped the drone with additional hardware.
This includes NVIDIA’s Jetson TK1[5]
which is an embedded
Linux development platform featuring a Tegra K1 SOC
(CPU+GPU+ISP in a single chip). Besides the quad-core
2.3GHz ARM Cortex-A15 CPU and the Tegra K1 GPU, the
Jetson TK1 board includes similar features as a Raspberry
Pi.
One feature becoming common in miniature UAVs is
the onboard camera which helps in performing isolated
vision based tasks and position control. We have mounted
a USB camera over the drone and is connected to the Jetson
TK1’s USB port. Images are acquired from the camera and
processed onboard.
C. Related Work
Various organizations and research groups have presented
their work on UAVs and recently it has become more
famous among researchers and hobbyists due its wide range
applications. Most similar to our work is [6] which also
focuses on detection of landing site using vision algorithm
and defines the orientation of the UAV. [7] also describes
an algorithm to detect a landing site which has a shape
of ”H”. This is one of the earliest works in this field. [8]
uses Moire patterns which are convincing not very robust in
outdoor surroundings. [9] and [10] also describe two different
approaches for identification of landing sites.
We have tried to overcome the two limitations in our
work: There has to be a specific pattern for the landing
site according to the detection algorithm for robustness and
orientation of the drone must be known for maneuvering it to
the destination. The next section describes our algorithm can
work for any generic or complex design and maneuvering is

2. independent of the orientation of the drone at any point of
time.
II. ROS AND VISION BASED LANDING
A. Identification of The Target
Our algorithm for detection of the landing site is pro-
grammed in Python and uses the OpenCV[11]
library. Taking
advantage of the computational capacity of the Jetson TK1,
algorithm runs on an image of resolution 1280x720 and
produces results in real time.
We have used SIFT[12]
feature extraction method for the
detection. It first identifies and stores the keypoints in the
input image. Fig 2 shows the detected keypoints in training
and input image. The image is convolved with Gaussian
filters at different scales,
L(x, y, kσ) = G(x, y, kσ) ∗ I(x, y) (1)
and then the difference of successive Gaussian-blurred im-
ages are taken (Fig. 2).
D(x, y, kσ) = L(x, y, kiσ) − L(x, y, kjσ) (2)
Keypoints are then taken as maxima/minima of the Dif-
ference of Gaussians (DoG) that occur at multiple scales.
Keypoints having low contrast or are poorly localized along
an edge are discarded and each of the remaining ones are
assigned a 128 elements long unique descriptor vector.
Next step is to match the keypoints obtained from input
image to the keypoints in training image (Fig 3). This feature
matching is done through a Euclidean-distance based nearest
neighbor approach. After this step we have all the keypoints
in the input image that matches to those in the training image
and will be used for distance estimation in the next section.
(a) Landing mark
(b) Input Image
Fig. 2: SIFT keypoints detected on training and testing
images
Fig. 3: Target detection using feature matching
B. Distance Estimation
Distance estimation using vision based algorithm has still
not been improved to a great extent. We have proposed here
a simple mathematical model for this purpose.
The SIFT feature matching returns all matched keypoints
across the landing site. To simplify the calculations we select
a single keypoint to represent the target. This keypoint is
selected as the arithmetic mean of all the matched keypoints.
xmean =
(x1 + x2 + ....xn)
n
(3)
ymean =
(y1 + y2 + ....yn)
n
(4)
Current position of the drone at the time of taking the
image will be at the center of the image. Therefore, for an
image of resolution 1280x720 the position of drone can be
assumed at (640,360) on the Cartesian plane and position
of the target is (xmean, ymean). We get the relative position
of the target w.r.t the drone by translation of the x-y axes.
This relative position will be in terms of pixels and hence to
get the actual distance, we have to multiply it with a scaling
factor.
xnew = (xmean − 640) ∗ (scalingfactor) (5)
ynew = (ymean − 360) ∗ (scalingfactor) (6)
Here ’*’ denotes the product. This position vector gives x
and y distance of the drone from the landing site in meters
and will be used in next section to guide the drone to the
target location.
C. Roll and Pitch Control
Matrice 100’s attitude mode gives control over drone’s
pitch and roll angle which in turn can be used for maneuver-
ing. Both these angles determine the velocity of the drone in
the forward and lateral directions. We have kept the velocities
low for better results as in attitude mode (non-GPS) drone
tend to drift away more. xnew and ynew determine the times
for which drone will move in the forward/backward and
lateral directions with their respective velocities. This guides
the drone to the landing site independent of the orientation of
the drone. The sign of xnew determines whether the target is

3. in forward or backward direction and that of ynew determine
whether it is towards left or right and maneuvers the drone
accordingly.
D. The ROS Framework
All of the above mentioned processes are integrated into
a single ROS[13]
package as different nodes and forms the
controller structure. All the nodes can communicate with
each other and guide the drone to the target location.
• The launch node setups serial communication between
the Jetson TK1 and the drone through UART port.
• The camera node acquires image and publishes it to the
other nodes.
• The vision node subscribes to the image and uses it for
detection and distance estimation.
• The drone node uses the estimated x and y coordinates
to generate control signals and maneuver the drone to
the precise location.
III. EXPERIMENTAL RESULTS AND CONCLUSIONS
A. Experiment
The proposed landing algorithm is validated by the results.
Experiments were performed at the height of 3 meters and 5
meters and the landing mark was printed on an A4 sheet. The
scaling factor in eq. (5) and (6) had to be set accordingly. In
the steady wind conditions results are quite precise with an
error of 10 cm. Whereas in windy condition the drone tends
to drift away from the target in attitude mode (non-GPS)
and cannot hold its position with high accuracy increasing
the error. Instead, if we use a bigger landing site (bigger than
A4), the error is reduced as the area for landing is increased
and takes in account the drift.
B. Conclusion
In this paper we proposed the approach for the autonomous
landing of drone using a vision based algorithm. Our ap-
proach uses only 1 frame for identification and distance
estimation and hence usb cameras, which do not have very
high data rate, can also be used. Also, limitation of using
only a particular type of pattern for detection, as used in
most of the literature, is also eradicated.
REFERENCES
[1] Amazon: Prime Drones https://www.amazon.com/b?node=8037720011
[2] Alexis C. Madrigal (28 August 2014). ”Inside Google’s Secret Drone-
Delivery Program”. The Atlantic.
[3] ”USPS Drone Delivery — CNBC” https://youtu.be/V9GXiXgaK34.
[4] DJI developers : https://developer.dji.com.
[5] NVIDIA Jetson TK1 Wiki : http://elinux.org/Jetson TK1
[6] Sven Lange, Niko Sunderhauf, Peter Protzel, ”A Vision Based On-
board Approach for Landing and Position Control of an Autonomous
Multirotor UAV in GPS-Denied Environment” in International Con-
ference on Advanced Robotics, ICAR 2009.
[7] S. Saripalli, J.F. Montgomery, and G.S. Sukhatme, Vision-Based
Autonomous Landing of an Unmanned Aerial Vehicle, in IEEE
International Conference on Robotics and Automation (ICRA),2002
[8] G.P. Tournier, M. Valenti, J.P. How, and E. Feron, Estimation and
Control of a Quadrotor Vehicle Using Monocular Vision and Moire
Patterns, in AIAA Guidance, Navigation, and Control Conference and
Exhibit, 2006
[9] P.J. Garcia-Pardo, G.S. Sukhatme, and J.F. Montgomery, Towards
vision-based safe landing for an autonomous helicopter, Robotics and
Autonomous Systems,2001
[10] S. Bosch, S. Lacroix, and F. Caballero, Autonomous Detection of Safe
Landing Areas for an UAV from Monocular Images, in IEEE/RSJ
International Conference on Intelligent Robots and Systems, Beijing,
2006
[11] ”The OpenCV Library” http://opencv.org/
[12] Lowe, David G. (1999). ”Object recognition from local scale-invariant
features” in Proceedings of the International Conference on Computer
Vision.
[13] Morgan Quigley, Brian Gerkey, Ken Conley, Josh Faust, Tully Foote,
Jeremy Leibs, Eric Berger, Rob Wheeler and Andrew Ng, ”ROS: an
open-source Robot Operating System”, in Proc. of the IEEE Intl. Conf.
on Robotics and Automation (ICRA), 2009