Hanging Drone: An Approach to UAV Landing for Monitoring

Alan Kunz Cechinel

1 a

, Juha R

oning

2 b

, Antti Tikanmaki

, Edson Roberto De Pieri

1 c

and Patricia Della M

ea Plentz

3 d

Automation and Systems Department, Federal University of Santa Catarina, Florian

opolis, Brazil

Biomimetics and Intelligent Systems Group, Faculty of Information Technology and Electrical Engineering,

University of Oulu, Oulu, Finland

Informatics and Statistics Department, Federal University of Santa Catarina, Florian

opolis, Brazil

Keywords:

Autonomous Landing, Unnamed Aerial Vehicle (UAV), Computer Vision, Wildﬁre Monitoring.

Abstract:

Wildﬁre has been an environmental, economic, and health problem worldwide. Technological advances have

led to the popularization of Unmanned Aerial Vehicles (UAVs) for personal and business use. One of the

Unmanned Aerial Vehicle (UAV) applications is monitoring. However, UAVs still have payload and battery

limitations. UAVs can be an ally for wildﬁre management, but their use is challenging considering their

restraints and the large size of monitored areas. Therefore, it is necessary to develop approaches to circumvent

UAV limitations. This work’s approach allows a drone to land in strategic locations for data acquisition,

resulting in signiﬁcantly less battery consumption. The method uses principles from stereo vision through

a monocular camera motion to estimate the relative position of a selected landing site, allowing a drone to

hang itself by a hook in an artiﬁcial (e.g., aluminum frame, power line) or natural (e.g., tree branch) location.

However, the system is limited to static landing sites where the FAST feature detector algorithm can detect

features. The results showed that the landing site estimation system achieves over 90% accuracy in controlled

scenarios. Moreover, the Landing Site Estimation System (LSES) allied with navigation controllers achieved

95% success in landing attempts with light and wind under control.

1 INTRODUCTION

Forest loss has been a global concern due to its re-

lation to the greenhouse effect, public health, and

economic aspects. However, the problem does not

rely only on deforestation. Wildﬁres represented 26-

29% of global 2001–2019 forest destruction (Tyukav-

ina et al., 2022). Besides that, the smoke generated

by these wildﬁres can negatively affect public health,

especially in children (Holm et al., 2021). Forest

ﬁres also affected the worldwide economy, costing

around US$ 68 billion and representing around 41%

of economic losses reported between 1998 and 2017

(EM-DAT, 2018). Therefore, wildﬁre monitoring and

management is an important research subject to de-

crease forest loss, impacting positively the economy

and health factors.

Unmanned Aerial Vehicles (UAVs) can play an

https://orcid.org/0000-0003-1465-3101

https://orcid.org/0000-0001-9993-8602

https://orcid.org/0000-0002-3787-7549

https://orcid.org/0000-0001-8029-9468

important role in helping to monitor wildﬁres. How-

ever, despite recent technological advances, using

UAVs is still challenging. The challenges can involve

the UAV’s design considering the mission, intended

weather of operation, legislation regarding the UAV

operator, human-computer factor, and infrastructure

available. These challenges are explored deeply in

(Kramar et al., 2022).

Regarding design challenges, UAVs have techni-

cal limitations such as battery efﬁciency, low ﬂight

time, communication range, and payload capacity

(Von Bueren et al., 2015; Kim et al., 2019). Bat-

tery and payload restrictions impact the sensors used

in UAVs. The sensors must be lightweight, small,

and have low energy consumption (Tsouros et al.,

2019). These limitations also impact the application

range. For instance, in large areas, only one Un-

manned Aerial Vehicle (UAV) might not monitor an

entire ﬁeld because of its limited energy resources

(Radoglou-Grammatikis et al., 2020).

In an agricultural ﬁeld, usually, a drone has

to perform full coverage to monitor the desired

Kunz Cechinel, A., Röning, J., Tikanmaki, A., DePieri, E. and Della Méa Plentz, P.

Hanging Drone: An Approach to UAV Landing for Monitoring.

DOI: 10.5220/0012154900003543

In Proceedings of the 20th International Conference on Informatics in Control, Automation and Robotics (ICINCO 2023) - Volume 1, pages 363-373

ISBN: 978-989-758-670-5; ISSN: 2184-2809

363

area. In area surveillance missions, while performing

permanent/long-term monitoring, the drone can act as

a static sensor while being landed in some location

with the advantage that it can change locations. This

case applies to a drone navigating among strategic lo-

cations under tree canopies in wildﬁre management

scenarios instead of sweeping the entire area. For in-

stance, positioning the drone against the wind can al-

low the smoke to move toward the drone. Thus, it is

not required for the drone to ﬂy seeking smoke. The

drone staying landed in these locations while acquir-

ing air-quality data leads to lower battery consump-

tion and allows a single drone to cover a larger area.

Drones with perching ability can save energy in

surveillance missions by staying landed while acquir-

ing data (Kirchgeorg and Mintchev, 2022; Kitchen

et al., 2020; Hang et al., 2019). However, these

approaches usually require a grasping mechanism to

hold itself to the landing site. These mechanisms add

weight and mechanical complexity to the drone’s de-

sign, which is undesirable for small drones. Using a

hook can be an alternative with a small weight, sim-

pler than a perching mechanism, and with the same

purpose.

This work proposes an approach to UAV landing

for monitoring, considering the wildﬁre management

requirements and battery limitations in drones. The

method uses principles from Stereo Vision (SV) with

a monocular camera’s motion to estimate the relative

position of a selected landing site. Therefore, it can be

used for landing in either constructed areas, such as an

aluminum frame, or natural ones, like tree branches.

After detecting the landing site’s position, the sys-

tem lands the drone on the site by hanging its hook.

The system is limited to static landing sites where the

FAST algorithm can detect features.

The remainder of this paper is organized as fol-

lows: Section 2 presents related works. Then, Section

3 shows the approach of UAV landing for monitoring.

After that, Section 4 addresses the experiments’ de-

scription, results, and discussion. Finally, Section 5

concludes the paper and proposes future works.

2 RELATED WORKS

Autonomous landing for UAVs is an utmost abil-

ity for several applications such as package delivery

(Feng et al., 2020), battery self-charging (Junaid et al.,

2017), and cooperation in search and rescue (Persson

and Wahlberg, 2019). Works regarding autonomous

landing usually rely upon two categories: Landing

on a static platform (Xuan-Mung et al., 2020; Junaid

et al., 2017; Giernacki et al., 2017), and Landing on

a moving platform (Gautam et al., 2022; Rodriguez-

Ramos et al., 2019; Feng et al., 2018; Borowczyk

et al., 2017).

In both cases, usually, the drone must land on

a platform with its landing gear touching it. Alter-

natively, researchers developed perching mechanisms

inspired by nature. Thus, drones can behave like

birds and perch by grasping a branch (Kirchgeorg and

Mintchev, 2022), power line (Kitchen et al., 2020),

and edges or corners of buildings (Hang et al., 2019).

To the best of the authors’ knowledge, this is the ﬁrst

work that addresses the landing maneuver utilizing

a hook, allowing the drone to hang itself in a tree

branch. Therefore, this work also creates a new re-

search problem in the autonomous landing topic: the

autonomous landing of a drone equipped with a hook.

This problem is relevant because a drone hung on a

branch, power line, or power post at a sufﬁcient height

can monitor the environment, spending much less bat-

tery than during the ﬂight.

An important step in autonomous landing is the

landing site’s location detection. It can depend sim-

ply on GPS and IMU data (Giernacki et al., 2017),

computer vision with downward (Gautam et al., 2022;

Xuan-Mung et al., 2020; Junaid et al., 2017) or gim-

baled (Feng et al., 2018; Borowczyk et al., 2017)

camera, and external sensors. In (Rodriguez-Ramos

et al., 2019), the authors utilized the OptiTrack Mo-

tion Capture system (MoCap) to estimate the drone

location and the moving platform, thus allowing the

drone controller to know the position of the landing

site precisely.

Another approach to detecting the landing site’s

location is to capture images using a camera and ap-

ply computer vision to estimate the relative location

of a target at the landing site. In these approaches, the

system intends to track the target and keep the drone

camera aligned during the descending step. These

targets can be AprilTags (Gautam et al., 2022; Feng

et al., 2018; Borowczyk et al., 2017) or have a spe-

ciﬁc shape and color (blobs) (Junaid et al., 2017). It is

also possible to combine AprilTags with blobs. Thus,

when the landing site is far, it is detected as a blob,

and when it is near as an AprilTag (Gautam et al.,

2022).

GPS sensors and vision systems can also be com-

bined. Considering a known landing site at a dis-

tant location, the drone can navigate toward it using

GPS localization. Then, when the drone reaches the

surroundings of the platform, it uses a vision sys-

tem to land (Junaid et al., 2017; Borowczyk et al.,

2017). Besides RGB cameras, equipping the drone

with an infrared (IR) camera and the landing site with

an IR beacon enables the system to detect the target

ICINCO 2023 - 20th International Conference on Informatics in Control, Automation and Robotics

364

and landing, even during nighttime and/or in light-

inconvenient environments (Xuan-Mung et al., 2020)

Besides the landing site’s location estimation, it is

interesting to detect when the drone touches the plat-

form instead of shutting off the motors directly below

a threshold altitude. Turning off the motors at an inap-

propriate moment might result in the toppling of the

vehicle and/or harm to onboard equipment (Gautam

et al., 2022). Authors have been using sensors mea-

suring pressure (Rodriguez-Ramos et al., 2019) and

distance (Feng et al., 2018) to detect when the drone

touches the landing site.

Besides the hook, the system developed in this

work differs from the literature by using computer vi-

sion on images from a ﬁxed forward camera. In ad-

dition, the vision system does not rely on shapes or

artiﬁcial markers to allow the detection of the land-

ing site. However, the landing site must have features

detectable by the FAST (Rosten et al., 2010) algo-

rithm and can operate only in light-convenient envi-

ronments. The drone employed in this work has op-

tical ﬂow localization, which limits the trust in the

drone’s localization. However, the system has proved

to be sufﬁciently good for a landing site detectable in

the camera. The proposed approach also embodied a

landing detection by using IMU information to detect

the impact, which also differs from previous works

presented in the literature.

3 APPROACH

This section presents the approach to landing a drone

autonomously by hanging itself by its hook in a land-

ing site selected by a user. First, it shows the drone

preparation and limitations detected to solve the prob-

lem. Then, it explains the adopted control architec-

ture. After that, the Landing Site Estimation System

(LSES) is explored to explain how the system detects

the landing site inside the area selected by a user. Fi-

nally, it presents the steps of the proposed landing

pipeline.

3.1 Drone Preparation

The Robomaster Tello Talent (RMTT) drone was used

in this work to develop a landing system for a drone

equipped with a hook. Figures 1 and 2 show de-

tails about its dimensions. The hook was ﬁxed in the

drone top using epoxy glue. After adding the hook,

the drone’s center of mass was calibrated to improve

ﬂight stability using the Tello app provided by DJI.

Figure 1 shows that the hook has approximately

10 cm of space to hit the landing site (green ar-

Figure 1: Drone’s front view.

Figure 2: Drone’s side view.

row). Therefore, the landing system always attempts

to maintain the drone height to hit the landing site

in the middle of this space (5 cm above the cage).

Since the protection cage has around 9 cm (blue ar-

row) length from its middle, according to Equation 1,

the landing site cannot have a roll angle greater than

. Therefore, considering a tolerance of +/- 2.5 cm

(Equation 2), the system lands only in locations with

a detected roll angle smaller than 15

for safety mea-

sures.

arctan





≈ 29

(1)

arctan



5 − 2.5



≈ 15.5

(2)

It is important that when the drone hits the landing

site, its hook is facing forward. However, errors in the

Hanging Drone: An Approach to UAV Landing for Monitoring

365

yaw angle do not affect the landing signiﬁcantly. The

observation has shown that errors around 10

do not

affect the landing ability.

3.2 Architecture

The proposed system architecture considers a direct

connection between the drone and the server. How-

ever, the system could have intermediary routers if

they are sufﬁciently fast to transmit information. Fig-

ure 3 shows the architecture used in this work. The

server system receives images and odometry data

from the drone through Wi-Fi. Then, based on this

data, it sends commands to the drone to perform the

landing.

Figure 3: System architecture. Green arrows represent data

transference. The blue arrow represents command sending.

The Drone system has Visual Inertial Odometry

(VIO) and camera modules. The VIO is a propri-

etary and closed system. It uses a camera, barometer,

ToF, and accelerometer to estimate the drone’s odom-

etry. This module provides a quaternion and three-

dimensional position estimation. The quaternion al-

lows the server to compute the drone’s yaw, pitch, and

roll angles. Unfortunately, the VIO’s precision is not

provided by DJI. The camera module delivers images

with 960x720 pixels at approximately 30 frames per

second, while the VIO system provides data at 10 Hz.

The development of the Server system utilized

Python and predominantly relied on two libraries:

OpenCV (Bradski, 2000) for image processing and

the reverse-engineered TelloPY for both data recep-

tion and command transmission. This system is com-

posed of four modules: Vision, Landing Site Estima-

tion System (LSES), Screen-based PID, and Position-

based PID. The Vision module is responsible for three

tasks:

• For each image received, the vision module ap-

plies the Sharpen Filter (SF) followed by the Bi-

lateral Filter (BF): the SF enhances the objects’

boundaries but increases image noise, then the BF

reduces the noise while maintaining the improve-

ment in the image. This process increases the

number of features detected in the image.

• Feature detection: given an area deﬁned by a

polygon in an image, the Vision module uses the

FAST algorithm (Rosten et al., 2010) to detect

features inside the delimited area.

• Feature matching: given a set of features and an

image, the vision system matches the informed

features in the provided image using brute force.

The landing starts with a user selecting a land-

ing site in an image from the drone’s camera during

the ﬂight. Then, the Landing Site Estimation System

(LSES) combines features from pictures taken by the

drone at different heights and VIO data to estimate the

relative position of the selected landing site. The fol-

lowing subsection presents further information about

this module. We chose to estimate the relative coor-

dinates instead of performing servoing of the landing

site due to the positions of the camera and the hook

on the drone. In the servoing, the landing site will

go out of the camera’s ﬁeld of view as the drone ap-

proaches it. Then, the problem would fall back on

estimating the drone’s relative position to the landing

site. Therefore, we decided to compute the relative

position at the start of the landing.

The PID modules control the movements of the

drone relying on camera and or VIO data. Given

a pixel coordinate (x, y) and a set of features, the

screen-based PID modules centralize the pixel co-

ordinate with the average position from the features

matched in the current picture provided by the drone.

The yaw PID keeps the features aligned with the x

coordinate, and the height PID with the y coordinate.

Each PID module works independently of each other.

The Server system was limited to running around

25 frames per second during the landing pipeline

execution, which allowed run feature detection and

matching at a continuous rate. This choice was es-

sential to ensure the proper functioning of the screen-

based PID modules.

The position-based PID modules are majorly used

for navigation. They bring the drone from its cur-

rent position to the chosen setpoint, which can be the

landing site. As in the screen-based modules, each

position-based PID module works independently of

the others.

3.3 Landing Site Estimation System

After the user deﬁnes the landing site in the received

images, the Server will use the Vision module to cap-

ICINCO 2023 - 20th International Conference on Informatics in Control, Automation and Robotics

366

ture features from the landing site at two heights.

Then, the Server forwards these features to the Land-

ing Site Estimation System (LSES), which uses prin-

ciples from Stereo Vision (SV) to compute the land-

ing site’s relative position to the camera. Usually, an

SV system has two parallel cameras. In this work,

we use SV from motion by taking photos at different

known heights with the same camera.

Figure 4 illustrates a simpliﬁed example of the ge-

ometry behind the LSES. In Picture 1, a feature re-

garding the object is detected in the Y coordinate rep-

resented by the blue arrow. After the drone moves

up, in Picture 2, the same object’s feature appears in

a new Y coordinate represented by the red arrow. The

difference between these two coordinates is the dis-

parity.

Based on the disparity, parameters from the cam-

era, and knowing how much the drone moved up,

the LSES estimates the real-world relative position

of each feature using Equations 3-5. The Equations

were formulated to give the features’ real-world po-

sition considering the drone is in the second picture

height. The variation in the features’ X coordinate

must be small for these equations to work. There-

fore, after the drone moves up, it must ensure that the

features are aligned with the same X coordinate from

Step 1. A tolerance of 10 pixels was implemented in

this work.

Figure 4: A simpliﬁed example of the geometry behind the

LSES. The blue arrow represents the object’s Y coordinate

in the ﬁrst image. The red arrow represents the object’s Y

coordinate in the second image.

Y =

· T

Y 2

− I

Y 1

(3)

X = Y ·

− C

(4)

Z = Y ·

− (I

Y 2

+ I

Y 1

) · 0.5

−

(5)

Where:

• X, Y , and Z: world coordinates relative to the

drone’s camera;

• T

: the variation in the drone’s height between the

ﬁrst and second features acquisition;

• F

and F

: focal length for X and Y axis in the

images;

• C

and C

: the coordinates X and Y in the image

that represents the position of an object central-

ized with the camera in the environment;

• I

: the feature’s X coordinate in the second im-

age;

• I

Y 1

and I

Y 2

: the feature’s Y coordinate in the ﬁrst

and second images;

After estimating the world’s relative position of

each feature, the LSES estimates the landing’s site

yaw and roll angle. This is done using the least-

squares technique. Considering Equation 6 to deﬁne

the yaw angle, it uses the features’ coordinates X and

Y , as x and y data respectively, to compute the coef-

ﬁcients a and b. Then, the tan

−1

(a) gives the yaw

angle. The process is similar for the roll angle, in this

case, the system uses the features’ coordinates X and

Z as x and y data.

y = a · x + b (6)

3.4 Landing Pipeline

The proposed system adopts a landing pipeline, which

consists in seven steps to land:

1. User selects landing site;

2. Height and Yaw adjustment using screen-based

PID controllers;

3. Save the drone’s location and landing site’s fea-

tures;

4. Height and Yaw adjustment using position-based

and screen-based PID controllers;

5. Save the drone’s location and the landing site’s

features, then estimate its relative position;

6. Height and X adjustment using position-based

PID controllers;

7. Navigation toward the landing site using position-

based PID controllers;

The landing process starts with a user selecting the

landing site inside a polygon (Figure 5). After that,

the Server will use the Vision module to detect fea-

tures in the landing site and save it.

Considering the features saved from Step 1, the

Server will use the Vision module to match them in

Hanging Drone: An Approach to UAV Landing for Monitoring

367

Figure 5: Step 1.

the current frame. Based on the matched features’ po-

sition, it will use the screen-based PID controllers to

centralize the features in the frame (Figure 6a). When

they are centralized, the drone will save the last fea-

tures matched and the drone’s current height (Figure

6b).

(a) Step 2. (b) Step 3.

Figure 6: Steps 2 and 3.

Then, the Server will use the position-based height

controller to move the drone up TY cm, and the

screen-based yaw controller to centralize the matched

features (Figure 7a). After adjusting the height and

yaw, the drone will save the last features matched and

the drone’s current height again (Figure 7b).

(a) Step 4. (b) Step 5.

Figure 7: Steps 4 and 5.

After the Server has the features and the drone’s

positions at the two heights, the Server forwards these

data to the LSES, which returns the real-world posi-

tion for each feature matched and the landing site’s

roll and yaw angle. Then, the Server computes the

average position of the Y coordinate and chooses the

real-world feature’s location nearer the average. Us-

ing this location, the Server uses the position-based

PID controllers to adjust the drone’s yaw and X po-

sition (Figure 8a). Following, the Server uses all

position-based controllers to navigate the drone to-

ward the selected real-world location (Figure 8b).

Finally, when the Server detects the impact on

the landing site, it holds the pitch controller forcing

forward and the height controller forcing down for

one second. Thus, the drone will land with the hook

touching the landing site (Figure 9). A video show-

ing the system performing the pipeline can be seen at

https://youtu.be/-aCFcoKEJI8

(a) Step 6. (b) Step 7.

Figure 8: Steps 6 and 7.

Figure 9: Hit in the landing site.

4 RESULTS AND DISCUSSION

This section presents the results and discussion. First,

it explores the RMTT’s battery consumption in differ-

ICINCO 2023 - 20th International Conference on Informatics in Control, Automation and Robotics

368

ent scenarios. After that, it presents an experiment to

prove that the landing site estimation system (LSES)

works in a controlled scenario. Then, it shows exper-

iments in ﬂight, where even with the LSES accuracy

impaired, the drone can land inside the landing site.

4.1 Battery Consumption

This experiment aimed to evaluate the RMTT’s bat-

tery consumption and was conducted using a power

supply. We connected the drone to the power supply

set at 4.1 V. Then, we measured the current consumed

in six scenarios for one minute.

Figure 10 shows the consumption results for each

case. Note that each scenario includes the previous

ones. The graph indicates that the drone needs a sig-

niﬁcantly smaller power consumption when not ﬂy-

ing. For instance, when the drone is with the motors

on to prevent overheating, it consumes 84.6% less

current than when it is hovering. Therefore, if sus-

pended in a tree, the drone can monitor the environ-

ment signiﬁcantly longer.

Figure 10: Current consumption by scenario.

4.2 Experiment in Controlled Scenario

This experiment aimed to analyze the accuracy of the

LSES. The landing site was positioned to the drone in

known relative positions. Since the LSES system can

also detect the yaw and roll angles of the landing site,

the landing site also had known yaw and roll angles.

Thus, the experiment evaluated the accuracy of rela-

tive position and angle estimations. Figure 11 shows

crops of images used in this experiment, the ﬁrst com-

pletely aligned, the second with a 16

roll angle, and

the third with a 45

yaw angle. The landing site size

was 40x4.5 cm with three tags drawn to ease the fea-

ture detection process.

The drone was positioned on a height-adjustable

table. Then it was taken two photos using its cam-

era. The second photo was taken 15 cm above the ﬁrst

one. A user selected the landing site in the ﬁrst pic-

ture. Then, the LSES system estimated the position in

the real world of the detected and matched features in

the selected landing site. After that, it computed the

average location of the features in the environment.

The yaw and roll angles are estimated based on the

set of points using the least squares. Therefore, they

are absolute values. The estimated average position

and angles were then compared with the center loca-

tion of the real-world landing site.

Table 1 presents the center location of the real-

world landing site for each experimented position and

angle. The X coordinate represents the drone’s left

and right distance, the Y coordinate represents the

drone’s forward and backward distance, and the Z co-

ordinate represents the height relative to the drone.

We chose the Y coordinate distances considering the

image quality in the RMTT’s camera. Preliminary ex-

periments have shown that the system does not detect

and match features properly for landing sites more

than 200 cm away.

Table 1: Real-world landing site center location relative to

the drone’s camera.

X (cm) Y (cm) Z (cm) Yaw (

) Roll (

)

15.18 100.00 -29.60 0.00 0.00

15.18 140.00 -29.60 0.00 0.00

15.18 180.00 -29.60 0.00 0.00

14.87 140.00 -26.13 0.00 10.00

14.40 140.00 -24.09 0.00 16.00

13.97 146.84 -29.60 20.00 0.00

9.32 154.14 -29.60 45.00 0.00

Figure 11: Examples of roll and yaw variations within the

experiment.

Hanging Drone: An Approach to UAV Landing for Monitoring

369

First, the landing site was positioned at 100, 140,

and 180 cm without yaw and roll angles. After that,

the landing site was ﬁxed at 140 cm and its roll angle

was set at 10

and 16

. In the last part, the landing site

was also ﬁxed at 140 cm and the yaw angle was set at

and 45

. Table 2 presents the estimated landing

site center location relative to the drone’s camera in

the second image.

Table 2: Estimated landing site center location relative to

the drone’s camera.

X (cm) Y (cm) Z (cm) Yaw (

) Roll (

)

17.38 96.85 -29.41 -0.24 -1.39

16.22 137.46 -29.70 0.18 0.18

17.49 177.31 -29.99 2.87 -0.18

16.57 134.47 -24.00 2.60 10.46

12.38 131.71 -22.87 1.48 18.53

12.73 143.00 -29.56 17.99 -0.78

7.96 146.97 -29.03 43.03 0.24

Based on Tables 1 and 2, Table 3 presents the ac-

curacy and angle error to the landing site center loca-

tion relative to the drone’s camera. The accuracy was

computed using the Equation 7 and the error in roll

and yaw is the absolute difference between the esti-

mated and the real angle.

Accuracy = 1 −



estimated − real

real



(7)

The worse accuracy result in Table 3 is related to

the X coordinate. This coordinate is the most affected

by the average position of the detected features. Since

the system detected more features in the middle and

right tags, the average of the X coordinate moved to-

ward the right. The distance to the landing site (Y )

and height (Z) had an accuracy of over 96%. The yaw

and roll angle estimations achieved errors above 3

more than enough to allow the drone hit the landing

site facing forward. This experiment showed that con-

sidering the accuracy and error on position and angle

estimations, the LSES can be used to detect a static

landing site.

4.3 Experiment in Flight

After evaluating the LSES in a controlled scenario,

it was embedded into the drone navigation system.

Combined with the navigation controllers, it allows

a drone to land by hanging its hook in a selected area.

For this experiment, the drone’s takeoff point was

positioned about 180 cm to the right end of the land-

ing site. The landing site had about -20

yaw and 0

roll. It is important to mention that after the takeoff,

the drone did not maintain the 180 cm distance be-

Table 3: Position accuracy and angle error to the landing

site center location relative to the drone’s camera.

Accuracy (%) Error (

)

X Y Z YAW ROLL

85.44 96.85 99.36 0.24 1.39

93.09 98.19 99.68 0.18 0.18

84.73 98.51 98.68 2.87 0.18

88.60 96.05 91.85 2.60 0.46

85.98 94.08 94.96 1.48 2.53

91.12 97.38 99.85 2.01 0.78

85.45 95.35 98.06 1.97 0.24

Mean

87.77 96.63 97.49 1.62 0.82

cause it moved around until it became stable in a po-

sition. The user selected the landing site after a drone

took off and stabilized in a position for each experi-

ment attempt. The RMTT drone has an optical ﬂow

system to estimate its relative position from the take-

off point. Since the lab’s ground reﬂects too much

light, pads were scattered on the ground so the optical

ﬂow could work better. Figure 12 shows the experi-

mental environment.

Figure 12: Experimental environment. The takeoff region

is marked with a blue square.

This experiment evaluates the landing success

rate, the drone’s routes, and the LSES’s accuracy dur-

ing ﬂight. Twenty landing attempts were performed,

and only one was unsuccessful because the height

controller did not adjust the drone’s height in time to

hit the landing site. For each successful landing, it

was recorded the minimum, maximum, and the esti-

mated landing point, the actual landing position, and

ICINCO 2023 - 20th International Conference on Informatics in Control, Automation and Robotics

370

the route performed during the landing. The esti-

mated position of the landing site is relative to the

drone’s coordinates, informed by the VIO in step 5

of the landing pipeline. The actual landing position is

the coordinates the VIO provides when the drone hits

the landing site.

Table 4 shows the error between actual and esti-

mated landing points for the 19 successful landing at-

tempts. The error was calculated simply by the differ-

ence between the actual and estimated landing points.

Columns X, Z, and YAW had minor errors. All the

landings had the impact point inside the landing site’s

maximum and minimum X coordinates. In absolute

values, the most signiﬁcant error in Z was 2.183 cm.

Considering that the hook has approximately 10 cm

space to hit the landing site, and the system sets the

landing site in the middle of this space, an error up to

4 cm would still allow the landing. Therefore, these

results are adequate to the system’s requirements.

The system adds 20 cm plus the drone’s abso-

lute movement in Y to the estimated landing site’s

Y coordinate to ensure that the drone hits the loca-

tion. Therefore, it was expected an error in the Y

column (Table 4) around -20 cm. The negative val-

ues mean that the drone hit the landing site in a Y

position nearer than expected, which makes sense.

The high errors, like -49 cm, are related to instabil-

ity in the drone position during the acquisition step.

In those cases, the drone moved forward, making the

disparity of the features larger, resulting in larger Y

distances, which were also increased by adding the

drone’s Y movement. If the drone had moved back-

ward, the disparity would be decreased, making the

Y distance smaller. However, adding the drone’s ab-

solute Y movements would help to ﬁx the estimation.

The logic behind this is that when the drone estimates

the landing site further than where it is, it is not a

problem because the system expects to impact it.

The three positive red values (column Y, Table 4)

mean that the drone landed in a position further than

expected. This was possible because when the drone

achieves the landing site position or impacts the land-

ing site, the system sets the Y controller to half of the

maximum forward force for one second. After that,

it also sets the height controller to half the maximum

downward force and then turns off the motors. This

behavior aims to make the drone keep leaning on the

site during the landing. However, if the drone has not

yet impacted the landing site and is near enough, this

behavior will ensure that the drone lands.

Figures 13 and 14 show the performed route for

the best result. Figures 15 and 16 show the performed

route for the worse result. The system adjusts the X

and Z coordinates after deﬁning the landing site’s rel-

Table 4: Error between actual and estimated landing points.

Error (cm) Error (

)

X Y Z YAW

4.45 -5.99 -0.23 0.60

2.90 -13.54 -2.18 -0.20

6.75 -18.60 0.26 -2.67

3.89 -8.32 0.13 -1.23

6.48 -15.79 1.07 0.10

6.57 -32.04 -0.65 -0.18

4.74 -12.66 1.01 -0.39

-1.10 -11.33 0.19 0.54

4.77 -71.87 1.45 -1.29

13.34 -22.81 1.75 -9.15

3.63 -6.80 -0.82 2.58

-2.59 1.66 -0.67 -0.62

-1.37 -1.62 1.74 -2.08

-0.66 11.14 -1.62 -1.00

0.19 -13.78 1.64 -10.30

1.92 -49.42 -0.40 -3.14

5.37 -0.62 -0.43 -0.28

-1.36 3.09 0.53 -1.77

-1.06 -2.14 0.47 -1.14

Average

2.99 -14.29 0.170 -1.66

ative position. Then, after achieving a certain thresh-

old, the drone navigates toward the desired landing

point. This behavior can be observed in both the best

and worst results.

After the initial adjustment in X and Z coordi-

nates, the drone navigates toward the landing point

while adjusting its yaw to impact the landing site fac-

ing toward. Since the landing site has about -20

yaw

angle relative to the drone, it adjusts its yaw by ro-

tating clockwise so the route moves towards the right

side. After adjusting the yaw, the roll and pitch con-

trollers ﬁx the drone position to impact the landing

site. This behavior can also be observed in both the

best and worse results.

In the worst result (Figures 15 and 16), the route

ends around the desired X and Z coordinates. How-

ever, the desired Y coordinate was around 70 cm from

the landing point. The acquisition step had an un-

stable ﬂight, which resulted in a poor Y estimation.

However, the poor estimation in Y did not signiﬁ-

cantly affect the estimation in X and Z, which allowed

the drone to land. Therefore, an error that results in

a larger Y distance still enables the landing to be ac-

complished. This is not true when estimating a nearer

Y coordinate with a large error. Thus, the safety mea-

sure that adds 20 cm plus the absolute traveled dis-

tance in Y during the acquisition step helps the drone

to land in unstable acquisition scenarios.

Hanging Drone: An Approach to UAV Landing for Monitoring

371

Figure 13: Route in X and Y axis for the best result.

Figure 14: Route in Y and Z axis for the best result.

5 CONCLUSIONS

This work proposed an approach to UAV landing for

monitoring. The results conﬁrmed that a drone can

monitor an environment signiﬁcantly longer when not

ﬂying. Besides that, the landing site’s relative posi-

tion estimation system showed an accuracy of over

90% in a controlled scenario. Thus, it has enough ac-

curacy to be used as a vision system during the land-

ing. The system evaluated during ﬂight allowed the

drone to land on 19 of 20 attempts. Besides that,

landing site position poor estimations were related

to instabilities during ﬂight. Consequently, further

research could be done to assess whether more sta-

ble ﬂight controllers would lead to improved land-

ing sites’ relative position estimations. The system

Figure 15: Route in X and Y axis for the worse result.

Figure 16: Route in Y and Z axis for the worse result.

is limited to the landing sites where the FAST algo-

rithm can detect features. Therefore, in future work,

it would be interesting to improve the feature detec-

tion system for scenarios such as tree branches and

power lines. In addition to enhancing the feature de-

tection system, further research could uplift the qual-

ity of estimations by implementing multiple estima-

tions as the drone advances toward the designated

landing site. Besides that, the experiments were con-

ducted within a controlled environment without per-

turbation (e.g., wind or smoke). Hence, improving

the system for outdoor scenarios such as lighting vari-

ations from sunlight, shadows, smoke, and occlusions

in tree branches, would be interesting.

ICINCO 2023 - 20th International Conference on Informatics in Control, Automation and Robotics

372

ACKNOWLEDGEMENTS

The authors are grateful for the ﬁnancial support

granted by the Conselho Nacional de Desenvolvi-

mento Cient

ıﬁco e Tecnol

ogico (CNPq) and the

Coordenac¸

ao de Aperfeic¸oamento de Pessoal de

ıvel Superior - Brasil (CAPES) - Finance Code

001. The authors wish to acknowledge the Academy

of Finland support via RoboMesh (Decision num-

ber: 336060) Beyond 5G Distributed Ledger Technol-

ogy driven Mesh for Industrial Robot and Collabora-

tion, FireMan (Decision number: 348008) Unmanned

aerial systems based solutions for real-time manage-

ment of wildﬁres and Aeropolis (Decision number:

348479) Sustainable and autonomous carbon-neutral

aerial ecosystems and energy solutions for future

metropolises.

REFERENCES

Borowczyk, A., Nguyen, D.-T., Phu-Van Nguyen, A.,

Nguyen, D. Q., Saussi

e, D., and Ny, J. L. (2017).

Autonomous landing of a multirotor micro air ve-

hicle on a high velocity ground vehicle**this work

was partially supported by cﬁ jelf award 32848 and

a hardware donation from dji. IFAC-PapersOnLine,

50(1):10488–10494. 20th IFAC World Congress.

Bradski, G. (2000). The OpenCV Library. Dr. Dobb’s Jour-

nal of Software Tools.

EM-DAT (2018). UNISDR and CRED report: Eco-

nomic losses, poverty & disasters (1998–2017).

Retrieved from https://www.unisdr.org/ﬁles/61119

credeconomiclosses.pdf.

Feng, K., Li, W., Ge, S., and Pan, F. (2020). Packages de-

livery based on marker detection for uavs. In 2020

Chinese Control And Decision Conference (CCDC),

pages 2094–2099.

Feng, Y., Zhang, C., Baek, S., Rawashdeh, S., and Mo-

hammadi, A. (2018). Autonomous landing of a uav

on a moving platform using model predictive control.

Drones, 2(4).

Gautam, A., Singh, M., Sujit, P. B., and Saripalli, S. (2022).

Autonomous quadcopter landing on a moving target.

Sensors, 22(3).

Giernacki, W., Skwierczy

nski, M., Witwicki, W., Wro

nski,

P., and Kozierski, P. (2017). Crazyﬂie 2.0 quadrotor as

a platform for research and education in robotics and

control engineering. In 2017 22nd International Con-

ference on Methods and Models in Automation and

Robotics (MMAR), pages 37–42.

Hang, K., Lyu, X., Song, H., Stork, J. A., Dollar, A. M.,

Kragic, D., and Zhang, F. (2019). Perching and rest-

ing—a paradigm for uav maneuvering with modular-

ized landing gears. Science Robotics, 4(28):eaau6637.

Holm, S. M., Miller, M. D., and Balmes, J. R. (2021).

Health effects of wildﬁre smoke in children and public

health tools: a narrative review. Journal of Exposure

Science & Environmental Epidemiology, 31(1):1–20.

Junaid, A. B., Konoiko, A., Zweiri, Y., Sahinkaya, M. N.,

and Seneviratne, L. (2017). Autonomous wireless

self-charging for multi-rotor unmanned aerial vehi-

cles. Energies, 10(6).

Kim, J., Kim, S., Ju, C., and Son, H. I. (2019). Unmanned

aerial vehicles in agriculture: A review of perspective

of platform, control, and applications. IEEE Access,

7:105100–105115.

Kirchgeorg, S. and Mintchev, S. (2022). Hedgehog:

Drone perching on tree branches with high-friction

origami spines. IEEE Robotics and Automation Let-

ters, 7(1):602–609.

Kitchen, R., Bierwolf, N., Harbertson, S., Platt, B., Owen,

D., Griessmann, K., and Minor, M. A. (2020). Design

and evaluation of a perching hexacopter drone for en-

ergy harvesting from power lines. In 2020 IEEE/RSJ

International Conference on Intelligent Robots and

Systems (IROS), pages 1192–1198.

Kramar, V., R

oning, J., Erkkil

a, J., Hinkula, H., Kolli, T.,

and Rauhala, A. (2022). Unmanned aircraft systems

and the nordic challenges. In Lipping, T., Linna, P.,

and Narra, N., editors, New Developments and Envi-

ronmental Applications of Drones, pages 1–30, Cham.

Springer International Publishing.

Persson, L. and Wahlberg, B. (2019). Model predictive con-

trol for autonomous ship landing in a search and res-

cue scenario. In AIAA Scitech 2019 Forum.

Radoglou-Grammatikis, P., Sarigiannidis, P., Lagkas, T.,

and Moscholios, I. (2020). A compilation of uav ap-

plications for precision agriculture. Computer Net-

works, 172:107148.

Rodriguez-Ramos, A., Sampedro, C., Bavle, H., de la

Puente, P., and Campoy, P. (2019). A deep reinforce-

ment learning strategy for uav autonomous landing on

a moving platform. Journal of Intelligent & Robotic

Systems, 93(1):351–366.

Rosten, E., Porter, R., and Drummond, T. (2010). Faster

and better: A machine learning approach to corner de-

tection. IEEE Transactions on Pattern Analysis and

Machine Intelligence, 32(1):105–119.

Tsouros, D. C., Bibi, S., and Sarigiannidis, P. G. (2019). A

review on uav-based applications for precision agri-

culture. Information, 10(11).

Tyukavina, A., Potapov, P., Hansen, M. C., Pickens, A. H.,

Stehman, S. V., Turubanova, S., Parker, D., Zalles, V.,

Lima, A., Kommareddy, I., Song, X.-P., Wang, L., and

Harris, N. (2022). Global trends of forest loss due to

ﬁre from 2001 to 2019. Frontiers in Remote Sensing,

Von Bueren, S. K., Burkart, A., Hueni, A., Rascher, U.,

Tuohy, M. P., and Yule, I. J. (2015). Deploying four

optical uav-based sensors over grassland: challenges

and limitations. Biogeosciences, 12(1):163–175.

Xuan-Mung, N., Hong, S. K., Nguyen, N. P., Ha, L. N.

N. T., and Le, T.-L. (2020). Autonomous quadcopter

precision landing onto a heaving platform: New

method and experiment. IEEE Access, 8:167192–

167202.

Hanging Drone: An Approach to UAV Landing for Monitoring

373