Ceppi Thesis 2020

Model-based Design of a Line-tracking Algorithm for a Low-cost Mini
Drone through Vision-based Control
BY
PAOLO CEPPI
B.S. in Mechanical engineering, Politecnico di Torino, Turin, Italy, 2018
THESIS
Submitted as partial fulfillment of the requirements
for the degree of Master of Science in Electrical and Computer engineering in the
Graduate College of the
University of Illinois at Chicago, 2020
Chicago, Illinois
Defense Committee:
Igor Paprotny, Chair and Advisor

Milos Zefran
Alessandro Rizzo, Politecnico di Torino
To my Uncle Damiano,
Happiness is to love an uncle like a big brother.
ii
ACKNOWLEDGMENTS
I dedicate this space to the people who, with their support, have helped me in this engaging research work
through which I applied the knowledge acquired during my academic studies.
Firstly and foremost, I would like to thank my UIC advisor, Dr. Igor Paprotny, who, in these five months
of work, has been able to continually guide me with tools, improvements, useful tips during the research, and
drafting of the paper. Thanks to him, I have improved my analysis and problem-solving skills, which will be
useful in my future career.
I am grateful to my advisor in Politecnico di Torino, Dr. Alessandro Rizzo, who allowed me to undertake
this research work at UIC and who has always been available to help me with practical suggestions. Thanks to
him, I acquired a working method that I will surely replicate in the future.
I sincerely thank my parents for their constant support, their patience, and their motivation during this
challenging period and for allowing me to complete academic studies with serenity, focusing on my goals.
I want to express my gratitude to my close relatives, and my best friends, who always made me feel
optimistic and encouraged me not to give up.
A special thanks to my mates of Micromechatronic Systems Lab for welcoming me in their united and
tight-knit group and for supporting me during these months.
Finally, I would like to dedicate this milestone made of sacrifices to myself. I hope it could be the beginning
of a long and brilliant professional career, made of determination, enthusiasm, and desire for learning.
PC
iii
TABLE OF CONTENTS
CHAPTER PAGE
1 INTRODUCTION………………………………………………………… 1
1.1 Brief History of UAV and advantages………….......................... 1
1.2 Applications…………………………………………………….. 2
2 MODEL-BASED DESIGN AND SPECIFICATIONS…………………. 7

2.1 Model-based software design…………………………………... 7
2.1.1 Model-in-the-loop test………………………………………….. 8
2.1.2 Optimization, code generation, and Software-in-the-loop test…. 9
2.1.3 Processor-in-the-loop test………………………………………. 9
2.1.4 Hardware-in-the-loop test………………………………………. 10
2.2 Project goals and specifications………………………………… 10
3 MODELING AND CONTROL OF QUADCOPTERS………………… 13

3.1 Main Components and Working Principles…………................. 13
3.2 Rigid body motion……………………………………………... 18
3.2.1 Kinematics …………………………………………………….. 18
3.2.2 Dynamics ……………………………………………................ 25
3.2.3 Torques and forces generation………………………................. 29
4 HARDWARE DESCRIPTION AND FIRMWARE SETUP…………... 31

4.1 Hardware and instrumentation………….…………..................... 31
4.1.1 PARROT mini drones……………………………….................. 32
4.1.2 Technical characteristics of MAMBO model and main
Components……………………………………………………. 32
4.2 SIMULINK Hardware support package and firmware setup….. 39
5 SOFTWARE TOOLS DESCRIPTION AND MODEL

CONFIGURATION……………………………………………………... 44
5.1 Description of Software Tools used in the MATLAB
environment ………….…………............................................... 44
5.2 Software configuration, Simulink project overview,
and Simulation Model description………….……..................... 45
5.3 Compiler Configuration………….…………............................. 51
5.4 Preliminary Test of the drone motors………………................. 52
5.5 Physical characteristics of Quadcopter model from
Aerospace Blockset………….………….................................... 55
6 SIMULATION MODEL DESCRIPTION AND CONTROL

PROBLEM SETUP……………………………………………………..... 57
6.1 Nonlinear and Linear models in a Model-based design approach 57
6.2 Control system Architecture (Hovering control)……................ 60
6.3 Simulink model structure description………….……................ 63
iv
TABLE OF CONTENTS (continued)
CHAPTER PAGE
6.3.1 Simulation model…………………………………….................. 63

6.3.2 Flight Control System…………………………………………... 63
6.4 Controllers Tuning…………………………………………….... 88
6.5 Hovering Hardware-in-the-Loop Test…………………….......... 92
7 1st LINE-TRACKING ALGORITHM DESIGN…………….................. 96

7.1 Image processing algorithm (Color Thresholder app and
sub-images analysis approach) ………….................................... 96
7.2 Path planning algorithm (Stateflow approach)………................. 101
7.3 Simplified Version of the previous algorithm………………….. 106
7.4 Model-in-the-loop test (standard and simplified
implementations)……………………………………………….. 108
8 2nd LINE-TRACKING ALGORITHM DESIGN……………................. 110

8.1 Image processing algorithm………….…………........................ 110
8.2 Path planning algorithm………….………….............................. 114
8.3 Model-in-the-Loop Test with Simulation…………………….... 118
9 HARDWARE-IN-THE-LOOP TEST AND VALIDATION…………... 120

9.1 Environment setup and HIL test of the first design version
(with issues explanation) ………….…………............................ 120
9.2 HIL Test of the second design version…………………………. 125
10 CONCLUSION AND RESULTS COMPARISON…………………….. 133
APPENDIX A…………………………………………………………….. 135
CITED LITERATURE…………………………………………………... 144
VITA………………………………………………………………………. 148
v
LIST OF TABLES
TABLE PAGE
I TUNED GAINS OF CONTROLLERS ………………………………………. 92

II PARAMETERS VALUE OF THE FIRST ALGORITHM…………………… 109
III ANALYSIS OF THE TIME OF COMPLETION OF THE TRACK
(1st ALGORITHM)……………………………………………………………. 123
IV ANALYSIS OF THE TIME OF COMPLETION OF THE TRACK
(2nd ALGORITHM)……………………………………………………………. 129
vi
LIST OF FIGURES
FIGURE PAGE
1.1 “aerial screw” prototype as designed by Leonardo da Vinci [1] ……........……… 1

1.2 An archaeologist monitoring a historical site [4]…………………………………. 3
1.3 A drone monitoring an industrial power plant [5]………………………………... 4
1.4 Airborne Hyperspectral sensor [6]………………………………………………... 4
1.5 A drone monitoring the construction of a building [7]…………………………… 5
1.6 the US Global Hawk model that flew over the nuclear power plant in
Fukushima after the earthquake, monitoring the reactors [8]…………………….. 5
1.7 Use of Swiss AI and drones to count African wildlife [9]………………………... 6
1.8 the DJI - M200 Series during Search and Rescue in Extreme Environments [10].. 6
2.1 Workflow of Model-based design [11]…………………………………………... 7
2.2 V-shaped Workflow of Model-based design [11]………………………………... 8
2.3 Model-in-the-loop test [11]……………………………………………………….. 9
2.4 Optimization, Code generation, and Software-in-the-loop test [11]……………… 9
2.5 Processor-in-the-loop test [11]……………………………………………………. 10
2.6 Hardware-in-the-loop test [11]……………………………………………………. 10
2.7 a drone flying in the competition arena [13]……………………………………… 11
2.8 rules on the path measurements in the competition arena [14]…………………… 12
3.1 Schematic of the model of a quadrotor [15]………………………………………. 13
3.2 type 1 and type 2 propellers [15]………………………………………………….. 14
3.3 “Plus” and “cross” configurations [15]…………………………………………… 14
3.4 Forces balance between thrusts and weight while Hovering [15]………………… 15
3.5 Pitch in “Plus” configuration [15]………………………………………………… 15
3.6 Pitch in “cross” configuration [15]……………………………………………….. 15
3.7 forces balance during Pitch maneuver [15]………………………………………. 16
3.8 Roll in “plus” configuration [15]…………………………………………………. 16
3.9 Roll in “cross” configuration [15]………………………………………………... 16
3.10 forces balance during Roll maneuver [15]……………………………………….. 17
3.11 Torques balance [15]……………………………………………………………… 17
3.12 Yawing maneuver and torques balance [15]……………………………………… 17
3.13 Rigid body motion model schematic [15]………………………………………… 18
3.14 Vector p in the Cartesian reference frame (inertial) [15]………………………… 19
3.15 quadrotor’s position vectors in the inertial frame [15]…………………………… 20
3.16 translation of the quadrotor [15]………………………………………………….. 20
3.17 quadrotor’s body reference frame and inertial frame [15]……………………….. 21
3.18 Euler angles and rotations [15]…………………………………………………… 23
3.19 quadrotor’s body reference frame and inertial frame [15]……………………….. 25
3.20 roto translation of the quadrotor [15]…………………………………………….. 25
3.21 torque description [15]…………………………………………………………… 27
vii
LIST OF FIGURES (continued)
FIGURE PAGE
3.22 external force on the quadrotor [15]……………………………………………….. 27

3.23 Actuators’ model [15]……………………………………………………………... 29
3.24 propeller model and air propulsion [17]…………………………………………… 30
4.1 PARROT Mambo Fly minidrone [20] and [22]…………………………………… 31
4.2 Safety goggles (left) [24], Bluetooth dongle adapter (center) [25], and USB
cable (right) [26]…………………………………………………………………... 32
4.3 black background (left and center) [42] and white tape for the track (right) [43]… 32
4.4 top (right) and bottom (left) view of PARROT Mambo Fly model [20]………….. 34
4.5 motherboard of Mambo fly model [27]…………………………………………… 35
4.6 IMU with gyroscope and accelerometer [28] and [29]……………………………. 36
4.7 IMU reference frame [29]…………………………………………………………. 36
4.8 propellers (left) and coreless motors (right) [23]………………………………….. 36
4.9 DC motor structure and components [30]…………………………………………. 37
4.10 Coreless DC motor structure [30] and rotor winding [29]………………………… 38
4.11 recognition of the drone through the USB cable [12]……………………………… 39
4.12 Firmware setup [12]………………………………………………………………... 40
4.13 RNDIS driver setup (1) [21]……………………………………………………….. 40
4.16 firmware checking (left) and
Bluetooth radio device setup (center and right) [20] and [44]……………………... 42
4.17 Bluetooth pairing [44]……………………………………………………………… 42
4.18 PAN connection setup [44]………………………………………………………… 43
4.19 Bluetooth connection test [12]……………………………………………………... 43
5.1 The Simulation model……………………………………………………………... 46
5.2 Workspace with the variables and constants (left and center)
and project folders (right)………………………………………………………….. 47
5.3 Simulink 3D environment used for simulation……………………………………. 47
5.4 Flight Control system……………………………………………………………… 48
5.5 Flag block………………………………………………………………………….. 48
5.6 Flag subsystem (1)………………………………………………………………… 48
5.7 Simulation commands (left) and pace settings (right)…………………………….. 48
5.8 Figure 5.8: Flag subsystem (2)……………………………………………………. 49
5.9 Simulink Visualization block……………………………………………………… 50
5.10 Simulink Visualization subsystem………………………………………………… 50
5.11 Commands subsystem…………………………………………………………….. 50
5.12 Modeling bar [12]…………………………………………………………………. 51
5.13 Model settings [12]………………………………………………………………... 52
viii
FIGURE PAGE
5.14 FCS subsystem for the actuators test……………………………………………… 52

5.15 Hardware functionalities………………………………………………………….. 53
5.16 Code generation [12]……………………………………………………………… 53
5.17 C code of the model [12]………………………………………………………….. 53
5.18 Flight control interface with flight time settings, stop and end of the test [12]…... 54
5.19 reference frame of Quadcopter model [31]……………………………………….. 55
6.1 Schematics of Feedback Control loop with a nonlinear model for simulation
(left) and with a linear model for controller design (right) [41]…………………. 58
6.2 Model-based design technique schematic [40]…………………………………… 59
6.3 Components of the airframe model [40]…………………………………………. 60
6.4 Feedback Control loop with a nonlinear model for simulation [34]……………... 63
6.5 Simulation model blocks…………………………………………………………. 64
6.6 Airframe model (nonlinear and linear)………………………………………….... 64
6.7 Quadcopter Linear model…………………………………………………………. 65
6.8 Quadcopter Nonlinear model……………………………………………………... 65
6.9 AC model (actuators)……………………………………………………………… 66
6.10 Gravity Force calculation (left) and Drag calculation (right)……………………... 66
6.11 Drag calculation subsystem……………………………………………………….. 67
6.12 Motor forces and torques………………………………………………………….. 67
6.13 Motor forces and torques (MotorsToW subsystem)………………………………. 67
6.14 Motor forces and torques (Matrix concatenation subsystem)…………………….. 67
6.15 Motor forces and torques (Matrix concatenation subsystem structure)…………... 68
6.16 Applied Force calculation………………………………………………………… 68
6.17 6 DOF model described through Euler angles (left) and change of the
reference frame from Body to Earth (right)………………………………………. 68
6.18 Environment model……………………………………………………………….. 69
6.19 Constant environment…………………………………………………………….. 69
6.20 Variable environment…………………………………………………………….. 70
6.21 Sensors model…………………………………………………………………….. 70
6.22 Sensors’ dynamics………………………………………………………………... 71
6.23 Sensor system (Sensors’ dynamics subsystem)…………………………………... 71
6.24 Camera model…………………………………………………………………….. 71
6.25 IMU and pressure sensor models…………………………………………………. 71
6.26 HAL acquisition creator model…………………………………………………… 72
6.27 the two options by which the FCS could be implemented [35]…………………... 73
6.28 Flight Control System block………………………………………………………. 74
6.29 FCS structure and its subsystems…………………………………………………. 74
6.30 “Control System” subsystem structure……………………………………………. 75
ix
FIGURE PAGE
6.31 Control mode (“1” for Position XYZ, “0” for Orientation Roll-Pitch-Yaw)……… 75
6.32 State estimator block………………………………………………………………. 76
6.33 State estimator model structure with its subsystems………………………………. 76
6.34 Sensor preprocessing subsystem block……………………………………………. 77
6.35 Sensor preprocessing subsystem structure………………………………………… 77
6.36 Sensor data group subsystem……………………………………………………… 77
6.37 X-Y position estimator block……………………………………………………… 78
6.38 X-Y position estimator structure and its subsystems……………………………… 78
6.39 velocity estimator subsystem block (part of the X-Y position estimator)………… 78
6.40 velocity estimator subsystem structure……………………………………………. 79
6.41 Acceleration Handling subsystem block (part of the velocity estimator)………… 79
6.42 Acceleration Handling subsystem structure (part of the velocity estimator)……... 79
6.43 Data Handling subsystem block (part of the velocity estimator)…………………. 79
6.44 Data Handling subsystem structure (part of the velocity estimator)……………… 80
6.45 X-Y position estimator subsystem block
(part of the general X-Y position estimator)……………………………………… 80
6.46 X-Y position estimator subsystem structure……………………………………… 80
6.47 Altitude (Z position) estimator block…………………………………………….. 80
6.48 Altitude (Z position) estimator structure…………………………………………. 81
6.49 Outlier Handling subsystem block (part of the Altitude estimator)……………… 81
6.50 Altitude (Z position) estimator block…………………………………………….. 81
6.51 Complementary filter block for orientation estimation…………………………... 82
6.52 Complementary filter structure…………………………………………………... 82
6.53 Complementary filter subsystem 1………………………………………………. 82
6.54 Complementary filter subsystem 2……………………………………………….. 83
6.55 gyroscope and accelerometer signals and complementary filter estimation [37]… 84
6.56 Controller block………………………………………………………………….... 85
6.57 Controller model structure………………………………………………………… 85
6.58 X-Y position controller block…………………………………………………….. 85
6.59 X-Y position controller structure…………………………………………………. 85
6.60 Attitude controller block………………………………………………………….. 86
6.61 Attitude controller structure………………………………………………………. 86
6.62 Yaw controller block and model structure………………………………………... 86
6.63 Altitude controller block………………………………………………………….. 87
6.64 Altitude controller structure……………………………………………………….. 87
6.65 Motor mixing Algorithm block and model structure (part 1)……………………... 88
6.66 Motor mixing Algorithm block and model structure (part 2)…………………….. 88
6.67 Model linearization for Controllers tuning [35]…………………………………… 88
x
FIGURE PAGE
6.68 Control system simplification for Altitude controller tuning [34]………………… 89

6.69 Linearized Control system used for Altitude controller tuning [35]……………… 90
6.70 Simplified Altitude Controller used for tuning [35]………………………………. 90
6.71 PID Tuner App on Simulink [35]…………………………………………………. 91
6.72 real-time Hovering test– Trajectory (3D view and Top view)……………………. 93
6.73 Real-time Hovering test– Motor speeds…………………………………………... 93
6.74 Real-time Hovering test – X, Y, Z Positions (left) and
corresponding linear speeds (right)……………………………………………….. 94
6.75 Real-time Hovering test – RPY angles (left) and
corresponding angular velocities (right)…………………………………………... 95
7.1 Image processing (left) and Path planning (right) subsystems……………………. 96
7.2 template model creation for Image processing……………………………………. 96
7.3 Color Thresholder app operations……………………………………………….... 97
7.4 Color Thresholder app…………………………………………………………….. 98
7.5 parametric representation of a line through SHT [45]……………………………. 99
7.6 user-defined function “FindAngle”……………………………………………….. 99
7.7 track lines with different angles (positive on the right)…………………………... 100
7.8 Image processing subsystem…………………………………………………….... 100
7.9 Image subdivision into areas for the path planning algorithm…………………..... 100
7.10 Chart 1 subsystem……………………………………………………………….... 102
7.11 Path planning subsystem………………………………………………………….. 103
7.12 Chart 2 subsystem………………………………………………………………… 105
7.13 Simplified version of the Image processing subsystem…………………………... 106
7.14 Simplified Path planning subsystem……………………………………………… 107
7.15 Chart1 (left) and Chart2 (right)…………………………………………………… 107
7.16 Model-in-the-loop test of the first algorithm - standard version…………………. 108
7.17 Model-in-the-loop test of the first algorithm – simplified version……………….. 109
8.1 “Binarization” (left) and “Filter” (right) function blocks………………………… 110
8.2 “Ctrl_radar” function with discrete filter (left) and gains for increments (right).... 111
8.3 Edge detection subsystem block………………………………………………….. 111
8.4 Edge detection subsystem model………………………………………………..... 111
8.5 Image processing subsystem……………………………………………………… 112
8.6 “Borderless” (left), “Filter_C_Det” (center) and “Ctrl_Circle” (right) functions... 113
8.7 Gains for positioning phase before landing……………………………………..... 114
8.8 Butterworth Filter implementation………………………………………………... 114
8.9 Path planning algorithm…………………………………………………………... 115
8.10 Altitude planning………………………………………………………………….. 116
8.11 Tracking function………………………………………………………………..... 116
xi
FIGURE PAGE
8.12 End of line detection (left) and logic for switching to landing phase (right)……… 116
8.13 Tracking with slow movements for positioning…………………………………... 117
8.14 Landing control block…………………………………………………………….. 117
8.15 position signals generation with increments addition…………………………….. 117
8.16 track of the simulation test………………………………………………………... 118
8.17 track of the simulation test………………………………………………………... 118
8.18 x-y increments signals filtered (up) and x-y-z position coordinates of
the drone (bottom)………………………………………………………………… 119
9.1 Example of the track in the environment used for the HIL Test………………….. 120
9.2 Optical flow estimation [40]………………………………………………………. 121
9.3 Issue related to wrong filtering of the track due to darker regions of the path…… 122
9.4 HIL test with 1st algorithm – Trajectory (3D view and Top view)……………….. 123
9.5 HIL test with 1st algorithm – Motor speeds………………………………………. 124
9.6 HIL test with 1st algorithm – X, Y, Z Position and Yaw angle…………………... 124
9.7 Screenshots of the HIL test with 2nd algorithm………………………………….... 125
9.8 HIL test with 2nd algorithm – Trajectory (3D view and Top view)……………… 126
9.9 HIL test with 2nd algorithm – Motor speeds…………………………………….... 126
9.10 HIL test with 2nd algorithm –X, Y, Z Positions (left) and corresponding
linear speeds (right)……………………………………………………………….. 127
9.11 HIL test with 2nd algorithm –RPY angles (left) and corresponding
angular velocities (right)………………………………………………………….. 127
9.12 2nd HIL test with 2nd algorithm – Trajectory (3D view and Top view)………….... 130
9.13 2nd HIL test with 2nd algorithm – Motor speeds…………………………………… 130
9.14 2nd HIL test with 2nd algorithm – X, Y, Z Positions (left) and corresponding
linear speeds (right)………………………………………………………………. 131
9.15 2nd HIL test with 2nd algorithm –RPY angles (left) and corresponding
angular velocities (right)………………………………………………………….. 132
xii
LIST OF ABBREVIATIONS
UAV Unmanned Aerial Vehicle
RPA Remote Piloted Aircraft
DOF Degrees of Freedom
RPY Roll, Pitch, and Yaw
MIL Model in the loop
HIL Hardware in the loop
MMA Motor Mixing Algorithm
PWM Pulse Width Modulation
SONAR Sound Navigation and Ranging
IMU Inertial Measurement Unit
DMP Digital Motion Processor
RNDIS Remote Network Driver Interface Specification
PAN Personal Area network
FCS Flight Control System
PID Proportional-Integral-Derivative
RGB Red Green and Blue (Color space)
SHT Standard Hough Transform
UI User Interface
xiii
SUMMARY
This Thesis research project aims to design a Line-tracking algorithm for a low-cost mini drone through
Vision-based control with image processing techniques. The design process is the application of the principles
of Model-based software design, which is a modern technique to design control systems, based on the
development of a model of the plant and the controller with enough detail to have a realistic representation of
its behavior to accomplish the specifications. The designed model is tested in a simulation environment
(Model-in-the-loop phase). Then, if it satisfies the requirements, it is tested in real-time, deploying the
algorithm on the Hardware to evaluate if its performances are still acceptable or if it requires to be updated.
A significant advantage that characterizes this technique is the auto-code generation, which allows us to
automatically translate the blocks of the model built through Simulink into a C-code executable by the
hardware, instead of writing it manually.
This research project is adapted from a competition organized by Mathworks, which aims to make a drone
follow a line of a specific color and land at the end of it on a circle. The task should be accomplished in as
little time as possible but at the same time remaining stable and following the path as precisely as possible
(within the low-cost limits of the mini drone used). The environment used to design and develop the control
system is MATLAB, with Simulink and their add-on toolboxes like Aerospace blockset, image processing,
computer vision, and Hardware support package for Parrot mini drone, which is the specific company that
made the drone model of this project.
Firstly, the preliminary goal is the accomplishment of the stabilization of flight maneuvers through a
suitable control system architecture and PID controllers tuning.
Then, the Flight Control System design proceeds with Image processing and Path planning subsystems
design. The line-tracking algorithm implementations developed are two. The first one is based on the analysis
of the pixels of the image acquired from the downward-facing camera and elaboration through image
xiv
processing techniques like color thresholding and edge detection. The path planning logic was implemented
through Stateflow, which is an add-on tool of Simulink, useful for State machines design. This first designed
control system also has another simplified version, useful because computationally lighter on the hardware
compared to the first standard version. The second algorithm, instead, is realized by using user-defined
functions, like thresholding operation for noise removal in the binary image, or like the function that searches
and detects the path and the line angles, and by some other already existing functions provided by the computer
vision toolbox.
Finally, their performances were both tested on the hardware and then analyzed and compared. The
validation phase was discussed, commenting on their limits, and highlighting other issues encountered, not
previously noticed within the simulation 3D environment during the Model-in-the-loop test.
xv
CHAPTER 1
INTRODUCTION
1.1. Brief History of UAV and Advantages
The Leonardo da Vinci’s “aerial screw” project in the Renaissance period (Figure 1.1) would have been
only a progenitor of what has been a continuous technological evolution until now dictated by the greatest of
man's desires: flying.
Today, not only can this action be carried out standardly, but it is also possible to control any aircraft while
sitting comfortably.
Figure 1.1: “aerial screw” prototype as designed by Leonardo da Vinci [1]
The “main character” that embodies this new concept of flight is the drone, which is a remotely piloted aircraft,
or more generally a flying device, which is not piloted on board.
As often happens in aviation history, the technique and quality of the aircraft have developed first in the
war field during the two world wars, and the same goes for drones. The progress of technology pushed more
increasingly during the Cold War, and it allowed us to reduce their sizes progressively up to what we know.
Because of the growing commercialization and due to the more excellent usability of the vehicle, the FAA
(Federal Aviation Administration) has decided to regulate multi-copters based on their features and to overall
performance: weight, range, and service ceiling are just some of the parameters taken into consideration [2].
1
2
Without a driver, the drone can perform certain types of missions between two points through an on-board
computer or a remote-control system, and for this reason, it is also called RPA (Remote Piloted Aircraft). They
are also known with other acronyms, many of which are of Anglo-Saxon derivation: in addition to RPA
(Remotely piloted aircraft) can also be referred to as RPV (Remotely piloted vehicle), UVS (Unmanned vehicle
system), ROA (Remotely operated aircraft), UAV (Unmanned aerial vehicle), etc.
Their use is now consolidated in the military field. It is increasing for civil applications, for example, in fire
counteraction and critical operations, for reconnaissance, for remote sensing, research, and, in general, in all
cases where these devices can permit the accomplishment of standardized, time-consuming, or dangerous
missions. They are often carried out with lower economic costs than traditional air vehicles and do not
encounter so many ethical issues as it could be expected a few years ago because of their reliability. Depending
on the range, they can be equipped for long distances with a video camera that allows them to follow in real-
time the movements in first-person: this class is called FPV (First-person view) [3].
The drone, thus, has become the everyday aircraft used for playful purposes as well as for professional
purposes, see aerial footage and surveys, or even short-range air raid. This variety of purposes immediately
defines its characteristics and construction solutions. In response to these purposes, there are numerous
advantages of a rotary-wing aircraft over a fixed-wing model, firstly the economic cost.
The considerable ease of piloting appears fundamental due to technological developments in
microelectronics and now reliable control systems, that make possible lots of varieties of missions, previously
carried out by helicopters with human pilots.
Furthermore, the possibility of flying indoors is not negligible, it has been made much safer now by the
modern control system, and it would have been unthinkable in the past. In recent years, lots of research has
been carried out all over the world regarding the simulation and the control of these devices. They are different
according to the requested task, for example, the drone's ability to be a carrier of a load. By realizing systems
of this form and implementing control systems, it can be possible to carry out missions in which previously
3
manned helicopters were employed.
Examples are the meausurement of the magnetic field of the Earth using a drone with a suspended
magnetometer, or the study of the seabed carried out by a drone with payload immersed in water.
Unlike traditional aircraft, RPA can be used in situations characterized by a great danger to human life and
in areas difficult to reach, also flying at low altitude. For this reason, they can be used during the monitoring
phases of areas affected by natural or artificial disasters (earthquakes, floods, road accidents, etc.).
1.2 Applications
In recent years, technologies related to the development of RPAs systems have undergone a rapid surge. In
particular, technological development in the field of sensors made it possible to equip RPAs with many
different loads, in the visible spectrum (compact digital cameras or professional), infrared (thermal cameras),
multispectral cameras, up to sensors more advanced such as sensors for monitoring air quality. Here are some
civil applications for RPAs:
 The monitoring of Archaeological sites, against the looting and illegal trade. A clear example is the
monitoring of the ancient necropolis of Fifa in Jordan with drones. It is a logical solution, also considering
the difficulty of hiring staff in militarily active areas, such as the Middle East.
Figure 1.2: An archaeologist monitoring a historical site [4]
 Monitoring of Thermal power plants and industrial plants. RPAs can also be used to monitor over time
the electricity production plants, or, more generally, industrial systems, using sensors (thermal imaging
4
cameras, multispectral cameras, etc.).
Figure 1.3: A drone monitoring an industrial power plant [5]
 Remote sensing. It is the technical-scientific discipline or applied science with diagnostic-investigative
purposes that allows us to derive qualitative and quantitative information on the environment and on
objects placed at a distance from a sensor through electromagnetic radiation measurements (emitted,
reflected or transmitted) that interacts with the physical surfaces of interest. Thanks also to the possibility
of flying at very low altitudes and of having small but suitable sensors available, RPAs classified as “mini-
drones” can be used for applications linked to remote sensing, such as the creation of agricultural crop
vigor maps and monitoring the health of vegetation. Another remote sensing useful application is in
creating coverage maps and land-use maps, for the analysis and the support in the phases that follow a
natural disaster immediately or for the monitoring and mapping of the thermal losses of private and public
buildings (houses, warehouses, industrial plants). This is very important in this period when there is much
talk about sustainable development and loss of land to be allocated to areas greens.
Figure 1.4: Airborne Hyperspectral sensor [6]

5
 Aerial photogrammetry and architectural survey: Photogrammetry is a survey technique that allows
acquiring metric data of an object (shape and position) through a pair of stereometric frames. With the
advent of small digital cameras (compact or reflex) which can guarantee a high-quality standard
concerning the image produced, photogrammetry can be approached to the RPA and to their use to create
digital soil models, for orthophoto production and for the architectural survey of infrastructures and
buildings to create 3D models.
Figure 1.5: A drone monitoring the construction of a building [7]
 Environmental monitoring and natural disasters: The RPAs have been actively used in monitoring areas
severely affected by earthquakes and floods. Some examples are the US Global Hawk RPAs that have
flown over the nuclear power plant in Fukushima Dai-ichi, Japan, entering the forbidden zone to monitor
the reactors after the explosions caused by the Tōhoku earthquake in 2011, also taking photos with
infrared sensors. The high radioactivity made the presence of humans impossible.
Figure 1.6: the US Global Hawk model that flew over the nuclear power plant in Fukushima after the
earthquake, monitoring the reactors [8]

6
 Biodiversity and fauna monitoring: RPAs can be used for monitoring wild animals. There are periodic
numerical controls of those species which are either endangered or, on the contrary, which have a high
rate of reproduction that could be a problem, both for the biodiversity of the environment in which they
live and for the economic damage caused to agricultural production and livestock in the area.
Figure 1.7: Use of Swiss AI and drones to count African wildlife [9]
 Search and Rescue Operations: RPAs can play an essential role in search and rescue operations for quick
reconnaissance and detection, after the occurrence of emergencies.
Figure 1.8: the DJI - M200 Series during Search and Rescue in Extreme Environments [10]
 Video footage and photographs for general purpose: RPAs in combination with the latest and lightest
digital video cameras, including general-purpose ones and not only professional, are making themselves
more and more competitive for all those needs of "aerial" shooting, that substituted other complicated and
expensive tools such as the dolly shot or at higher altitudes the helicopter.
CHAPTER 2
MODEL-BASED DESIGN AND SPECIFICATIONS
2.1 Model-based software design
We can imagine that we should design a controller for an industrial robot, a drone, a wind turbine, an
autonomous vehicle, an excavator, or a servo motor. If the code is written manually and the search for
requirements data is document-based, the only way to respond to the previous questions will be through “trial
and error” procedure and tests on a physical prototype, which can also be expensive.
If a single requirement changes, the whole system will have to be redesigned, delaying the delivery of days
or even weeks.
Figure 2.1: Workflow of Model-based design [11]
Following the Model-Based design (MBD) procedure (Figure 2.1), instead of developing a hand-written
code and use hand-written documents, a model of the system can be developed, subdividing it into different
subsystems of variable complexity. For a control system, they are generally the plant, actuators, sensors, and
the controller.
7
8
Model-Based Design consists of a technique empowering quick and cheap design of dynamic systems,
control systems, communication systems, and signal processing. In this method, the model of the system is at
the focal point of the development procedure, beginning from requirements, then implementation, and finally
testing. It can be continuously refined throughout the development process, and it can be simulated at any time
to get an instant view of the system behavior. Multiple scenarios can be tested without any risk, without any
delay and without using any expensive machinery directly. Various techniques can be applied to create the
mathematical model to be inserted into the Simulink project, and the physical system can be represented as
much detailed as the task that must be accomplished requires.
In this project, the ready-made general model of a quadcopter is provided by MathWorks together to a three-
dimensional graphic simulator that represents the system to be controlled (Simulink 3d environment).
In the Figure below (Figure 2.2), the approach that describes Model-based software design is shown, which
is also known as “V-shaped” development flow.
Figure 2.2: V-shaped Workflow of Model-based design [11]
2.1.1 Model-in-the-loop test
In this first stage, both the Plant (systems to be controlled) and the Controller (algorithm controlling the
Plant) are modeled. The Simulation helps in refining both models and evaluate design alternatives.
The model exists entirely in native simulation tools (Simulink and Stateflow), and this phase is useful for
control algorithm development.

9
Figure 2.3: Model-in-the-loop test [11]
2.1.2 Optimization, code generation, and Software-in-the-loop test
In this stage, different transformation rules produce hardware and software. The implementation is co-
simulated with the plant model to test its correctness, and it is still executed on a PC. A portion of the model
is implemented in the native simulation tool (e.g., Simulink / Stateflow), and another portion as executable C-
code. This phase is suitable to test the controller implementation written in C-code.
Figure 2.4: Optimization, Code generation, and Software-in-the-loop test [11]
2.1.3 Processor-in-the-loop test
In this stage, the implementation is deployed and runs on the target hardware (e.g., EVB or ECU). The
implementation is co-simulated with the plant model to test its correctness. However, the system is not running
in real-time yet.
A portion of the model is implemented in the native simulation tool (e.g., Simulink / Stateflow), and another
portion as executable C-code, running on target hardware or rapid prototyping hardware.

10
Figure 2.5: Processor-in-the-loop test [11]
2.1.4 Hardware-in-the-loop test
In this stage, the controller implementation is co-simulated with the plant model to test its correctness. The
controller runs on the target hardware, the plant model on rapid prototyping hardware, and the system is
running in real-time. A portion of the model runs in a real-time simulator, and another portion could be
implemented as physical hardware (ECU). This phase is suitable to test the interactions with hardware and
real-time-performance. The controller algorithm is sometimes distributed among two devices: the ECU and
rapid prototyping hardware.
Figure 2.6: Hardware-in-the-loop test [11]
2.2 Project goals and specifications
This Thesis research project aims to study and to design a model-based algorithm on Simulink and Matlab
through simulation, and to test it by deploying on the Hardware of a Parrot Mambo Fly mini-drone through
the interfaces provided with its Hardware support package in Simulink.
The first goal to be accomplished is to maintain the drone stable during the flight maneuvers adopted to
11
follow the path. In addition to the previous objective, the drone should acquire images through its downward-
facing camera, then elaborate them with several image processing techniques and move the drone along the
track according to a specific path planning control algorithm that receives the vision-based data and elaborate
them, obtaining in this way an autonomous vision-based control system.
The line specifications are adapted from the rules of the competition organized by MathWorks: the
multinational private company specialized in the production of software for mathematical calculations. Its
main products are MATLAB and Simulink. In addition to government and industrial purposes, the software
products made by the company are also used in teaching and research in many universities throughout the
world.
The rules of this competition are based precisely on the same purpose of this Thesis, which is to make the
drone follow autonomously a path of a particular color (specification useful for the image processing algorithm
that should have a filter with a threshold). The project in Simulink is the one that Mathworks provided to the
participants who had to modify the general model of the Quadcopter according to the aim of the competition.
Figure 2.7: a drone flying in the competition arena [13]
Some of the rules describing the track are the following:
 The line track is about 4 inches in width.
 The circular landing figure has 8 inches in diameter.
 The path is made of connected line segments only and does not have any smooth curves at the connections.
 The end of the path is about 10 inches far from the landing circle.
 For the Thesis project, it was chosen either a random color like red or, more conveniently, for the image
12
processing algorithm, a white line on a black background to enhance the contrast useful in the image
processing algorithm.
 The background is not made by a single color (black), but it has texture (different color papers were used
in the tests). This texture helps the optical flow estimation algorithm, through which it is possible to
estimate the speed and the displacement of the drone. This is not necessary for the Simulation 3d
environment, which does not influence the sensors as the physical environment does in real-time.
Figure 2.8: rules on the path measurements in the competition arena [14]
The circle is used to land the drone as part of the competition, but for this Thesis project, this specification
was an optional goal, not necessary to the accomplishment of the main task, which is the line tracking, but it
is considered only to improve the algorithm.

CHAPTER 3
MODELING AND CONTROL OF QUADCOPTERS
3.1. Main Components and Working Principles
In this chapter, the working principle of a Quadrotor will be explained, and its mathematical model will be
analyzed, deriving its kinematics and its dynamical equations (useful for the implementation on Simulink).
The approach followed is the one used by Orsag et al. [16] and Totu [15].
The main components of quadrotors are:
 The frame, which gives physical support for the other components, consists of a center and four arms.
 Electronics and a battery in the center of the frame.
 A motor and a propeller on each arm.
Figure 3.1: Schematic of the model of a quadrotor [15]
A spinning propeller creates thrust, a force that is perpendicular to the propeller’s rotation plane. Besides
the thrust force, a spinning propeller produces a turning effect (or torque) on the quadrotor frame. It is in the
opposite direction to the propeller’s rotation.
There are two types of propellers:
 Type 1, or right-handed propeller, produced thrust in the upward direction when rotating
13
14
counterclockwise.
 Type 2, or left-handed propeller, produces thrust in the upward direction when rotating clockwise.
A quadrotor has two Type 1 and two Type 2 propellers (figure 3.2).
Figure 3.2: type 1 and type 2 propellers [15]
The quadrotor is an underactuated system. It has four propellers, but there are six degrees of spatial freedom.
 Translational degrees of freedom: up/down, forward/backward, left/right.
 Rotational degrees of freedom: heading, pitch, and roll.
Since the actuators (4) are less than the D.O.F. (6), then some directions are not controllable at any given
time. For example, the quadrotor is not capable of moving right without rotating in that direction. The same
goes for forward and backward motions as well.
It is possible to overcome this underactuation problem by designing a control system that couples rotations
and thrusts to achieve the overall tasks.
The two possible quadrotors configurations are the “Plus” configuration and the “Cross” configuration.
Figure 3.3: “Plus” and “cross” configurations [15]
Let us assume that the frame of the quadrotor is perfectly level with the ground.
If the same commands are given to the motors, the overall thrust is in the vertical direction, and it can
15
compensate for the gravity to generate a movement up (Figure 3.4).
Figure 3.4: Forces balance between thrusts and weight while Hovering [15]
If the overall thrust is less than the force of gravity, then the quadrotor will move down.
To pitch, i.e., rotate around the left-right axis, we must create an unbalance in the forward-side and
backward-side forces. Pitching forwards is done by decreasing the force in the forward side and/or increasing
in the backward side.
Pitching backward is done by decreasing the force on the backward side and/or increasing on the forward
side.
Figure 3.5: Pitch in “Plus” configuration [15]
Figure 3.6: Pitch in “cross” configuration [15]
The pitching rotation is coupled with translation on the forward/backward direction. When the quadrotor is
16
pitching forward/backward, it will also move forward/backward.
Figure 3.7: forces balance during Pitch maneuver [15]
To roll, i.e., rotate around the forward-backward axis, we must create an unbalance in the left-side and
right-side forces. Rolling right is done by decreasing the force on the right side and/or increasing on the left
side. Rolling left is done by decreasing the force in the left part and/or increasing in the right part.
Figure 3.8: Roll in “plus” configuration [15]
Figure 3.9: Roll in “cross” configuration [15]
The rolling rotation is coupled with translation on the left/right direction. When the quadrotor is rolling
left/right, it will also move left/right.

17
Figure 3.10: forces balance during Roll maneuver [15]
Yawing is, instead, the rotation around the up/down direction. It is important to remember that rotating
propellers are causing an opposite torque/turning effect on the frame. If all propellers were rotating in the same
direction, the frame would rotate around its up/down axis in the opposite direction, and it would be spinning
in place. The quadrotor has two types of propellers, such that they rotate in the opposite direction in pairs, and
the reaction effect is canceled.
Figure 3.11: Torques balance [15]
The input is increased on the pair of clockwise propellers and decreased on the counterclockwise pair to
create a controlled counterclockwise yawing rotation, whereas the input is increased on the pair of
counterclockwise propellers and decreased on the clockwise pair to create a controlled clockwise yawing
rotation.
Figure 3.12: Yawing maneuver and torques balance [15]

18
Summarizing it is possible to say that Pitching and forward/backward motion are coupled, Rolling and
left/right motion are coupled, Up and down motion is independent, and Yawing (changing of heading) is
independent.
It could be useful to implement a motor mixing algorithm that can convert the roll, pitch, yaw, and thrust
commands into the motor speeds, and which can be useful in the model representation on Simulink.
The principles to generate Thrust, Roll, Pitch, Yaw maneuvers described before are applied in the same way
to the MMA (Motor mixing algorithm) according to the convention followed by Douglas [19].
motorfront right = thrust cmd + yawcmd + pitchcmd + rollcmd (3.1)
motorfront left = thrust cmd − yawcmd + pitchcmd − rollcmd (3.2)
motorback right = thrust cmd − yawcmd − pitchcmd + rollcmd (3.3)
motorback left = thrust cmd + yawcmd − pitchcmd − rollcmd (3.4)
3.2. Rigid body motion
A rigid body object can be seen as a system consisting of a vast (in the limit infinite) number of small (in
the limit infinitesimal) point-mass particles, with the property that the relative positions of the particles relative
to each other are constant (rigidly).
As the following picture (3.13) shows, the Rigid body motion model of the drone is needed to understand
how from generalized forces in input, it is possible to derive the position and the attitude of the drone in the
space.
Figure 3.13: Rigid body motion model schematic [15]
3.2.1. Kinematics
In the Kinematics of the rigid body, there are different concepts introduced and explained. Some of them are
19
the translational and rotational motion, position, linear velocity, and linear acceleration vectors, the rotation
matrix that defines the body frame attitude relative to a fixed inertial frame, angular velocity, and angular
acceleration, motion composition between frames, etc.
Firstly, the cartesian coordinate system must be defined to describe the geometrical vectors like the position
vector p.
Figure 3.14: Vector p in the Cartesian reference frame (inertial) [15]
The vector p can be described either through a geometrical description, like the following:
𝐩 = p0x 𝐢 + p0y 𝐣 + p0z 𝐤 (3.5)
or through an algebraic description:
p0x 1 0 0
𝐩0 = [p0y ] , 𝐢0 = [0] , 𝐣0 = [1] , 𝐤 0 = [0] (3.6)
p0z 0 0 1
It could be possible to define an inertial cartesian reference frame centered in O, that can be called e-frame
(for “external”) and to define the points A and C respectively in the propeller center and in the center of mass
of the drone, which should coincide approximately with the center of the drone frame.
Three positions vector could be introduced:
 Position of C relative to O, p
 Position of A relative to O, r
 Position of A relative to C, s
Furthermore, it is possible to put in evidence the following relationship between these vectors:
20
𝐫=𝐩+𝐬 (3.7)
Figure 3.15: quadrotor’s position vectors in the inertial frame [15]
Under a translational motion, there will be a variation of the position vectors r and p, but not of vector s
that will be constant.
Figure 3.16: translation of the quadrotor [15]
The velocity and acceleration of the quadrotor are the time derivative of vector p:
d𝐩(t)
𝐯P (t) ≜ 𝐩̇ (t) = dt
(3.8)
d𝐯(t) d2 𝐩(t)
𝐚P (t) ≜ 𝐯Ṗ (t) = p̈ (t) = dt
= dt
(3.9)
Now it will be interesting to see how the vector r change in time:
𝐯r = 𝐫̇ = 𝐩̇ + 𝐬̇ = 𝐩̇ = 𝐯p , 𝐯r = 𝐯P (3.10)
𝐚r = 𝐫̈ = 𝐯ṙ = 𝐯ṗ = 𝐚p , 𝐚r = 𝐚p (3.11)
Under a translation movement, it can be noticed that all points of the moving object; in this case, the drone
has the same velocity and acceleration.
Under a rotational motion, instead, is possible to observe that p(t) = 0, and a body reference frame, “b-
21
frame”, that rotates together with the quadrotor, around an axis passing through its center of mass, and the
position vector r defined by OA, which points to a fixed point A on the body frame that could be the propeller
center.
Figure 3.17: quadrotor’s body reference frame and inertial frame [15]
The goal is to represent the speed and acceleration of vector r.
The b-frame can be denoted with respect to the e-frame basis vectors through the following notation:
𝐢b = ieb,x 𝐢e + ieb,y 𝐣e + ieb,z 𝐤 e (3.12)
𝐣b = jeb,x 𝐢e + jeb,y 𝐣e + jeb,z 𝐤 e (3.13)
𝐤 b = k eb,x 𝐢e + k eb,y 𝐣e + k eb,z 𝐤 e (3.14)
It is possible to express vector r in the b-frame
𝐫 = rxb 𝐢b + ryb 𝐣b + rzb 𝐤 b (3.15)
and to refer this equation in the e-frame:
𝐫 e = rxb 𝐢eb + ryb 𝐣eb + rzb 𝐤 eb (3.16)
The vector r can be represented either in b-frame or in e-frame, and the passage between them is done by
multiply a rotation matrix R:
𝐫 e = 𝐑eb 𝐫 b (3.17)
Where:
22
rxe rxb ieb,x jeb,x k eb,x

e
𝐫 = [rye ] ,
e
𝐫 b = [ryb ] , 𝐑eb = [ib,y jeb,y k eb,y ] (3.18)
rze rzb ieb,z jeb,z k eb,z
The rotation matrix consists of 9 elements, but only three are required to specify a rotation/orientation in
3D space. Some alternative ways of representing the orientation in a more compact way instead of the Rotation
matrix are the Unit Quaternions (also known as “Euler parameters”) and the Euler angles.
Unit Quaternions are a set of 4 numbers (4D vectors). They are equivalent to rotation matrices, and they
are the most efficient and numerically stable tool to express orientations and rotations. For this reason, as it
will be shown in chapter 6, the Simulink quadcopter simulation model also has the unit quaternions
representation, provided by the Aerospace blockset add-on tool.
Quaternions are 4-dimensional vectors, and they are the combination of a 3-dimensional vector with a
scalar.
s v3 ]T
𝐪 = [ ] = [s v1 v2 (3.19)
𝐯
It is possible to express a rotation matrix as a function of a quaternion is the following way:
−𝐯 T
𝐫 e = [−𝐯 s𝐈3 + [𝐯]x ] [ ] 𝐫 b = 𝐑eb (𝐪)𝐫 b (3.20)
s𝐈3 + [𝐯]x
s 2 + v12 − v22 − v32 2v1 v2 − 2sv3 2v1 v3 + 2sv2

𝐑eb (𝐪)𝐫 b = [ 2v1 v2 + 2v3 s s − v12 + v22 − v32
2
−2sv1 + 2v2 v3 ] = (3.21)
2v1 v3 − 2sv2 2sv1 + 2v2 v3 s − v12 − v22 + v32
2
s 2 + v12 − 0.5 v1 v2 − sv3 v1 v3 + sv2

= 2 [ v1 v2 + v3 s s 2 + v22 − 0.5 −sv1 + v2 v3 ] (3.22)
v1 v3 − sv2 2sv1 + 2v2 v3 s 2 + v32 − 0.5
The derivative of the quaternion is the following:
−v1 ωbx − v2 ωby − v3 ωbz

1 −𝐯 T 1 sωbx − v3 ωby + v2 ωbz
𝐫̇ = [ ] ωb = (3.23)
2 s𝐈3 + [𝐯]x 2 v3 ωbx + sωby − v1 ωbz
b b b
[ −v2 ωx + v1 ωy + sωz ]
23
The other alternative way to represent orientation is through Euler Angles. They are a set of 3 numbers that
are equivalent to a rotation matrix, but they have an intrinsic disadvantage because they have singularities in
operation (divisions by zero in expressions), and these need to be treated as special cases.
The rotation of the frame can be split up in a series of three individual rotations around the coordinate
system axes, that is a product of three 2D rotations
𝐑eb = 𝐑 z (φ)𝐑 y (θ)𝐑 x (ψ) (3.24)
where:
cφ −sφ 0 cθ 0 sθ 1 0 0
𝐑 z (φ) = [sφ cφ 0], 𝐑 y (θ) = [ 0 1 0 ], 𝐑 x (ψ) = [0 cψ −sψ] (3.25)
0 0 1 −sθ 0 cθ 0 sψ cψ
The relation between the rotation matrix and the Euler angles can be expressed in the following way:
cθcφ sψsθcφ − cψsφ cψsθcφ + sψsφ

𝐑eb = [cθsφ sψsθsφ − cψcφ cψsθsφ − sψcφ] (3.26)
−sθ sψcθ cψcθ
ψ is the roll angle, φ is the yaw angle, and θ is the pitch angle, and. This intrinsic sequence is also known as
Cardan angles.
Figure 3.18: Euler angles and rotations [15]
In total, there are 12 extrinsic combinations and 12 intrinsic combinations.
As far as the derivation of velocity and acceleration for rotational motion is concerned, it can be observed
that rb is constant (because the b-frame rotates with the drone) and re is changing with time. If the derivative
of a rotation matrix is used, the linear velocity in the e-frame is written as:
𝐯 e (t) = 𝐑eḃ (t)𝐫 b (t) (3.27)
Instead, the derivative of the rotation matrix can be denoted as:

24
𝐑eḃ (t) = Ωe 𝐑eb (t) = 𝐑eb (t)Ωb (3.28)
Matrices Ω are skew-symmetric, and they have the following form:
ωx 0 −ωz ωy
𝛚 = [ωy ] 𝛀 = [ ωz 0 −ωx ] = [𝛚]x (3.29)
ωz −ωy ωx 0
The multiplication by a skew-symmetric matrix is another way to represent the vector cross-product. The
vector ω is the angular velocity, which is a physical vector.
The derivative of the rotation matrix is expressed as:
𝐑eḃ (t) = [𝛚e ]x 𝐑eb (t) = 𝐑eb (t)[𝛚b ]x with 𝛚e = 𝐑eb 𝛚b (3.30)
The angular rate in the rotating body frame ωb can be obtained from the measurements of a MEMS
gyroscope sensor placed on the b-frame.
The linear velocity vector is written with the following notation:
𝐯 e = [𝛚e ]x 𝐫 e (3.31)
The angular acceleration α is the derivative of the angular velocity vector ω, and its unit is [rad/s].
The linear acceleration in the e-frame ae can be written as:
𝐚e = [𝛂e ]x 𝐫 e + [𝛚ⅇ ]x [𝛚e ]x 𝐫 e (3.32)
Instead, ab is the output of the quadrotor’s accelerometer sensor:
𝐚b = 𝐑be 𝐚e (3.33)
Under a roto-translational motion, we have the two motions combined, which is the generalized case, and
the most important in the drone application.

25
Figure 3.19: quadrotor’s body reference frame and inertial frame [15]
Figure 3.20: roto translation of the quadrotor [15]
The following position vector is obtained, combining translation and rotation and translation motion:
𝐫 e = 𝐩e + 𝐬 e = 𝐩e + 𝐑eb 𝐬 b (3.34)
Moreover, the following linear velocity of point r relative to e-frame is expressed as
𝐕re = 𝐕pe + 𝐑eb [ωb ]x 𝐬b = 𝐕pe + [𝛚e ]x 𝐬e (3.35)
The linear acceleration of point r is obtained deriving the linear velocity:
𝐚er = 𝐚ep + [𝛂b ]x 𝐬e + [𝛚ⅇ ]x [𝛚e ]x 𝐬 e (3.36)
3.2.2 Dynamics
The following laws are considered for the derivation of the equations describing the dynamics of the
quadrotors.
 Newton’s laws of motion for point masses/particles
 Euler’s laws of motion for the rigid body (translational and rotational equation of motion)
Among the first ones the Second law tells that the sum of forces on a particle object is equal to the mass of
26
the object times the acceleration of the object in an inertial reference frame (in this project the hypothesis is
that the earth is like as an inertial frame)

e
fk,x ax p̈ ex
e e
ftotal = ∑k fke = m𝐚e = m𝐩̈ e ⇔ ∑ [fk,y ] = m [ay ] = [p̈ ey ] (3.37)
e
fk,z aez p̈ ez
k
According to the Third principle of Dynamics, instead, when a particle applies a force on a second particle
upon some form of interaction, contact, or at-a-distance), the second particle simultaneously applies a force
equal in magnitude, but opposite, onto the first particle.
The direction of the two forces is along the straight line joining the point masses. If i and j are two particles,
fij with which particle i acts upon particle j, and ri and rj are position vectors then:
𝐟ij = −𝐟ji (3.38)
𝐟ij = ±‖𝐟ij ‖(𝐫i − 𝐫j ) (3.39)
If a system of particles is considered, with fixed distance one each other, that is a rigid body, the Euler’s
laws of motion tell the dynamics behavior of this system. The resultant of the internal forces applied to these
particles is, therefore, null and the only nonzero forces are the external ones, that can be thought to act on the
center of gravity of the object. The resultant of external forces equals the center of gravity acceleration
multiplied by the total mass of the system.

1
𝐩 = m ∑i mi 𝐫𝐢 (3.40)
e
𝐟ext,total = m𝐩̈ e = m𝐚ep (3.41)
Here r is the center of the gravity position vector, and m is the body mass. Forces not only push or pull a
rigid body (translation) but also tend to rotate it, and the torque expresses this effect.
27
Figure 3.21: torque description [15]
𝛕O = [𝐬]x 𝐟 = ‖s‖‖f‖ sin(θ)𝐧 = ‖f‖ ⋅ d ⋅ 𝐧 (3.42)
If a quadrotor is considered like in the following picture, the external torque about O can be expressed in the
following way:
Figure 3.22: external force on the quadrotor [15]
𝛕O,ext = [𝐩]x 𝐟total,ext + τC,ext (3.43)
Furthermore, as observed for the resultant of the forces on a rigid body, the torque expression has also only
the external component, because the sum of internal torque is zero.
Combining the last equation with the (3.41), the following equation is obtained:
𝛕e0 = ∑i mi [𝐫ie ]x 𝐚ei (3.44)
e
𝛕eO − [𝐩e ]𝐟total,ext = 𝐉 e αe + [𝛚e ]x 𝐉 e 𝛚e (3.45)
And combining equation (3.43) with (3.45), we can notice that the resultant of Torques about C (center of
mass of the quadrotor) is a sum of two components as the following expression shows:
𝛕eC = 𝐉 e αe + [𝛚e ]x 𝐉e 𝛚e (3.46)
Where:
28
𝐉 e = − ∑i mi [𝐬ie ]x [𝐬ie ]x (3.47)
that is the global inertia matrix.
The rotational equation of motion can also be represented in body-frame coordinates:
𝛕bc = 𝐉 b 𝛂b + [𝛚b ]x 𝐉b 𝛚b (3.48)
And the global inertia matrix is represented in relation to the body inertia matrix as:
𝐉 b = 𝐑be 𝐉 e 𝐑eb (3.49)
0 −zi yi 0 −zi yi 𝑥𝑖
b
𝐉 = −∑ mi [𝐬ib]x [𝐬ib ]x ∑ z
= − i mi [ i 0 −xi ] [ zi 0 −xi ] with 𝒔𝑏𝑖 𝑦ⅈ
=[ ] (3.50)
i −yi xi 0 −yi xi 0 𝑧𝑖
yi2 + zi2 −xi yi −xi zi

𝐉 b = ∑i mi [ −xi yi xi2 + zi2 −yi zi ] (3.51)
−xi zi −yi zi xi2 + yi2
∫V(yi2 + zi2 ) dm ∫V −xi yi dm ∫V −xi zi dm

b
𝐉 = [ ∫V −xi yi dm ∫V(xi2 + zi2 ) ∫V −yi zi dm] (3.52)
∫V −xi zi dm ∫V −yi zi dm ∫V(xi2 + yi2 )
Furthermore, it can be noticed that while 𝒔𝒃𝒆 is variable in time, vector 𝒔𝒃𝒊 is constant, meaning that Je is
time-dependent, while Jb in the body frame is time constant.
The differential equations of motion can be summarized and put together to form the dynamical model of
the quadrotor, and they can be written either with the rotation matrix notation:
𝐩̇ e = 𝐯 e (3.53)
1 e 1
𝐯̇ e = m 𝐟total,ext = m 𝐑eb 𝐟total,ext
b
(3.54)
𝐑̇eb = 𝐑eb [𝛚b ]x (3.55a)
−1
𝛚̇b = (𝐉 b ) (−[ωb ]x 𝐉b 𝛚b + τbc ) (3.56)
Alternatively, with the Unit Quaternion notation, having the equation (3.55a) been replaced by the
following one:
29
−v1 ωbx − v2 ωby − v3 ωbz

𝑠̇ 1 −𝐯 T 1 sωbx − v3 ωby + v2 ωbz
𝒒̇ = [ ] = 2 [ ] ωb = 2 (3.55b)
𝒗̇ s𝐈3 + [𝐯]x v3 ωbx + sωby − v1 ωbz
b b b
[ −v2 ωx + v1 ωy + sωz ]
3.2.3 Torques and forces generation
The following schematic (figure 3.23) shows the part of the system which involves motors and propellers,
and, in this paragraph, the mechanism of the force generation on the propeller from the motor’s torque will
be analyzed.
Figure 3.23: Actuators’ model [15]
Firstly, the code on electronics (firmware/embedded code) sends a digital command to the motors, and then
it is transformed into an analog PWM (Pulse Width Modulation) signal that commands the coreless motors of
the quadrotor. This digital command consists of four 16-bit integer numbers (from 0 to 65535), one for each
motor.
The physical principles that regulate the movement of the quadrotor (the same principles are valid for every
aircraft) are the following ones:
 Bernoulli's theorem: Daniel Bernoulli experimented that pressure inside a fluid, liquid, or gas, decreases
as much as the speed of the fluid increases, in other words: "in a moving fluid, the sum of pressure and
speed in any point remains constant. "
 Venturi effect: Giovanni Battista Venturi proved experimentally that a fluid particle, passing through a
narrowing, increases its speed.
 3rd Newton's Law: there is always an equal and opposite reaction to each corresponding action.
The concept of lift can summarize the previous principles applied to an apparatus with an airfoil, for
30
example, a wing or a propeller. If it moves in the air (which has its atmospheric pressure and speed), at a
certain speed with a specification upward inclination (called the angle of attack), it will pass through an airflow
splitting this into two flows. The one running along the top of the profile will do it with a higher speed than
the other, which flows through the lower part (Venturi effect). Moreover, higher speed implies lower pressure
(Bernoulli's theorem). Therefore, the surface of the upper wing is subjected to a lower pressure than that
inferior.
The airfoil is typically obtained from a body that has a specially designed shape to obtain most of the force
generated by the speed and pressure variation of the fluid when it is flowing in an air stream.
Figure 3.24: propeller model and air propulsion [17]

CHAPTER 4
HARDWARE DESCRIPTION AND FIRMWARE SETUP
4.1 Hardware and instrumentation
The mini drone that was used in this project is the " Mambo " model, produced by the French
company Parrot. This model can be considered a low-cost drone, considering its cost, from 40 to 80 dollars,
especially if compared to other models from the same company that can also reach up to 500 dollars.
Figure 4.1: PARROT Mambo Fly minidrone [20] and [22]
The other instruments are the following.
 Safety goggles ensure safety during the flight.
 micro-USB cable used both for charging the mini-drone and for uploading the firmware released from
Mathworks to the drone.
 USB dongle compatible with Bluetooth Low Energy used to deploy the model from Simulink to the
hardware and make it run in real-time.
 USB dongle drivers usually provided on a CD included with the dongle.
 Additional batteries to be optionally purchased besides the one included with the mini-drone, due to
their short duration (about 8-10 minutes of flight)
 Charger.
It is possible to run the model alternatively also on a PARROT Rolling Spider mini-drone model
because it is also compatible with the Simulink Hardware support package.
31
32
Figure 4.2: Safety goggles (left) [24], Bluetooth dongle adapter (center) [25], and USB cable (right) [26]
The environment is made by background and a path that must be followed by the mini drone. For the
background, a non-glossy backdrop, like the ones of a photo studio, can be put on the ground. White tape was
used to create the track such that there is more contrast with the background, and in addition to this, some
pieces of paper or other casual objects were randomly put on the background to improve the Optical flow
estimation.
Figure 4.3: black background (left and center) [42] and white tape for the track (right) [43]
4.1.1 PARROT mini drones
PARROT mini drones are tiny and cheap quadrotors (4 motors), which can be commanded through a
smartphone or a tablet. They are considered among the most stable quadcopters available to buy, because of
autopilot on their original firmware. However, this will not be used in the project because the new firmware
supports a customized controller in the Simulink model. They are also very safe, and they can also fly indoor
thanks to the “cut-out” system, which activates in case of impact and shuts down suddenly the motor. They
are provided with a 3-axis gyroscope and a 3-axis accelerometer, a down-facing camera useful for optical flow
estimation and image processing, and a SONAR sensor and a pressure sensor for altitude measurements.
4.1.2 Technical characteristics of MAMBO model and main components
Some of the main technical characteristics [20] will be shown below:

33
 Dimensions: 7.1 x 7.1 in.
 Weight: 2.22 oz
 Energy:
- 660 mAh LiPo battery
- 8 minutes of autonomy with accessories
- 30 minutes of charging with a 2.1 Ampere charger
- 10 minutes of autonomy without accessories
 IMU sensors:
- to measure speed and accelerations. The creation of a sensory system capable of perceiving the
movement of an object in the surrounding environment is a relevant problem in mobile robotics.
Control algorithms are designed from these sensory data, and their availability and reliability are
essential requirements. An inertial navigation system (INS) is a system capable of acquiring the object
navigation information (linear and angular positions and speeds) in its body frame and of transforming
it into the fixed (inertial) reference frame.
 Vertical Stabilization:
- Ultrasound sensor: SONAR (sound navigation & ranging) systems are used in mobile robotics as a
low-cost solution for measuring the distance between the robot and the surrounding environment (in
the case of this project to measure the distance between the drone and the ground). Their working
principle is based on the calculation of the flight time of an ultrasound wave to travel the distance back
and forth from the sensor: they are both signal generators and transducers. The quadcopter uses it to
measure altitude. It emits a high-frequency sound wave, and it measures the time that the wave takes
to reflect on the ground and to be received back by the sensor. From the measured time, the distance
between the floor and the drone can be calculated. The maximum altitude that can be estimated is
about 13 ft.
34
- Pressure sensor: it is an aid to the ultrasonic sensor for calculating the altitude of the drone. As the
drone increases the altitude, the air pressure decreases. The pressure variation can be used to estimate
the change in altitude.
 Horizontal Stabilization:
- Camera sensor
 Speed Measurement:
- 60 FPS (Frame per second) vertical camera
 SDK (OS of the Firmware):
- OS Linux
Figure 4.4: top (right) and bottom (left) view of PARROT Mambo Fly model [20]
On the front of the mini drone, there are the two indicator lights that change color depending on the state
of the drone:
 Steady orange: The Parrot Mambo is starting up.
 Steady green: Parrot Mambo is ready to fly.
 Steady red: Parrot Mambo has detected a problem.
 Blinking red: The Parrot Mambo battery is running out.
In the upper part of the drone, there is a connector for any external accessories, such as the FPVfirst-person
view camera and the clamps.
It is also possible to protect the propellers by installing the hulls on the appropriate supports.
35
In the rear part, there is the place that houses the battery, the micro-USB port, together with an LED
indicator that indicates the charging status (Red: charging; Green: 100% battery). With the same USB port, it
is possible to communicate the PCB to the computer.
In addition to the on and off button, below the drone, there are several sensors:
 Pressure sensor (black protrusion at the tip of the drone).
 Ultrasonic sensor (Sonar).
 60 FPS camera: It uses an image processing technique called optical flow to determine the change in
shape or translation of objects from a frame and the next one. From these modifications in the position
of objects and shapes, the drone can recognize horizontal motion and compute speed.
In figure 4.5, the drone motherboard is shown. It is equipped with the latest SiP6 chipset Parrot with an 800
MHz ARM A9 (in the gray box of figure 4.5 on the right), which has a size of about 12 × 12 mm. This
motherboard runs on Linux and is equipped with a pressure sensor, accelerometer, gyroscope, and ultrasonic
sensor. It also includes a 60 frame-per-second vertical camera.
Figure 4.5: motherboard of Mambo fly model [27]
The 6-axis motion detection device (Figure 4.6) is an MPU-6050 from the US manufacturer
InvenSense. Inside this, there is a Digital Motion Processor (DMP) interacting with a three-axis accelerometer
and a three-axis gyroscope. The device dimensions are about 0.16 × 0.16 × 0.035 inches. This device is
mounted just above the processor (in the blue box of Figure 4.6 on the left).
36
Figure 4.6: IMU with gyroscope and accelerometer [28] and [29]
The axes of the motion sensor (Figure 4.7) are not concordant with the axes of movement of the drone
(figure 5.32 in the next chapter).
Figure 4.7: IMU reference frame [29]
To conclude this brief overview of the components of the Parrot Mambo, the motors and their propellers
are shown (Figure 4.8). An important feature to classify these motors is their spin direction.
It can also be noticed that the drone propellers are of two different colors: the white ones are the ones
mounted in the front while the black ones are in the back.
Figure 4.8: propellers (left) and coreless motors (right) [23]
The drone has two different types of motors (Figure 4.8 on the right):
 Motor A (Counterclockwise) is identified with a black and white wire together with a circular sign black
printed on the upper part (where the rotation axis comes out), and it is located on the front right and rear
left parts.
37
 Motor C (Clockwise) is identified with a red and black wire, and it is located on the front left and right
rear parts.
Both motors are Coreless type and dimensions of 8.5 × 20 mm. Similarly, there are two types of propellers,
identifiable by the same word "C" (clockwise) and "A" (counterclockwise):
 Front left and rear right: "C”.
 Front right and rear left: "A".
These plastic propellers have been designed to minimize the electrical consumption of the engines and, at
the same time, ensure maximum thrust to the engines.
To better understand what a Coreless motor is, it is useful to briefly describe a typical DC motor with a
collector.
Figure 4.9: DC motor structure and components [30]
It consists of several main parts:
 Stator: typically composed of a permanent magnet.
 Rotor: composed of a set of ferromagnetic plates, in the shape of a circular crown where the various
copper windings are wrapped.
 Collector: it is a set of blades fixed on a cylindrical drum to the coil terminals.
 Brushes: made of graphite. They remain in contact with the collector even when it is in rotation, and they
give the main power to the motor.

38
In Coreless motors (Figure 4.10 on the left), there are all the components of the previous motor, except for
the core of laminations in the rotor. The approach used to describe them is adapted from Karnavas et al. [22].
Figure 4.10: Coreless DC motor structure [30] and rotor winding [29]
The rotor winding (Figure 4.10 on the right) is wound obliquely, or honeycomb, to form a self-supporting
hollow cylinder. Since there is no rigid structure that supports it, the windings are covered with epoxy resin.
The stator which generates a magnetic field is located inside the rotor and is composed of rare earth
magnets, such as Neodymium, or Al-Ni-Co (Aluminum, nickel, cobalt), or Sm-Co (Samarium, cobalt).
The Coreless motor usually has tiny dimensions: between 6 millimeters to less than 75 millimeters in
diameter. Their power limit, in general, is around 250 watts or even less.
The following properties make Coreless motors valid solutions for different applications:
 Coreless rotor technology ensures smooth operation (regular dynamics)
 The rotor has lower inertia than iron-core DC motors, and this leads to more significant accelerations.
 The absence of iron losses implies higher efficiency (about 85% compared to 50% of the iron-core
ones), longer life, and less overheating.
 Lighter than iron-core DC motors.
 At low speeds, the starting voltage is lower (0.3 V).
 Faster dynamical response (low mechanical time constant, about 10 ms.
 Longer life of the brushes and the collector because of a reduced inductance.
39
4.2 SIMULINK Hardware support package and firmware setup
In this paragraph, the procedure to install the firmware on the drone from the Hardware support package of
Matlab will be explained. The firmware released by Mathworks makes it possible to deploy the model from
the Matlab project to the drone board and run it through Bluetooth connection.
The material needed for this operation is the following:
 Parrot Mambo drone (with the charged battery inserted).
 USB- micro-USB cable.
 Bluetooth USB Dongle.
 Bluetooth USB CSR driver already installed.
 Add-ons: “Simulink Support Package for Parrot Minidrones” installed.
The following method explains how to make the drone communicate with the PC via Bluetooth in a
Windows 10 operating system. From the main MATLAB window, open the Add-Ons Manager (Manage Add-
Ons) and start the procedure by clicking on the gear icon. Connect the drone with the USB cable (Figure 4.11)
and wait a few moments for it to be recognized by the PC. Click Next and wait for the recognition of the drone.
Press Next again, the firmware file will be loaded into the drone's memory.
Figure 4.11: recognition of the drone through the USB cable [12]
It is necessary to disconnect the USB cable from the drone to start the setup procedure (figure 4.12). During
40
installation, the front LEDs will flash orange. Once completed, the LEDs will flash green (make sure that for
at least 10 seconds, this last state occurs to proceed with the Setup).
Figure 4.12: Firmware setup [12]
Reconnect the USB cable to the drone (wait a few moments for it to be recognized) and proceed by
clicking Next (Figure 4.13).
Figure 4.13: RNDIS driver setup (1) [21]
Before doing Next, the following procedure must be completed to install the RNDIS driver (figure 4.14).
Open “Device Manager” and click on “Other devices”. From the list, select “RNDIS” and “Update Driver
Software...”.
41
In the new screen, the driver software must be picked from the list of devices. Among the “Common
hardware types”, select “Network adapters”.
Look for “Microsoft” among the Manufacturers, “Remote NDIS Compatible Device” among the Models
(figure 4.16). In the “Update Driver Warning” dialog box, select “Yes”. Once installed, the RNDIS device can
be found among the devices inside “Network adapters”. Now, it is possible to continue with the Setup
procedure.
By clicking on Next (Figure 4.16), a new window shows up, where there are instructions to connect the
drone via Bluetooth. First, you must disconnect the USB cable from the drone again.
42
Figure 4.16: firmware checking (left) and Bluetooth radio device setup (center and right) [20] and [44]
Now, from the file explorer window, click on "This PC" and then double click on "My Bluetooth
devices". Then, open the "Add device" dropdown and select "All".
The search for Bluetooth devices starts. The name of the device to be connected is called "Mambo”
followed by several numbers (relative to the hardware serial number). It is necessary to choose the one that
has the icon of a Joystick (figure 4.17); otherwise, the communication between MATLAB and the drone will
not be possible.
Figure 4.17: Bluetooth pairing [44]
Once pairing is complete, double click on the device just added (Figure 4.19), and after pressing with the
right text above the icon accompanied by name Personal Area Networking (NAP), click Connect.
43
Figure 4.18: PAN connection setup [44]
Now press Next (Figure 4.19), and in the next screen, click on "Test connection" to ping to verify that the
communication occurs.
Figure 4.19: Bluetooth connection test [12]
Once completed and successful, click Next, and then Finish to end the Setup. This procedure only needs to
be done once. If we want to connect the drone other times, it must be done via Bluetooth with PAN (Personal
Area Network) in the services of the device, keeping the Bluetooth dongle plugged in the PC and turned on.
CHAPTER 5
SOFTWARE TOOLS DESCRIPTION AND MODEL CONFIGURATION
5.1 Description of Software Tools used in the MATLAB environment
For this Thesis project, the leading software and tools used are MATLAB 2019b, Simulink, Stateflow, and
some of their add-on tools like Aerospace Blockset, Simulink 3D Animation, which lets us visualize
simulations in 3D, the Simulink Hardware Support Package for PARROT Minidrones, and Simulink Coder.
Simulink is an extension of Matlab that allows us to build control systems models with a graphical
representation through blocks, to simulate them, and to adopt Model-Based Design technique. It provides
automatic code generation, and hardware test and validation. It also offers solvers for dynamical systems and
customizable block libraries.
MathWorks developed the Simulink support package for PARROT mini drones for various purposes. One
is to make people interested and aware of the influence of Model-Based Design in modern engineering
applications. Another reason is to assist developers in the industry in acquiring Model-Based Design
techniques through an innovative application. Finally, to support instructors and professors in training students
on this design method through drones as an admired hardware platform [18].
The add-on tools and toolboxes used in Matlab 2019b environment are the following:
 Simulink Hardware Support Package for Parrot Minidrones: used for firmware setup, for the connection
of the drone via Bluetooth, and algorithm deployment through MATLAB. It allows us to employ different
onboard sensors for creation, simulation, and test of flight control algorithms. Through this package, we
can have a low-cost and ultra-compact drone to conduct experiments with its control system.
 Aerospace Blockset: for modeling, simulation, and analysis of aerospace vehicles. It provides blocks for
44
45
modeling and simulation of crewless airborne vehicles, propulsion systems, aircraft, and spacecraft, that
are subjected to environmental and atmospheric conditions.
 Optimization Toolbox: a toolbox for solving linear optimization problems, quadratic, integer, and
nonlinear.
 Signal Processing Toolbox: for the processing and analysis of signals.
 Simulink 3D Animation: it allows us to visualize the behavior of dynamic systems in a virtual reality
environment.
 Simulink Coder and Matlab Coder: C++ automatic code generation from Simulink models and execute
them. We can use the generated source code in non-real-time and real-time applications like rapid
prototyping, simulation, and HIL (hardware-in-the-loop) tests. It can also be used to record flight data.
 Embedded Coder: C ++ code generation for embedded processors (Extension of Simulink Coder).
 Simulink Control Design: model linearization and control systems development.
 Image Processing Toolbox: to use the App Color Threshold.
 DSP System Toolbox: Design/Simulation of signal processing systems.
5.2 Software configuration, Simulink project overview, and Simulation Model description
The project should be configured to work on the model design, and for this purpose, the project
environment of the Mathworks mini-drone competition organized by Mathworks can be used by entering the
command window the word “ParrotMinidroneCompetitionStart”.
Firstly, the project environment must be configured, and once done it, three windows appear:
 Tree model with all the subsystems and the variables (figures 5.2) of the project named
“ParrotMinidroneCompetition”.
 Mini Drone simulation model (figure 5.1).
 Simulink 3D Animation: visualization of the drone dynamics during the simulation (figures 5.3)
46
Figure 5.1: The Simulation model

47
Figure 5.2: workspace with the variables and constants (left and center) and project folders (right)
Figure 5.3: Simulink 3D environment used for simulation
Within the project framework, we can organize user-defined tasks related to this competition, files, and
settings. The simulation model (figure 5.1) is constituted by six main blocks that embrace the mathematical
descriptions of the relative dynamical subsystem: the airframe, the environment, the flight control system
(FCS), the sensors, and the subsystems used to show a visualization output or to feed input commands during
the simulation.
Four of the six subsystems can be defined as variant subsystems because they allow us to choose between
different properties of the subsystem. However, the FCS (Flight Control System) block (figure 5.4) should not
be considered as a variant subsystem, but as a modeled subsystem, because its elements are referred to another
Simulink model. The FCS subsystem will be explained more deeply in the next chapter.
48
Figure 5.4: Flight Control system
Furthermore, other blocks in the Simulation model are employed to set the simulation pace. Instead, the
“Flag” subsystem is used to terminate the simulation if an unwanted flight situation happens (figures 5.5, 5.6,
and 5.8).
Figures 5.5: Flag block
Figure 5.6: Flag subsystem (1)
The “play” icon must be pressed to run the simulation. The system dynamic behavior can be observed for
the time defined through “TFinal” variable, and then can be stopped (figure 5.7 on the left). The “set pace”
option can be used to modify the speed of the simulation (figure 5.7 on the right).
Figure 5.7: Simulation commands (left) and pace settings (right)
For example, the simulation can be run tenth times slower, and the sample time can be increased to 0.1 ms
to see the mini drone moving at slow-motion. Once it runs, the mini drone 3D model takes off and hovers.
49
Figure 5.8: Flag subsystem (2)

50
Other options can be found by double-clicking on the “Flight visualization” subsystem. The signals can be
output as shown on a cockpit display, provided by Aerospace Blockset toolbox, with some standard flight
instruments like heading, percent RPM, airspeed and climb-rate indicators, and altimeter. The “extract flight
instruments” (figure 5.9) subsystem contains the applicable states.
Figures 5.9: Simulink Visualization block
Figure 5.10: Simulink Visualization subsystem
The “Command” subsystem can be used to change the mini drone input signals in the simulation (figure
5.11). Four variant subsystems constitute this one: data input, Signal builder, joystick, and spreadsheet file
reading.
Figure 5.11: Commands subsystem

51
Signal builder is set as default in the command subsystem, but, in this project, the signals driving the drone
come from the path planning subsystem in the Flight controller, which receives the data from the image
processing subsystems. The mode selected to drive the drone is according to position coordinates X-Y-Z and
yaw, but alternatively, it can be driven by pitch and roll signals coming from the Image processing subsystem.
However, in this last case, it would become more complex because the camera orientation has to be ensured
to point right to the track, and this would be easily solved by using position coordinates rather than pitch and
roll angles. In the Model-Based Design method, the simulation model is the focal point, because it aids to
improve the control development before deploying it on the hardware, therefore avoiding damages and crashes.
A part of the FCS subsystem is adapted from the model of Professor Sertac Karaman, and Fabian Riether
developed at MIT. Since the project aims to design a flight controller able to track the path while keeping
stable itself, the controller subsystem is designed tuning the gains of the PID controllers, and the image
processing and path planning subsystems are designed according to our project goal.
Finally, once the FCS is developed and tuned through simulations, the code can be generated and tested on
the mini-drone hardware.
5.3 Compiler configuration
To generate the C ++ code to be loaded into the aircraft, the exact compiler must be chosen. What by default
it does not allow the download of MAT files from the drone, after each flight simulation.
This procedure can only be performed from the model window opened automatically after generating the
code with the "Generate Code" button. Go to the "Modeling" bar (Figure 5.12) and click on the gear icon. It
will open a window, shown in Figure 5.13.
Figure 5.12: Modeling bar [12]

52
Figure 5.13: Model settings [12]
Click on "Hardware Implementation", then move the mouse over the three dots to bring up the word
"Advanced Parameters", click on it to show the menu, then tick the "Use Simulink Coder features" option.
Complete the configuration by pressing "OK".
5.4 Preliminary Test of the drone motors
Before performing any flight tests, it is possible to run a test in which the propellers are only moved without
generating the minimum thrust to flight (figures 5.14). In this test, the model spins the motors 1 and 3 for 2
seconds and then the motors 2 and 4 for the other 2 seconds. The test model is a template included in the
package, and it can be set up by entering the command “parrot_gettingstarted” in Matlab and by selecting
“Deploy to Hardware” (figure 5.16).
Figure 5.14: FCS subsystem for the actuators test

53
Figure 5.15: Hardware functionalities
The window called “diagnostic viewer” (figure 5.16) allows us to follow and to troubleshoot the
compilation, the file transfer, the code generation (figure 5.17), and the execution. There is another way the
model can be executed: the external mode (“Monitor and Tune”). Through this execution method, some values
can be updated from the block diagrams throughout its execution, besides building and deploying the model
on the hardware. It is particularly useful during the test of the image processing algorithm because the real-
time visualization of the images acquired from the downward-facing camera allows us to tune the threshold
filter for path detection.
Figure 5.16: Code generation [12]
Figure 5.17: C code of the model [12]

54
After the “diagnostic viewer” confirms that code generation has completed, the model can be run on the
hardware by opening the flight control user interface from the diagnostic viewer, and the flight time can also
be set as the simulation time (figure 5.18). At this point, to prevent an uncontrolled flight, we should be careful
not to increase the power gain of the thrust of the propellers (it is recommended to begin by setting the gain at
the 10% of the rated one).
In this test, signals are fed to the motors to ensure that the Bluetooth communication and the toolchain are
working correctly.
Figure 5.18: Flight control interface with flight time settings, stop and end of the test [12]
If we click on “Stop”, the test can be interrupted, and the flight log and the MAT file with sensor data can
be downloaded to the working directory.
It is useful to remember that the hardware target must be set up correctly (in our case “Mambo” model)
before deploying the code, and data logging must be enabled to take advantage of Simulink Coder capabilities
to capture sensor and controller commands signals for inspection. They can be used to plot trajectory, motor
output, sensor signals, altitude, velocities, orientation, position, optical flow velocities, and percentage of
55
battery charge. The data that can be stored by the mini drone during flight test is limited, and memory allocation
can be controlled through the flight time variable “TFinal”. The model cannot be executed if there is an excess
of memory needed beyond the hardware capacity.
Furthermore, it is possible to list other important guidelines:
 The space where flight tests are conducted must be large at least 20 feet by 20 feet by 10 feet high to
avoid damages to the mini drone, the environment, and any observer.
 Small rooms and specific ground materials (like a carpet) could cause flight stability issues because of
bouncing or absorption of ultrasound signals.
 Optical flow estimation algorithm, image processing algorithm, and stability during flight can be
affected by lighting conditions and shapes on the ground.
5.5 Physical characteristics of the Quadcopter model from Aerospace Blockset
The following picture (5.19) shows the quadcopter model with its body reference frame, taken from
Aerospace Blockset.
Figure 5.19: reference frame of Quadcopter model [31]
The origin of the mini drone reference frame coincides with its center of mass. “VehicleVars”
folder contains the mass and inertias in the Simulink project.
Body Frame:
 The x-axis is oriented in the forward direction.
 The y-axis is oriented to his right.
 The z-axis is oriented down.

56
Motors directions convention:
 Motor 1 spins positively relative to the z-axis, and its position is at -45° from the x-axis.
 Motor 2 spins negatively relative to the z-axis, and its position is at -135° from the x-axis.
 Motor 3 spins as motor 1, and its position is at 135° from the x-axis.
 Motor 4 spins as motor 2, and its position is at 45° from the x-axis.
The approach used to define the convention is the same as expressed by Prouty [32] and by Ponds et al. [33].
CHAPTER 6
SIMULATION MODEL DESCRIPTION AND CONTROL PROBLEM SETUP
6.1 Nonlinear and Linear models in Model-based design approach
In the design flow, nonlinear and linear models are both needed for designing and testing the flight control
software. The following method is the one treated by Douglas [41].
In figure 6.1, the upper block is the flight control software that represents all the control system
software. This code must interface with the rest of the mini-drone firmware, and it has two inputs, the raw
sensor readings and the reference commands or set points, and two outputs, the motor speed commands, and
the stop flag. The reference commands are put inside the flight code. The lower block called ‘Nonlinear model’
represents everything else, anything that is not the flight control code. This includes the rest of the mini-drone
firmware, the hardware, the atmosphere flying in, etc. At an elementary level, the model inputs motors
commands, and the stop flag, then it makes a few calculations and outputs sensor measurements. In this way,
the model is wrapping around the flight code and provides the feedback loop. If the model were so accurate
that it perfectly represents reality, it would be indistinguishable whether the results came from the actual
hardware or they came from this perfect model. Therefore, the mini drone performance could be simulated
using the model, and we can be very confident that when the flight code is run after on the actual hardware, it
would have the same result.
However, a perfect model of reality is impossible to create, and it is not necessary to model everything. The
trick is to figure out what to include in the model and what to leave out. Some of that knowledge comes easily
by just understanding this system and how it will be operated. For example, the code that turns on and off the
front LEDs does not need to be modeled. They will not impact the control system.
Nevertheless, there are a lot of other things that are not as obvious and knowing what to model requires a
57
58
little experience and investigation. One example is whether to model the airframe structure as a rigid body or
as a flexible body. It is hard to precisely know what to model and what to leave out initially. Usually, what
happens is starting with the best guess, and then over time, the fidelity of the model will grow until the match
between the experimental results and simulation are satisfying. Therefore, simulation is a way to verify the
system in hard situations or time-consuming to physically test, as long as the model adequately reflects reality.
The excellent advantage of simulation is that the model can be reset quickly, and the drone can be put in any
situation of interest, and if it performs poorly, we make the necessary changes, and we do not damage any of
the hardware in the process.
However, a model is also used for control system design, also with the linear analysis tools. Unfortunately,
the previously created nonlinear model used for simulation does not lend itself to a suitable design and linear
analysis. Thus, we also need a linear model. Essentially, we should remove the nonlinear components in the
model or estimate them as linear components. The linear model will not reflect reality as accurately as the
nonlinear model, but it should still be accurate enough that it can be used to design the controllers.
Summarizing, two different models are considered: a lower fidelity linear model that is useful for
determining the controller structure and gains and a higher fidelity nonlinear model that is useful for simulating
the result and verifying the system. The approach used in the method description
Figure 6.1: Schematics of Feedback Control loop with a nonlinear model for simulation (left) and with a
linear model for controller design (right) [41]

59
To summarize, the steps taken to design the flight control software using Model-Based Design techniques
are the following:
1. Create a high-fidelity model of everything the flight control software needs to interact with. This very
likely will be a nonlinear model.
2. Verify that the model matches reality with several test cases.
3. Once had a model that reflects reality, create a linear version.
4. The linear model and our linear analysis tools are used to design and analyze our control system.
5. The nonlinear model is used to verify the performance of the control system.
6. The flight control software can be run with enough confidence in the actual hardware for final
verification.
Figure 6.2: Model-based design technique schematic [40]
Instead of thinking about the plant model as a single block or as a monolithic set of calculations, it is
generally easier to break it up into several smaller models that represent specific and discrete systems.
For the mini drone, it might be broken up into the airframe structure, the actuators, the environment, and
the sensors (figure 6.3). Then, within these models, there are even smaller subsystem models, like the gravity
model or the IMU model.

60
Figure 6.3: Components of the airframe model [40]
There are many reasons to approach modeling in this hierarchical way, rather than lumping all the plant
components together into a single model:
 For instance, it allows multiple people and teams to build different parts of the model simultaneously.
 Portions of the model can be upgraded, based on which area needs more fidelity, without impacting the
rest.
 Each sub-model can be developed using different modeling techniques. It can be chosen as the modeling
technique that makes the most sense or is the easiest for each system. Then when they are put
all together, there will be an entire model that can be used for wrapping around the flight control
software and simulating the results.
6.2 Control system architecture (Hovering control)
In this chapter, it will be described how to design a control system architecture to control the quadcopter
during the flight maneuvers needed to accomplish the task. The first goal is to control hovering. It should be
ensured it is stable while hovering, given a three-dimensional reference, X-Y position, and altitude, which in
this task are all constant. While tracking the path, the mini-drone will have only a change of reference in the
X-Y position, which is generated and continuously updated from the Path planning algorithm. If we assume
that during these motions, the mini drone should maintain the camera faced down almost orthogonally to see
the track with the right orientation, then there will be low pitch and roll angles, which means that the hovering
61
control system architecture can be used for the path tracking task as well. The control system is designed
according to the approach of Douglas [19].
In this architecture, the mini drone is considered as the plant. It receives as inputs the four motor speeds,
spinning the propellers and generating forces and torques as plant outputs. The control goal requires the mini
drone to hover at a specified altitude. Therefore, the four motors should be controlled autonomously to obtain
this output. The motor mixing algorithm (MMA) can be used to control yaw, pitch, roll, and thrust directly,
rather than controlling in terms of the four motor speeds.
Thrust is always oriented along the Z-axis of the body frame. If the drone flies at steep roll or pitch angles,
then, if the thrust is commanded, there is a coupling between the altitude variation and the horizontal motion.
If a controller for a racing drone must be built, this coupling must be considered because it could fly at steeper
roll and pitch angles. However, for this simple hover controller, low roll and pitch angles should be
assumed. Therefore, a change in thrust only significantly influences the altitude rate.
Firstly, we should think of adjusting the altitude through a PID controller that commands the thrusts. The
drone altitude is the state to measure and to feedback, comparing it to the reference. Then the altitude controller
determines how to command the thrust through the obtained error. If, for example, the mini drone hovers at a
lower altitude, the error becomes positive, and the thrust command is split evenly among the four motors, and
the drone rises. However, the effect of disturbances, like wind gusts, can cause the drone to roll or pitch a
little, and so the thrust could also induce to move horizontally, making the drone drift away from the reference
position. Therefore, the controller is not suitable, and it can be improved through the roll, pitch and yaw angles
control, keeping them constant at zero degrees, so that thrust can only impact altitude, preventing the drone
from wandering away. Thrust, roll, pitch, and yaw commands can be independently controlled because they
are decoupled. For this reason, three more feedback controllers can be added respectively for yaw, pitch, and
roll. Now, these angles are also the plant outputs and the states to estimate and to feedback.
Therefore, in this way, altitude is kept fixed, and the mini drone faces forward. This hover controller is
62
improved, but there are still some limits. If, for example, multiple wind gusts, as disturbances, flows in the
same direction, the controller still allows the drone to meander away slowly. In this control system
implementation, the drone is not brought back to its reference position.
The improvements can be made by thinking that pitch and roll angles are not necessarily zero, but they can
also assume non-zero values while hovering. For example, the mini drone could hover in a strong constant
wind, and in these conditions, it leans at a certain angle into the wind to keep its reference position. Therefore,
we should implement a ground position controller that recognizes when the quadcopter wanders off and makes
the required changes to take it back to the original point in X-Y coordinates.
Due to the coupling between forwarding, backward, left, and right movements with roll and pitch
maneuvers, also position error is coupled with roll and pitch motion. The measured X-Y position should be
fed back and compared to the reference signal to obtain the position error. If there is no signal having the
sequence of position points to follow generated by the path-planning algorithm, the reference position may be
set temporarily to (0, 0). It corresponds to hovering above the take-off point. The position error is taken as
input by the position controller, and roll and pitch angles are sent as outputs. These angles are the reference
signals for the roll and pitch controllers. In this way, the position controller creates them.
Summarizing, the roll and pitch controllers in the inner loop receive the reference signals from the position
controller in an external loop. They are cascaded loops. It should be noticed that also the measured yaw angle
should be fed into the position controller. The X-Y position errors are expressed relatively to the environment
reference frame, but pitch and roll angles are expressed relatively to the drone frame. This means that a
movement in the X and Y world directions can be obtained through pitch and roll only by knowing the yaw
angle because it is needed by the position controller for the conversion from the drone X-Y frame to the
environment reference frame.
The next picture shows the structure of the architecture described before (figure 6.4).
63
Figure 6.4: Feedback Control loop with a nonlinear model for simulation [34]
In the next paragraph, where the Simulink model will be described, it can be noticed that the controller
model has the architecture of the previously explained feedback control system.
6.3 Simulink model structure description
In this paragraph, the Quadrotor model used for simulation purposes will be analyzed, focusing on the part
of the model briefly anticipated in chapter 5. Two main parts are described: Simulation model and Flight
control system for code generation.
6.3.1 Simulation model
The following picture (6.5) represents the subsystems that constitute the simulation model: the multi-copter
model, the sensors model, the environment model, and the flags to stop the simulation, all wrapped around the
Flight Control Software.

64
Figure 6.5: Simulation model blocks
Starting from the multicopter (or airframe) model (Figure 6.6), it can be noticed that the inputs are the
actuators' signals and the data coming from the environment model, while the outputs are the states that must
be estimated by state estimator algorithm.
Figure 6.6: Airframe model (nonlinear and linear)
The airframe model is implemented as a variant subsystem, which means that before the model is run, we
can select which version of airframe we want to run with: either nonlinear model, that can be used for
simulating the flight, or the linear airframe model, that can be used to tune the controllers (Figure 6.7).
The linear model structure can be represented through the typical state-space description and through the
blocks corresponding to “trimLinearizeOpPoint” functions that, through Simulink Control Design, turn the
65
mini drone nonlinear model into a linearized one.
Figure 6.7: Quadcopter Linear model
Figure 6.8: Quadcopter Nonlinear model
Looking inside the nonlinear model (figure 6.8), it can be noticed that there are two main blocks. The AC
model on the left consists of the actuators models and a model of how the environment disturbances impact
the system (these subsystems are shown in figure 6.9).

66
Figure 6.9: AC model (actuators)
Anything that can create a force or torque on the mini drone is calculated in this block. The forces and
torques are then fed into the 6DOF model. This is a rigid body model that comes with the aerospace blockset.
It is an example of using an existing model rather than going through the effort of writing out the dynamical
equations ourselves.
However, it is still needed to determine the specific parameters for this rigid system like masses and inertias.
More than likely, the developer pulled this information from a CAD model of the mini drone, but a physical
test could be set up to calculate this information, or it could be done through system identification techniques.
The following pictures show respectively the block diagrams of the “gravity force calculation” (figure 6.10
on the left), “drag calculation” (figure 6.10 on the right, and 6.11), “motor forces and torques” (from figure
6.12 to 6.15) and “applied force calculation” (figure 6.16) of the AC model.
Figure 6.10: Gravity Force calculation (left) and Drag calculation (right)
67
Figure 6.11: Drag calculation subsystem
Figure 6.12: Motor forces and torques
Figure 6.13: Motor forces and torques (MotorsToW subsystem)
Figure 6.14: Motor forces and torques (Matrix concatenation subsystem)

68
Figure 6.15: Motor forces and torques (Matrix concatenation subsystem structure)
Figure 6.16: Applied Force calculation
Instead, the 6DOF model, which can be described either through Euler angles or Quaternions, is shown in
figure 6.17. This block type is included in the Aerospace blockset add-on tool. The block in figure 6.17 on the
right models the change of the reference frame from the mini drone’s one to the fixed inertial frame.
Figure 6.17: 6 DOF model described through Euler angles (left) and change of the reference frame from
Body to Earth (right)
If we consider, instead, the environment block (figure 6.18) at the top level, it is possible to see that this,
again, is a variant subsystem, and we have the option of choosing constant environment variables or variables
that change based on position. In this block, different Aerospace Blockset environment models are
implemented, for example, the atmosphere and gravity models. The “VSS_ENVIRONMENT” value from the
69
workspace can be changed to choose between the two previously described models.
However, for this project, the constant variables (figure 6.19) will be selected because things like gravity
and air pressure are not going to change between ground level and altitude at less than a meter. Nevertheless,
if the objective is to simulate how high the mini drone could fly, then selecting the changing environment
(figure 6.20) will lower air pressure and density as it gets higher, which will eventually stall the drone at some
maximum altitude. So, choosing one model or another depends on the aim of the test. For our purposes, we
will use the Constant Environment model.
Figure 6.18: Environment model
Figure 6.19: Constant environment

70
Figure 6.20: Variable environment
Lastly, going into the sensors block (figure 6.21), which is also a variant subsystem, the two alternatives
to select are either dynamic sensors with noise or feedthrough sensors. The feedthrough option will be select
for tuning the controllers, but for simulation purpose, the sensors should be chosen to behave as much like the
real things as possible. Inside this subsystem, there are some hardcoded sensor calibration data and a block
called sensor system that houses the models for the camera, the IMU, the ultrasound, and the pressure sensor
(figure 6.22). As previously explained, the following sensors determine the states of the system.
 An Inertial Measurement Unit (IMU). It measures the translational accelerations and angular rates.
(figures 6.23 and 6.25).
 A down-facing camera. It is needed for Image processing and optical flow estimation (figure 6.24)
 An ultrasound sensor. It is needed for altitude measurements (figure 6.25).
The file “sensorVars” contains the sensors’ characteristics. The “VSS_SENSORS” variable in the
workspace can be changed to consider the sensors’ dynamics in these measurements.
Figure 6.21: Sensors model

71
Figure 6.22: Sensors’ dynamics
Figure 6.23: Sensor system (Sensors’ dynamics subsystem)
It is a crude simulation of the “optical flow (vx,vy,vz)” function as fed into the c-function
rsedu_optical_flow() on the mini-drone by internal, inaccessible code of Parrot’s firmware. This is assumed
to be far from what “optical flow (vx,vy,vz)” actually is.
Figure 6.24: Camera model
Figure 6.25: IMU and pressure sensor models

72
Figure 6.26: HAL acquisition creator model
The model can also be used to safely simulate a failure and see how the system does. For example, we
could go in the sensor block, into the IMU model and change the gyro bias. Let us say we estimated the bias
poorly, and it is three times worse than expected. Then the simulation can be run. It takes off just about the
same, but the gyro bias error quickly causes the drone to roll away from level, and then it runs away and
eventually crashes into the ground. If the gyro bias is three times worse, the drone will not perform well, and
if there is worry about being this far off on estimating bias, then the stop flag logic should be changed to
recognize that it is drifting away and so shut the drone down before the hardware would be hurt.
6.3.2 Flight Control system
A technique for the implementation of the flight controller could be through hand-written C code, then
compiling the entire flight code with controller changes and finally deploying the code onto the mini drone
(Option 1 in figure 6.27). However, this approach has some drawbacks during the development of the feedback
control software. The controllers cannot be easily tuned because they should be tweaked on the hardware.
73
Moreover, the architecture of control systems realized through C code could be challenging to explain to other
people, and the impact of changes on the whole system can be harder to understand than using a graphical
method. Instead, the second approach (Option 2 in figure 6.27) consists of a graphical description of the flight
controller through block diagrams. In this way, Simulink can be used for development; then, the FCS can be
turned automatically into C code, can be compiled, and deployed onto the mini drone.
The Simulink code is easier to read, as well as the performance simulation, and the controller gains tuning
through existing tools. Most of the flight code does not have to be written because the Simulink Support
Package for Parrot Mini drones will be used to design the customized flight control software. In this way, we
can load a new flight firmware to the device, preserving all the standard operating functions, but allowing us
to modify only the firmware control part. The same output and input signals should be maintained to ensure
that any written code will be placed correctly after Simulink programming, and it will interface appropriately
with the rest of the mini-drone firmware.
Figure 6.27: the two options by which the FCS could be implemented [35]
During the Hardware-in-the-loop test, the only part that must be deployed on the Hardware is the “Flight
Control system” block, which is the orange block (Figure 6.28), which is also the block to be customized to
design the flight controller.

74
Figure 6.28: Flight Control System block
Inside the Flight Control System, there are two subsystems (Figure 6.29):
 Image Processing System: containing the graphic processing part (the green block).
 Control System: containing the flight logic (the orange block).
The two subsystems work at different rates: The Image Processing System at almost 20 milliseconds (60
Hz as the camera frame rate), while the Control System at 5 milliseconds (200 Hz).
Therefore, they are connected through the "Rate transition" block, which allows the data transfer between
systems with a different rate.
Figure 6.29: FCS structure and its subsystems
Within the Control system subsystem block, there are other different blocks (Figure 6.30):
 Path planning: (the orange one) where the logic for the line-tracking algorithm will be designed.
 Controller: where all the PIDs of the flight controller reside.
 State estimator: contains the state observer.

75
 Crash Predictor Flags: this contains the logic to turn off the drone in case of anomalies in flight.
On the left, it can be noticed instead of a “constant” block with the value “1” assigned (figure 6.31).
It is used to define the type of drone piloting: if equal to 1: the drone will base its movements according to
the X and Y input values, while if the constant is equal to 0: the drone will be piloted changing the Pitch and
Roll values. In any case, it is possible to act on the value of the Yaw (in radians).
Figure 6.30: “Control System” subsystem structure
Figure 6.31: Control mode (“1” for Position XYZ, “0” for Orientation Roll-Pitch-Yaw)
The State estimator block is the block shown in the next two figures (6.32 and 6.33). A complementary
filter is used for attitude estimation and Kalman filters for position and velocity estimations. The variables
characterizing the estimator are contained in the “estimatorVars” file.

76
Figure 6.32: State estimator block
Figure 6.33: State estimator model structure with its subsystems
The acquisition of raw sensor measurements and the state estimation is conducted in two steps. At the first
stage, the measurements are processed and then blended with filters to estimate the control states. Therefore,
the first stage is the sensor processing block (figures 6.34 and 6.35). The sensor data group is needed to
extract the individual sensor values from the sensor bus to manage them in the code (figures 6.36).
Firstly, the previously determined bias is subtracted to the gyro and acceleration data to calibrate them.
Through bias removal, zero measurement results from zero acceleration and zero angular rates. Next, we
should express the measurements from the sensors’ frame to the drone frame by performing a rotation
transformation. Finally, the measurements must be low pass filtered to cut off the noise at high frequency.
77
Likewise, we should remove bias from the sonar sensor. The optical flow data is based on a pass/fail criterion:
if it has a valid position estimate, then the validity flag is set as true by this block.
Summarizing the sensor preprocessing consists of bias removal, coordinate transformation, and filtering.
Figure 6.34: Sensor preprocessing subsystem block
Figure 6.35: Sensor preprocessing subsystem structure
Figures 6.36: Sensor data group subsystem
After filtering and data calibration, we can proceed with combining measurements for state estimation
useful for the controller subsystem. The following pictures show the blocks used to estimate the X-Y position
(from figure 6.37 to 6.46) and altitude (from figure 6.47 to 6.50). It can be noticed they employ a Kalman
filter to combine the measurements, to predict the dynamical system behavior, and to obtain an optimal
78
estimation.
Figure 6.37: X-Y position estimator block
Figure 6.38: X-Y position estimator structure and its subsystems
Figure 6.39: velocity estimator subsystem block (part of the X-Y position estimator)
79
Figure 6.40: velocity estimator subsystem structure
Figure 6.41: Acceleration Handling subsystem block (part of the velocity estimator)
Figure 6.42: Acceleration Handling subsystem structure (part of the velocity estimator)
Figure 6.43: Data Handling subsystem block (part of the velocity estimator)
80
Figure 6.44: Data Handling subsystem structure (part of the velocity estimator)
Figure 6.45: X-Y position estimator subsystem block (part of the general X-Y position estimator)
Figure 6.46: X-Y position estimator subsystem structure
Figure 6.47: Altitude (Z position) estimator block

81
Figure 6.48: Altitude (Z position) estimator structure
Figure 6.49: Outlier Handling subsystem block (part of the Altitude estimator)
Figure 6.50: Altitude (Z position) estimator block
The block in figures 6.51 and 6.52 (subsystems in figures 6.53 and 6.54) is used to estimate roll, pitch, and
yaw angles through a complementary filter. This type of filter can simply combine measurements from two
sensors, and it is suitable for the subsystem.

82
Figure 6.51: Complementary filter block for orientation estimation
Figure 6.52: Complementary filter structure
Figure 6.53: Complementary filter subsystem 1

83
Figure 6.54: Complementary filter subsystem 2
The complementary filter is a more straightforward solution to implement compared to the very capable
but intricate Kalman filter, which is often harder to implement on specific microcontrollers. The following
description of the complementary filter was adapted from Jouybari et al. [38].
The gyroscope and accelerometer measurements are employed for the same objective: determine the
drone’s angular position. We can achieve this with the integration over time of the gyroscope angular velocity.
Instead, the angular position can be obtained through the accelerometer, using the gravity vector position. We
can simply achieve this with the “atan2” function. However, there are two problems, justifying the need for a
filter to process data. The first issue concerns the accelerometer. Since it measures all the forces acting on the
drone, it can also perceive more components than only the gravity vector. All the small forces could disturb
the measurement. The sensor can also sense the forces driving the drone since the system is actuated.
Therefore, we should use a low-pass filter because of the reliability of accelerometer measurements exclusively
in the long term.
The second issue affects the gyroscope. We can easily obtain accurate measurements not significantly
influenced by external forces, but due to the integration over time, there is a tendency of the measurement to
drift, that is, not coming back to zero value after a return of the system to its original position. Due to this
tendency of drifting in the long term, gyroscope measurements are reliable only in the short term.
Therefore, a complementary filter can be the right tradeoff solution to the two issues. Because of the
gyroscope precision, not sensitive to external forces, we can use its measurements in the short term. Instead,
84
we can use the accelerometer measurements in the long term, for the absence of drift. The filter can be
implemented according to the following simple form, adapted from Van de Maele [37].
𝑎𝑛𝑔𝑙𝑒 = 0.98 ∗ (𝑎𝑛𝑔𝑙𝑒 + 𝑔𝑦𝑟𝐷𝑎𝑡𝑎 ∗ 𝑑𝑡) + 0.02 ∗ (𝑎𝑐𝑐𝑒𝑙𝐷𝑎𝑡𝑎) (6.1)
In this way, we integrate the gyroscope measurement (gyrData) every timestep with the current value of
the angle (angle). Next, we combine it with the filtered measurement from the accelerometer (accelData), after
processing it with the “atan2” function. The total sum of the constant parameters in the formula should be 1,
and they should be chosen according to our filter tuning. This function implemented through Simulink blocks
must be used with a loop. Every cycle, we update the roll and pitch angles with the new measurement from
the gyroscope with time integration. Then, the angles are updated with the accelerometer measurement
considering 98% of the actual measurement and summing up 2% of the angle computed from accelerometer
data. In that way, the measurement is ensured not to drift, and it is accurate in the short term.
Figure 6.55 shows a graph created with GNUPlot that represents the signals from gyroscope and
accelerometers and their estimation through a complementary filter. We can notice that the filtered signal (in
red) follows the gyroscope (in blue) for quick variations, and it follows the mean value of the accelerometer
measurements (in green) for slow variations, not sensing the noise and not drifting away too.
Figure 6.55: gyroscope and accelerometer signals and complementary filter estimation [37]
Now that State estimators have been treated, it can be possible to introduce the controller subsystem (figure
85
6.56). It takes the reference signals generated from the Path planning algorithm and makes a comparison
between them and the estimated states to obtain the errors. They are fed into the PID controllers to produce
the actuators’ commands (figure 6.57). We have the X-Y position outer loop controller (figures 6.58 and 6.59),
sending the signals to the pitch /roll (or attitude) internal loop controller (figures 6.60 and 6.61). Moreover,
independently of them, there is a yaw controller (figure 6.62) and the height controller (figures 6.63 and 6.64).
Overall, the position and the attitude of the mini drone are controlled by 6 PID controllers.
The “controllerVars” folder includes all the variables relative to the controller.
Figure 6.56: Controller block
Figure 6.57: Controller model structure
Figure 6.58: X-Y position controller block

86
Figure 6.59: X-Y position controller structure
Figure 6.60: Attitude controller block
Figure 6.61: Attitude controller structure
Figures 6.62: Yaw controller block and model structure

87
The altitude controller (figures 6.71 and 6.72) is set up as PID. In this implementation, we multiply the
proportional gain by the altitude error derived from the sonar sensor, and the derivative gain by the altitude
rate measurement of the gyroscope, which is a less noisy signal compared to the ones from the ultrasound. It
would be useful to remember that the z-axis points down in the drone’s coordinates system, so in the control
system, there will always be a negative sign in front of the altitude number (expressed in meters).
Figure 6.63: Altitude controller block
Figure 6.64: Altitude controller structure
Those PID and PD controllers have as outputs the force and torque commands that are subsequently sent
to the MMA (figure 6.65). It produces the necessary motor thrusts, and, through the blocks in figure 6.66, the
commands are turned into motors’ speeds.

88
Figure 6.65: Motor mixing Algorithm block and model structure (part 1)
Figure 6.66: Motor mixing Algorithm block and model structure (part 2)
6.4 Controllers Tuning
As previously stated, to tune the controllers, we should build a linear model, because nonlinear models are
not suitable for controller design, despite their accuracy in simulation. In this example, the “PID tuner” app
of Simulink will be used for tuning the height controller, and the same approach can be applied for the other
controllers according to the technique explained by Douglas [35].
Figure 6.67: Model linearization for Controllers tuning [35]
Once the system is linearized, we can apply the superposition principle, so it is possible to tweak and adjust
altitude without influencing yaw, pitch, and roll because the altitude controller is independent of the other
ones. We can also assume that sensor dynamics and noise effects are negligible in the controller design.
89
Therefore, we can remove the state estimation logic and the sensors model. Moreover, we should assume the
controller knows the actual height.
Once the controllers are tuned, they will be tested on the original nonlinear model and check that the
assumptions were valid.
Figure 6.68: Control system simplification for Altitude controller tuning [34]
We should linearize the altitude controller and adjust the gains to obtain the desired dynamics.
After that, the yaw controller will be tuned, with roll, pitch and altitude constant, and a similar procedure
for pitch and then roll. Then, the outer loop position controller will be tuned while the inner loop controllers
are working and keeping the orientation.
For the linearization of the altitude loop example, we can remove the scope and the sensor block since there
is a feedback of a perfect altitude state. In the flight controller block, we can keep the altitude reference, the
controller, the MMA, and the thrust-to-motor command block. The RPY torques are set to 0 to ensure that
without external forces and torques as disturbances on the drone, it can only increase or decrease altitude. This
can be assumed because the environment block does not model external disturbances.
Finally, the perfect altitude state must be fed back to the altitude controller. The linearized system is shown
in figure 6.69.
90
Figure 6.69: Linearized Control system used for Altitude controller tuning [35]
The altitude controller can be implemented with a PD controller through the relative block in Simulink
because then it can be used for autotuning. The structure consists of a comparison between the altitude
reference and the actual altitude, a PD controller that receives the error, and an addition of a feedforward
gravity term.
The last one models the necessary thrust to offset the drone weight. In this way, the controller can simply
command the altitude positively to rise and negatively to descend. An alternative to the use of this term could
be the addition of an integral term to the controller. The next figure (6.70) shows the linearized controller
model used for tuning.
Figure 6.70: Simplified Altitude Controller used for tuning [35]
The “autotuner” can be run by opening the PD block. Firstly, it linearizes the control loop. Then, the app
plots the closed-loop response of the linearized version, and it is possible to adjust response time and transient
behavior of the system (figure 6.71).

91
Figure 6.71: PID Tuner App on Simulink [35]
The dashed line of the response signal has not the same behavior of simulation because of the removal of
nonlinear components, but it is still suitable for tuning. After gain selection and a quick test with the hardware,
it can be noticed that the hardware does not reflect the expected behavior because it cannot take off properly.
The problem, in this case, affects the feedforward term. If it is too little, the system believes the drone
weight is smaller than in reality, or the thrust is higher than the actual one. Therefore, the controller has more
difficulty in handling the remaining weight because of a reduction of the proportional path, so it cannot take
off. If the value is raised to about 25%, the drone can finally takeoff.
The same tuning procedure could be repeated for the other controllers until not just the simulation
performances, but also the hardware test performances, are considered acceptable.
Table I shows the gains of the six controllers tuned through simulation and real-time tests by following the
technique explained previously.

92
TABLE I: TUNED GAINS OF CONTROLLERS
CONTROLLER TYPE GAIN TYPE VALUE (sign according to the

reference frame convention)
P for X coordinate error -0.24
X-Y position PD controller P for Y coordinate error 0.24

D for X coordinate error 0.1
D for Y coordinate error -0.1
P for pitch error 0.013
P for roll error 0.011
Attitude PID controller I for pitch error for attitude 0.01
loop
(pitch/roll) I for pitch/roll error in the 0.24
altitude loop
D for pitch error 0.002
D for roll error 0.003
Yaw PD controller P for yaw error 0.004
D for yaw error 0.0012
Altitude PID controller (Z) P for altitude error 0.8
D for altitude error 0.5
6.5 Hovering Hardware-in-the-Loop Test
The Hovering test was performed to check the real-time performances after controllers tuning. During the
test, the drone was pushed slightly, with pulses acting as disturbances on the airframe. The figures on the next
page show the graphs obtained plotting the data recorded from sensors during flight. They represent the
trajectory (figure 6.72), the motor speeds (figure 6.73), x-y-z positions with their linear speeds (figure 6.74),
yaw-pitch-roll angles with their angular velocities (figure 6.75). It is easy to identify on the graphs the moments
when disturbances occurred (5 times in total), due to a sudden variation in velocities. The signals show that
the drone comes back to the reference point with a response suitable to our control goals.
93
Figure 6.72: real-time Hovering test– Trajectory (3D view and Top view)
Figure 6.73: Real-time Hovering test– Motor speeds

94
Figure 6.74: Real-time Hovering test – X, Y, Z Positions (left) and corresponding linear speeds (right)
95
Figure 6.75: Real-time Hovering test – RPY angles (left) and corresponding angular velocities (right)
The performance of the control system response to the disturbances in the HIL test validates the controllers'
gains tuned in the MIL simulation test, showing that the system is stable. Thus it is possible to proceed with
the image processing and path planning algorithms design.

CHAPTER 7
1st LINE-TRACKING ALGORITHM DESIGN
The path-tracking algorithm to be designed is constituted by the image processing subsystem (figure 7.1)
and the path planning subsystem within the control system (figure 7.2), which will generate the reference
commands for the flight controllers. The other components of the Flight Control system like the estimator
algorithm and the controllers can be left unchanged at this point, after having already tuned those in the
previous design stage.
Figure 7.1: Image processing (left) and Path planning (right) subsystems
7.1 Image processing algorithm (Color Thresholder app and sub-images analysis approach)
It is better to use another model to design and test the algorithm that processes images from the camera. A
new Simulink project can be opened to create it, from the main MATLAB screen, by following the procedure
explained in the next lines.
Firstly, click on "New" and then "Simulink model" (figure 7.3 on the left).
Figure 7.2: template model creation for Image processing
96
97
By opening the menu "Simulink Support Package for Parrot Minidrones", select the template "Code
Generation Template for Image Processing and Control" (figure 7.3 on the right). Through this simplified
model, it is possible to test the code without flying the drone independently; in fact, it will be possible to
manually hold it above the path to analyze the behavior of the image processing algorithm. It would be useful
to download an image of the route from the drone camera.
Firstly, the app called “Color Thresholder”, provided by the Computer vision add-on tool in Simulink can
be used to calibrate a threshold filter that creates a binary mask containing only the path. The tool can be
opened by typing "colorThresholder" in the Command Window. Once the screen is open, we can select the
image to be loaded (figure 7.4), or we can look for it inside the PC or from the Workspace. For our purpose, it
is preferable to upload an image coming directly from the camera of the drone (with at least one route line
included).
After that, we can choose between 4 different spaces: YCbCr, RGB, L*a*b*, and HSV. For our purpose, a
color space with a very robust control is HSV, but since the algorithm should be made less complicated, it can
be used merely RGB. HSV is a three-dimensional representation of the color based on the tonality components
(Hue), Saturation (Saturation), and Luminosity (Value). It was defined in 1978 by Alvy Ray Smith is a
representation closer to the form in which we humans perceive colors and their properties, unlike RGB, color
shades are grouped.
To the right of the image (in the App window), we have the cursors to adjust the components of the filter.
Once we reach the desired segmentation (Figure 7.4), we export the function in a .m file.
Figure 7.3: Color Thresholder app operations

98
Figure 7.4: Color Thresholder app
This phase can be simplified a lot if the path is made with white tape on a black background, which is
already a binary scenario. However, if other elements of different colors are added to help and to improve the
Optical flow estimation algorithm, the threshold filter should be able to cancel those out.
After this procedure, we are interested in detecting the edges of the path and their angles. An edge can be
modeled as a curve that shows a sudden variation in intensity. The "edge" function can be used to detect the
borders. The function seeks for areas in the picture where there is an abrupt variation in intensity, through the
following two criteria, as explained by Zhang et al. [39].
 The Points where the value of the 1st derivative of the intensity function exceeds a selected threshold.
 The Points where the 2nd derivative of the same function becomes null.
The “edge” function uses different estimators that implement one of the previous methods for identification.
In some of them, it is possible to define if the action must be sensitive to rather vertical, horizontal, or both
kinds. The function output is a binary picture assuming “1” close to the object borders, otherwise “0”.
Furthermore, another useful line detection method that can be used is the “Canny” method. The peculiarity
of this technique is the use of two thresholds, one for “strong” and another one for “weak” edges. In particular,
it highlights the weak ones only if there are connections to strong ones. The advantage of this technique is that
99
it can be influenced less likely by the noise, and it is more likely to find accurate weak contours.
For edge detection, we can use the “hough” function. It applies the SHT (Standard Hough Transform), and
it can be expressed through the following parametric representation:
𝜌 = 𝑥・𝑐𝑜𝑠(𝜃) + 𝑦・𝑠ⅈ𝑛(𝜃) (7.1)
The parameter “ρ” is the distance between the edges and the reference frame’s origin. “Θ”, instead, can be
interpreted as the orientation of the line, according to the convention shown in the next picture (Figure 7.5).
Figure 7.5: parametric representation of a line through SHT [45]
The “Hough” function generates an array of parameters θ and ρ, located respectively on the columns and rows
of the matrix. Once the picture is analyzed with the SHT, the “houghpeaks” function is used in the parameter
space to detect the peaks, which are potentials lines in the binary image. The “FindAngle” (Figure 7.6 and
Appendix B for the script) function can be employed to estimate the edges’ angles, which are then fed to the
path planning algorithm.
Figure 7.6: user-defined function “FindAngle”
If we look at the next pictures, we can notice that the transform calculates a null angle in the case in which
the line is oriented as in the picture in the center, whereas it will calculate a negative degree if it is oriented as
in the picture on the left, vice versa, a positive angle will be sent as output if the line is oriented as in the picture
on the right.
100
Figure 7.7: Track lines with different angles (positive angle on the right)
Figure 7.8: Image processing subsystem
As can be seen from figure 7.8, as input to the sub-system, we have the Image Data that comes processed
by the PARROT Image Conversion block, which has an Output with the components of the image, coming
from the camera, in RGB. The filter generated by the Color Threshold App has been loaded into a MATLAB
block Function and has as input the RGB three-dimensional matrix and returns an image to the Output binary
(in a matrix of size 160 × 120 pixels) containing “1”s where our track is present, while “0”s elsewhere.
The image was divided into various sub-matrices to analyze their pixels' binary values. In figure 7.9, it can
be seen how the image was subdivided into the different areas useful for the path planning algorithm logic.
Figure 7.9: Image subdivision into areas for the path planning algorithm
Through this technique the path planning logic can establish when the drone should land above the circle.
101
In fact, in every sub-image of the binary image, the number of the “1”s is counted and then compared with a
constant; therefore, we will have a Boolean value at the output. There are three Boolean signals which are
conveyed to an AND block. This causes that the condition must be true on all three sub-images. In this case,
the value of the constant to be compared with the sum of each sub-matrix is low enough to consider the residual
noise from the filtered image. Through Hardware tests, this value will be adjusted so that the landing happens
in the right condition.
Regarding the orientation, i.e., the part of code that acts on the Yaw angle of the aircraft, two different gains
are used to control it: the first one has a higher gain to align the orientation with the line approximately, and
then the second gain lower, to align the drone more precisely above the line (using the Hough transform).
However, when the line is positioned horizontally (from the image of the camera), the transform calculates
an angle that could oscillate rapidly between −90° and 90°. If in input to the X-Y displacement function, we
give these very oscillating values, the drone will not be able to follow the path. Therefore, a function has been
placed to align faster the drone as soon as it sees the line and then continues with a more precise alignment.
Two portions of the image have been dedicated to this function, called “Up_left” and “Up_right” (purple areas
in Figure 7.9). A Matrix Sum block has been placed in output to the two sub-matrices to count the number of
pixels; this will tell us how many white pixels of the path are present in that portion.
Instead, the portion of the image dedicated to the Hough transform analysis is the one called “ForDirection”
(blue area in Figure 7.9).
7.2 Path planning algorithm (Stateflow approach)
Stateflow is a Simulink tool that consists of a schematic language made of truth tables, transition diagrams,
flow diagrams, and state transition blocks. It allows us to design and develop supervisory logic, communication
protocols, GUI (user interfaces), state machines, and hybrid systems. Through this powerful tool, it is possible
to realize the sequential supervisory logic that can be modeled as a subsystem in Simulink. Various checks
102
during modification and execution ensure consistency and completeness of the project before implementation.
In the part of the movement management system, there are two Stateflow charts, two Bus Selector, two Bus
Creators, two Switches, two Gain, a Constant block, a Delay block, and a block conversion to "Single" type.
A bus arrives from the input states “estim” which is divided by a Bus selector; in this case, there is only
one way out because we only need the Yaw estimate. The same goes for input Vision-based Data, but the Bus
selector has multiple outputs: “Land it”, “DegAngle”, “UpLeft” and “UpRight”.
In Output, instead, we have another bus that is composed of the input bus ReferenceValueServerCmds
together with the pos ref and orient ref channels, which respectively receive the values yes set-points x, y, z,
and those of Yaw, Pitch, Roll via a Bus creator. In this case, being the constant “controlModePosVsOrient”
(Figure 6.31) equal to 1, the values of Pitch and Roll cannot be given as references.
The Delay block, together with the Add block, manages the task of varying the Yaw angle. The block Delay
keeps in memory the previous value that will be added (Add type block) to the variation provided by Switch1.
The Switch, instead, activates and deactivates the rotation, varying between the value of the constant (equal to
zero) and the output value from Chart 2. The operation of the Switch, together with the two Gain, will be
explained more in-depth in the section dedicated to Chart2.
In Chart 1 is implemented the task of managing the drone's x, y, and z coordinates in space. It can be seen
in figure 7.10 that it is modeled as a State machine and is divided into three states: Take Off Cruise and
Landing. The picture on the next page (figure 7.11) shows the Path planning subsystem structure.
Figure 7.10: Chart 1 subsystem

103
Figure 7.11: Path planning subsystem

104
The first state, which is the default one (the first to be performed) with the name “TAKEOFF” has the task
of raising the drone up to 0.9 meters from the ground and assigns the value False to the variable rot. This
variable, which will be explained in detail in the section of Chart 2, is used to "activate" the change of heading
of the drone (i.e., the yaw angle). During the take-off, it is necessary to turn off rotation due to the image
recorded by the camera when it is too far near the line; with the image completely white or completely black,
the video processing algorithms are not accurate and could lead the drone to instability. After hovering for 3
seconds, it switches to the "CRUISE" state (this time it is defined on the body of the arrow that connects the
“TAKE OFF” state with the “CRUISE” state: "[after (3, sec)]"). During the change, the value "rot" is assigned
the value True, thus activating the rotation.
In the entry part of this state (which is performed only once), they are initialized to zero the variables dx,
dy, and land. While during, the increase values dx and are calculated dy calculated as follows:
𝑑𝑥 = 𝑣𝑒𝑙・𝑐𝑜𝑠(𝑦𝑎𝑤ⅈ𝑛) (7.2)
𝑑𝑦 = 𝑣𝑒𝑙・𝑠ⅈ𝑛(𝑦𝑎𝑤ⅈ𝑛) (7.3)
Where the variable speed is a multiplicative constant equal to 4.5 · 10−4. At each iteration, the values “dx”
and “dy” are added respectively to the variables “xout” and “yout”.
In the last three lines, the algorithm checks if the landing flag is active; if this condition occurs, the “land”
variable is increased. The reason for this variable is intended to delay the landing slightly and not to create
false landings if the filter does not detect the path for short moments.
When the land variable reaches at least the value 150, it enters the LANDING state. During this state, the
xout and yout variables are reassigned (in a redundant form) their value, to keep the drone in the last assigned
x, y coordinates. Plus, the value “zout” is increased by 2.5 · 10−3 meters at each iteration; in this way, the
drone is lowered of altitude more slowly until it touches the floor and land above the circle.
105
Figure 7.12: Chart 2 subsystem
Chart 2 (shown in figure 7.12) realizes the task of aligning the drone above the line of the path. It is divided
into two states: “General Steerer” and “Fine steerer”.
The default state is the GENERAL STEERER. In the "entry" part, the variables: “OUT”, “p” and “r” are
initialized to zero, and then, “SpeedSelector” will set as False. Instead, in the iterations of this state, two types
of checks are made: the first acts on a cursor, represented by the variable p, which is decremented if the number
of pixels present in the upper-left area of the picture (UpLeft) is greater than or equal to 150; conversely, the
variable p will be incremented if the number of pixels contained in the upper-right area of the picture (UpRight)
will be greater than or equal to 150.
In the second control, the cursor value is analyzed. If the variable p is higher than the threshold value
(Threshold), the output “OUT” will be equal to 1. On the contrary, if the variable p is less than the threshold
value, the output “OUT” will be equal to -1. Instead, if it is between the two values −Threshold <p <Threshold,
the OUT variable will be zero. This state here is used to start the drone spinning, for the Hough transform has
angle limits. We have a reasonable estimate when the angle is between −77° and 77°. When the line reaches
an angle between −55° and 55°, the “END” state is activated in the Steerer. Upon entry, the variable “q” is
initialized to zero, and the variable “SpeedSelector” becomes true. Now the “OUT” variable directly receives
106
the angle value calculated from the Hough transform. Finally, a check is carried out, that is, if the drone is
aligned with the path with an error defined by the “Precision” variable, which is equal to 2°. The
“SpeedSelector” Flag acts on a switch (Figure 7.12), which selects the suitable gain depending on the state
assumed in Chart 2. In fact, in the “GENERAL STEERER” state, the variable OUT can only assume the
values: -1, 0, and 1, while in the “FINE STEERER” state, the variable OUT takes the values between −55°
and 55°. This method solves the problem of very oscillating values for the Hough transform, and it also makes
the rotation speed decrease as the error angle decreases, in other words, the smaller the error angle, the more
the drone turns slowly so as not to exceed the zero angle. To be considered valid, the algorithm must work on
different path configurations.
7.3 Simplified Version of the previous algorithm
We can simplify the model by removing the function for angle detection and the Canny method and make
the image processing algorithm rely just on the pixel count of the left and right sub-images to make the drone
turn. The next figure (7.13) shows the model of the image processing subsystem.
Figure 7.13: Simplified version of the Image processing subsystem
We can also simplify the Path planning algorithm, using just one gain for turning with yaw, instead of
having two (one more precise than the other for small angles). Moreover, the algorithm can allow the drone to
stop before turning by changing the yaw angle, and in this way, it would be slower but more precise. Figure
7.14 shows the simplified path planning algorithm.

107
Figure 7.14: Simplified Path planning subsystem
The next picture (7.15) shows Chart1 and Chart2. In Chart1, we can notice that there is a new state called
“Rotation,” and it is activated only for half of a second, and it let the drone keep the same coordinates, as we
said before, without changing the position while turning. It allows for more accurate maneuvers. In Chart 2,
there is only a general state for steering with just one gain value with two outputs. They give the sign to the
gain, depending on the direction of turning, if it is left, then “-1”, otherwise “1”.
Figure 7.15: Chart1 (left) and Chart2 (right)

108
7.4 Model-in-the-loop test (standard and simplified implementations)
When the two algorithms were tested, they showed a satisfying performance, proving that the drone can
follow the track until the end, with their design and their parameters tuned. The next pictures represent some
screenshots taken from the Model-in-the-loop simulation of the standard implementation (first). The drone can
precisely track the line, and it can deal with curves of different angles. Its performance relies on the tuning
combination of different parameters that characterize the algorithm. There is the altitude at which the drone
has to fly to see clearly the path; moreover, there are the sub-images dimensions for steering control coupled
with the threshold on the pixel sum that activates the command to change yaw orientation (increasing or
decreasing the angle). Another critical parameter is the drone speed, which has to be set according to the
previous parameter; otherwise, the drone could oscillate too much or even lose the vision of the path in the
sub-images.
Figure 7.16: Model-in-the-loop test of the first algorithm - standard version
The simplified model is also able to track the line, but it does so with less fluidity in movements above the
curves because, as was explained in the previous paragraph. If the steering control activates, the drone stops
to update the x-y position with increments and change just the yaw angle, having a slower but more precise
maneuver while tracking the curve. On the next page, there are tables II and III with the tuned values of the
parameters that characterize the algorithm (respectively, the standard and the simplified).
109
TABLE II: PARAMETERS VALUE OF THE FIRST ALGORITHM
SPEED (“VEL” CONSTANT) 4.5 * 10-4
ALTITUDE (“Z”) 1.1 m (3.61 ft)
SUB-IMAGES DIMENSION FOR STEERING CONTROL Up_Left: rows 1-39/columns 1-55
Up_Right: rows 1-39/columns 103-157
THRESHOLD ON THE SUM OF TRACK PIXELS (≥) 140
As we can see in the next pictures (figure 7.17), the simplified algorithm makes the drone follow the curve
slower and with more precision compared to the standard version.
Figure 7.17: Model-in-the-loop test of the first algorithm – simplified version

CHAPTER 8
2nd LINE-TRACKING ALGORITHM DESIGN
8.1 Image processing algorithm
The 2nd Image processing algorithm implemented is made differently; in fact, it is built mainly with
functions defined through MATLAB. The functions’ scripts are provided in the Appendix. The only
components that this algorithm has in common with the first Image processing algorithm are the first two
stages. These are the “Binarization” function (figure 8.1 on the left and Appendix A.1.1), done through the
“Color Threshold App” that allows us to obtain the Binary image together with path and background detection.
It is followed by another filter (figure 8.1 on the right and Appendix A.1.2) that removes the noise pixels and
makes the contours of the path smoother and continuous. The aim is to improve the image for the operations
of the following stages.
Figure 8.1: “Binarization” (left) and “Filter” (right) function blocks
At this point, the image is processed by two different branches: one needed for the path detection and
tracking, the other one which tells the drone that the path has ended and identifies the landing point that, in
this case, is a circle. The branch aimed to track the path is made with a function called “Ctrl_Radar” (figure
8.2 on the left and Appendix A.1.3), which tells the drone the direction to follow, generating the signals
corresponding to the increments in position, once the shape of the path has been analyzed. During its cycles of
path analysis, the output signals “x_incr” and “y_incr” are processed through a discrete digital filter and then
110
111
fed back as inputs for the next analysis. The increments are then normalized (figure 8.2 on the right) before
being sent to the Path planning algorithm to allow more precision in the movements and a “smoother”
trajectory.
Figure 8.2: “Ctrl_radar” function with discrete filter (left) and gains for increments (right)
The signal “Start_End_Flag” tells us if the path is ended in that direction of movement or not and it is fed
to the Edge detection subsystem (figures 8.3 and 8.4) that as output has a signal which indicates if the End of
the line was detected, and it is included in the Bus “Vision-based data” provided as input to Path planning
algorithm.
Figure 8.3: Edge detection subsystem block
Figure 8.4: Edge detection subsystem model
The complete schematic of the Image processing subsystem is shown in figure 8.5 on the following page.
112
Figure 8.5: Image processing subsystem

113
As previously anticipated, the other branch that takes the filtered binary image as input is used to detect the
circle to land on and to lead the drone to land safely on the circle. The first stage of this branch is the function
“Borderless” (figure 8.6 on the left and Appendix A.1.4), which removes the objects that touch the borders of
the image (only closed shapes in the image will be shown). This is needed when the image has the circle
centered and the rest of the path line that does not have to influence the path planning algorithm once the path
has ended, and so during the landing phase, the final portion of the path will not be visible in this image.
Then, the function “Filter_C_Det” (figure 8.6 in the center and Appendix A.1.5) is a filter that threshold
the image removes noise pixels, and detects the final circle when almost centered (in the central area of the
image).
Next, the function “Ctrl_circle” (figure 8.10 on the right and Appendix A.1.6) counts the pixels of the
circle, subdividing the image into four quadrants and counts the “unbalanced” pixels between left and right
halves and between upper and lower halves. So it gives the errors of the X-Y position of the circle center
compared to the center of the image (this algorithm works with all symmetric shapes like squares or rectangles,
not just circles).
Figure 8.6: “Borderless” (left), “Filter_C_Det” (center) and “Ctrl_Circle” (right) functions
The signal “C_DET” is a boolean value that is true if the circle is detected in the center of the image, and
it is fed to the Path planning algorithm to activate the landing phase on the circle. The signal “Sum_end”
represents the sum of the pixels of the circle.
Finally, the output signals “x_Last” and “y_Last”, as the increments of tracking function are normalized
before being fed to the Path planning algorithm, so these signals are multiplied by the gains needed to control
the X-Y movements more precisely while centering the circle (figure 8.7).
114
Figure 8.7: Gains for positioning phase before landing
8.2 Path planning algorithm
The 2nd Path planning algorithm is also very different than the 1st solution because the latter was
implemented through State machine representation and Stateflow, while the following one is implemented
through user-defined Matlab functions. The Vision-based data bus coming from the Image processing
algorithm contains the X-Y increments during tracking and the circle increments for positioning during
landing. They are filtered with a Butterworth 1st order filter with a cut-off frequency at 0.3 Hz (figure 8.8). It
is used to filter out the noise from the signals before they are processed in the Path planning subsystem logic.
Figure 8.8: Butterworth Filter implementation
Figure 8.9 on the next page shows the complete schematic of the Path planning subsystem.
115
Figure 8.9: Path planning algorithm

116
The function “Z_Ctrl” (figure 8.10 and appendix A.2.1) provides a constant Z reference for the drone that
must hover at a fixed altitude.
Figure 8.10: Altitude planning
After 5 seconds (1000*Ts) the tracking state is activated, and if both the signals “Enab_Go” and
“End_Track_Detection” are both true (1 if the end of the track is not detected), the tracking function “XY_Ctrl”
(figure 8.11 and appendix A.2.2) is activated and multiplies the position signals by the gains selected in the
function.
Figure 8.11: Tracking function
If the circle is detected for more than 1 second, and the end of the track is detected with the signal becoming
0 (figure 8.12), the slow movement tracking function is activated to center the circle before and during landing
(figure 8.13 and appendix A.2.3).
Figure 8.12: End of line detection (left) and logic for switching to landing phase (right)
117
Figure 8.13: Tracking with slow movements for positioning
The function “XY_Ctrl” for landing (figure 8.14 and appendix A.2.4) disables the slow movements tracking
and make the drone reduces the altitude until it touches the ground.
Figure 8.14: Landing control block
At the end of the path planning stages, all the increments in position signals X, Y, and Z are summed up
and generates the reference signals for the controller subsystem.
Figure 8.15: position signals generation with increments addition

118
8.3 Model-in-the-Loop test with Simulation
The algorithm was tested with a simulation in the Simulink 3d environment, and the next pictures show the
track used (figure 8.16) and the screenshots taken from the Model-in-the-loop test, showing that this algorithm
does not rotate the drone but make it track the line just according x-y position. The algorithm can follow the
line, and it lands on the circle at the end, satisfying the requirements.
Figure 8.16: track of the simulation test for the second algorithm
Figure 8.17: Model-in-the-loop test of the second algorithm
The pictures on the next page (figure 8.18) represent the signals over time of filtered increments (dx and
dy) coming from the image processing subsystem and the position coordinates of the drone relative to the earth
frame.
119
Figure 8.18: x-y increments signals filtered (top) and x-y-z position coordinates of the drone (bottom)
CHAPTER 9
HARDWARE IN THE LOOP TEST AND VALIDATION
9.1 Environment setup and HIL test of first design version (with issues explanation)
The last phase of Model-based design consists of deploying the algorithm on the Hardware and test it in
real-time. The environment for the test was prepared, as shown in figure 9.1. The only available space has the
minimum required dimensions to fly the drone safely, and the presence of other objects around the path and
lighting conditions not easy to adjust has created many difficulties in preparing the tests. This is due to the
extreme sensitiveness of the ultrasound sensors and camera also because of the low-cost hardware according
to the project goal.
Figure 9.1: Example of the track in the environment used for the HIL Test
The path was made with white tape on a black background so that the contrast between these two is
optimized, and these conditions can favor the Image processing algorithm. The use of the black cloth was also
justified because initially, the floor was too glossy, and it created too much noise and disturbance for the
camera and the optical flow estimation algorithm.
Moreover, the optical flow estimation was helped by putting the red pieces of paper randomly on the floor
120
121
to create a pattern that can be a reference for the optical flow algorithm (figure 9.2). In fact, without these
pieces, the absence of other objects and shapes other than the path could confuse the estimator, and the flag
can be activated terminating the flight, or if we deactivate the flag previously, we can see the drone drifting
away from the path. However, the pattern should be filtered out by the image processing algorithm, and the
binary image must not show the pattern shapes as white pixels.
Figure 9.2: Optical flow estimation [40]
Dealing with sensors measurements and estimation algorithms in the real-time test is very different than in
simulation. In the hardware-in-the-loop test, the most problematic aspects are the disturbances on the
ultrasound sensor that can lead the drone to drift up to the roof or the noise of the low-resolution camera
coupled with light conditions.
While the first design implementation worked satisfactorily in the simulation, the real-time test highlighted
issues connected with the factors previously treated, and it also showed how sensitive was the image processing
algorithm to them. This implementation relies on pixels count. As it was explained in chapter 7, the drone only
turns when the number of pixels overcomes a certain threshold, meaning they are pixels of the track and not
noise pixels. The issue occurs if the camera detects, for example, a large bright zone of noise due to light
reflection on an object. It also worsens if we notice that the drone orientation can influence the colors’ shades,
in our case the white color of the path could appear as grey and can be filtered out by the binary filter (figure
9.3), or some zones of the pieces of paper of the pattern could appear brighter, and so they could be detected
instead as part of the path by the image processing subsystem.

122
Figure 9.3: Issue related to wrong filtering of the track due to darker regions of the path
Another issue is related to altitude estimation. In the model-in-the-loop test, the drone was set to fly at a
constant altitude. Therefore the parameters, like the threshold value or the dimensions of the sub-images used
for steering, were tuned accordingly to the path dimensions seen by the camera at that particular altitude.
However, during the real-time test, the altitude is subject to variations or oscillations due to the ultrasound
sensitivity, and so if the altitude changes coupled with pitch and roll motion for disturbances, the track shape
can be seen very distorted, greater or smaller, especially if the drone hovers too close to the ground. Therefore
the pixel count technique is less robust, and the assumptions according to which the drone hovers at a fixed
altitude and roll and pitch angles do not affect path shape could not be valid in this case. Furthermore, the other
characteristic parameters of the algorithm, like speed, sub-images dimensions, and threshold, are tuned
according to the altitude. Thus even a small change in altitude could meaningfully affect the performances and
the effectiveness of the algorithm.
The sensor data plots of the real-time test of the first implementation are shown in the pictures on the next
page (figure 9.4 to 9.6). The trajectory plot from the test data shows that in the first curves of the track, the
drone approximately follows the path, oscillating frequently and changing yaw attitude in an irregular way
along the path, resulting in less precision and more time taken to accomplish the task. Furthermore, the drone
was often shut down during the test by the flag activated due to the wrong altitude estimation of the ultrasound
sensor for the oscillations in the yaw motion. Therefore, to improve the poor performances in the real-time
test, it was decided to design another version that tries to overcome the weaknesses of the first algorithm with
123
another implementation of Image processing and Path planning subsystems (2nd line-tracking algorithm
explained in chapter 8).
TABLE III: ANALYSIS OF THE TIME OF COMPLETION OF THE TRACK (1st ALGORITHM)
TIME OF COMPLETION OF THE TRACK [s]

AT LOWER SPEED AT MEDIUM SPEED AT HIGHER SPEED
“vel” = 0.00040 “vel” = 0.00045 “vel” = 0.00050
TEST 1 60 48 43
TEST 2 57 50 41
TEST 3 55 45 46
TEST 4 59 54 50
TEST 5 54 49 49
MEAN VALUE 57 49.2 45.8
STANDARD DEVIATION 2.28 2.92 3.49
Figure 9.4: HIL test with 1st algorithm – Trajectory (3D view and Top view)
124
Figure 9.5: HIL test with 1st algorithm – Motor speeds
Figure 9.6: HIL test with 1st algorithm – X, Y, Z Position and Yaw angle
125
9.2 HIL Test of the second design version
Due to the lower reliability of the first algorithm in the real-time test, the second algorithm was implemented
according to the technique explained in chapter 8. In this test, the 2nd algorithm revealed to be affected
minimally by the issues described in the first paragraph, validating the effectiveness and reliability of it
compared to the first that showed more difficulty to follow the track in a real-time test.
The next picture (9.13) shows some screenshots taken from a video that proves how the drone is capable of
following the line, and land at the end of the line on the circle satisfying the project goal.
Figure 9.7: screenshots of the HIL test with 2nd algorithm
(Link to the video: https://youtu.be/ehN1Mk-clvs)
The sensor data plots of the real-time test of the first implementation are shown in the next pictures.
126
Figure 9.8: HIL test with 2nd algorithm – Trajectory (3D view and Top view)
Figure 9.9: HIL test with 2nd algorithm – Motor speeds

127
Figure 9.10: HIL test with 2nd algorithm –X, Y, Z Positions (left) and corresponding linear speeds (right)
Figure 9.11: HIL test with 2nd algorithm –RPY angles (left) and corresponding angular velocities (right)
128
Figure 9.11 (continued): HIL test with 2nd algorithm –RPY angles (left) and corresponding angular velocities
(right)
The next table (II) shows the time measurements of 5 different tests for 3 different speeds. They are defined
by the gains corresponding to X and Y increments in the Tracking function block of the Path planning
algorithm explained in chapter 8 and written in appendix A.2.2. The higher is the gain, the higher is the speed.
Three different gains were picked. The optimal gain chosen is 0.001, and it was considered as the value in the
middle, whereas 0.006 and 0.0014 were respectively chosen as the lower and the higher values.
During the simulation test, it was already noticed that gains higher than 0.0015 lead the drone to lose the
track, for example in correspondence of right angles, because the image processing algorithm and the path
planning algorithm cannot generate the reference signals on time to make the drone slow down and turn
following the curve.

129
TABLE IV: ANALYSIS OF THE TIME OF COMPLETION OF THE TRACK (2nd ALGORITHM)
TIME OF COMPLETION OF THE TRACK [s]

AT LOWER SPEED AT MEDIUM SPEED AT HIGHER SPEED
Gains of the increments Gains of the increments Gains of the increments
in x and y position in x and y position in x and y position
coordinates = 0.0006 coordinates = 0.0010 coordinates = 0.0014
TEST 1 47 41 35
TEST 2 48 43 36
TEST 3 46 45 43
TEST 4 49 40 34
TEST 5 48 42 41
MEAN VALUE 47.6 42.2 37.8
STANDARD 1.02 1.72 3.53

DEVIATION
Looking at the mean values, we can notice that the average time of track completion decreases if the gain, and
therefore the speed, is increased, as it could be easily predicted. However, it could be more interesting to notice
that the variance and standard deviation increased with speed. It means that at higher speeds, the uncertainty
on the dynamics during the path tracking grows, and some curves require the drone to slow down to allow to
follow the path and to oscillate while positioning to center the line. The motions in these conditions take
additional time, and in some of the tests with higher gain, we can notice that the drone took the same time or
even longer than in the tests with the medium gain value. This is the reason why the medium gain value was
chosen as the optimal one for this implementation after the Hardware-in-the-loop test, and the implementation
was validated.
The next figures (9.12 to 9.15) show the results of a test with a higher gain that gave slightly different
results. In particular, the drone was less precise in tracking, having some issues at the first curve because it
130
was losing the track, but then it recovers oscillating, and it was still able to complete the track. This is the proof
of the observation made before: a higher gain leads to a higher speed but with less precision in tracking the
line.
Figure 9.12: 2nd HIL test with 2nd algorithm – Trajectory (3D view and Top view)
Figure 9.13: 2nd HIL test with 2nd algorithm – Motor speeds
131
Figure 9.14 : 2nd HIL test with 2nd algorithm – X, Y, Z Positions (left) and corresponding linear speeds (right)
132
Figure 9.15: 2nd HIL test with 2nd algorithm –RPY angles (left) and corresponding angular velocities (right).
CHAPTER 10
CONCLUSION AND RESULTS COMPARISON
This Thesis project was useful to consolidate the awareness of the importance of a Model-based design
technique in Control system applications, like the design of a drone Flight Control system, and to show its
potential and its advantages during the development process. In particular, it was useful to design and tune the
line-tracking algorithms in a more accessible way, allowing to simulate and test with real-time performances
the effectiveness and the reliability of the control system.
Model-based design method gives the idea of how in the design process, the model performances change
drastically from simulation to real-time execution, and it shows how each phase is essential to validate the
control system and to ensure the robustness of the model.
The two different implementations with their variants showed how differently the system could behave
between the simulation and real-time conditions, and especially in the latter, they revealed their weaknesses
but also their strengths. In this conclusion, a comparison will be made to summarize the project work.
The first algorithm shows the advantages of being able to follow the line by changing its yaw angle and its
orientation. Its image processing subsystem has a balanced load on the Hardware. Simple operations to detect
the track directions with an analysis of the sub-images, based on the binary pixels count and the comparison
of this sum with a certain threshold, are combined with the more complex operations of Hough transformation
and Canny method to detect the values of the angles. The path planning algorithm has an easy way to
implement the logic using State machine diagrams realized through Stateflow.
The second version of the first algorithm is characterized by even less complexity, without having the
Hough and Canny functions and having a simplified steering control that stops the drone every time it changes
the yaw angle, allowing more precision but decreasing the fluidity of movements.
133
134
However, these implementations have a weakness not to be neglected, which is highlighted in the real-time
test: they strongly rely on the altitude, and they are susceptible to its little variations. In the Model-in-the-loop
test, the hovering altitude was set as a constant, and the other parameters, like the drone speed and the sub-
images dimensions, are tuned according to the altitude value. In the Hardware-in-the-loop test, the drone may
not maintain the same altitude constant due to its estimation based on ultrasound sensors. It means the
environment has a strong influence on the performances, and if it is not set with the proper condition, the
algorithm does not work correctly, as it was proved by real-time test observation.
This issue does not affect the second algorithm implementation, whose image processing subsystem has a
more adaptive and reliable way to detect track shape and to guide the path planning algorithm. Small altitude
variations caused by the ultrasound sensors do not affect its performance significantly as in the first
implementations, and the second algorithm was realized to overcome this issue that occurs in real-time
performance. It has a different way of tracking the line, which is only based on x-y position in the space and
not on yaw angle, as needed in the first algorithm to turn to change orientation in the curves.
However, this second version is more complex, and it weighs more on the Hardware computational power.
Moreover, the second algorithm design is susceptible to the gain that controls the increments of position
coordinates, meaning that the image processing algorithm is sensitive to speed. If the gain is high, the drone
can still track the line, but it oscillates more, and it could more likely lose the position on the path if we think
that in a real-time test, the states are estimated and are not exact as in simulation.
Finally, it is possible to conclude that the design satisfies the requirements through the low-cost hardware
used in the experiments and through the final version of the line-tracking algorithm that can adequately perform
the tasks even in a noisy environment used for real-time tests.

APPENDIX A
USER-DEFINED FUNCTIONS IN MATLAB
This Appendix contains the codes of all the user-defined Matlab functions of the implemented algorithms
in Image processing and Path planning subsystems.
A.1 Image processing algorithm for the 2nd software implementation
A.1.1 “Binarization” function (using Color Thresholder App)
function BW = Binarization(R,G,B)
I = cat(3, R, G, B);
% Define thresholds for channel 1 based on histogram settings

channel1Min = 254.000;
channel1Max = 255.000;


% Create mask based on chosen histogram thresholds

sliderBW = ( (I(:,:,1) >= channel1Min) | (I(:,:,1) <= channel1Max) ) & ...
(I(:,:,2) >= channel2Min ) & (I(:,:,2) <= channel2Max) & ...
(I(:,:,3) >= channel3Min ) & (I(:,:,3) <= channel3Max);
BW = sliderBW;
end
A.1.2 “Filter” function
function BW_F = Filter(BW)
%%%% Tuning Parameters

% The lower is the Threshold, the more filtered is the image
Threshold_Filter = 3;
%%
% Output Image Dimensions
135
136
Width = 158;
Height = 118;
BW_F = zeros(Height,Width);
% Filter with a 3x3 box
for i = 2:(Height + 1)
for j = 2:(Width + 1)
BW_F(i-1,j-1) = sum(BW(i-1:i+1,j-1:j+1),'all');
if BW_F(i-1,j-1) > Threshold_Filter
BW_F(i-1,j-1) = 1;
else
BW_F(i-1,j-1) = 0;
end
end
end
end
A.1.3 “Ctrl_Radar” function
function [x_incr,y_incr,Start_End_Flag] = Ctrl_Radar(BW,xp,yp)
%%%% Tuning parameters

% Radar case size (must be odd dimensions)
% The larger is the size, the more robust is the system
box = 9;
% Detection threshold
Thrs = box*box/2;
% Search Resolution
Red_Srch = 1; %in degrees (1/360)
%%
% Image Dimensions
Width = 158;
Height = 118;
Mag_vec = (Height/2)-((box-1)/2);
Angle_detection = zeros(1,(359+1)/Red_Srch); %it has as many elements as the total

degrees (360) divided by the resolution
Angle_detection_centers = zeros(1,(359+1)/Red_Srch);
%Variable for centroid detection

Start_line = 0;
%square and centered matrix

BW2 = BW(:,20:139);
% Circular radar search

for i=0:Red_Srch:359
y_b_pos = round(Mag_vec*sin(i*pi/180) + Height/2);

if y_b_pos+((box-1)/2) > Height
y_b_pos = Height-((box-1)/2);
137
elseif (y_b_pos-((box-1)/2) < 1)

y_b_pos = 1 + ((box-1)/2);
end
x_b_pos = round(Height/2 - Mag_vec*cos(i*pi/180));

if x_b_pos+((box-1)/2) > Height
x_b_pos = Height-((box-1)/2);
elseif (x_b_pos-((box-1)/2) < 1)
x_b_pos = 1 + ((box-1)/2);
end
Angle_detection(i+1) = sum(BW2(x_b_pos-((box-1)/2):x_b_pos+((box-1)/2),...
y_b_pos-((box-1)/2):y_b_pos+((box-1)/2)),'all');
if Angle_detection(i+1) >= Thrs
Angle_detection(i+1) = 1;
else
Angle_detection(i+1) = 0;
end
end
% Search for line centers

Angle_detection_Concatenated = [Angle_detection, Angle_detection];
Off_set = 1;
while Angle_detection_Concatenated(Off_set) == 1
Off_set = Off_set + 1;
end
Angle_detection_Offset = Angle_detection_Concatenated(Off_set:Off_set+359);
White_Line = 0;
Pulse_rise = 0;
Pulse_down = 0;
if (Angle_detection_Offset(i+1) == 1)
White_Line = White_Line + 1;
Pulse_rise = 1;
if (i == 359)
Pulse_down = 1;
end
elseif (Pulse_rise == 1) && (Angle_detection_Offset(i+1) == 0)
Pulse_down = 1;
Pulse_rise = 0;
else
Pulse_rise = 0;
Pulse_down = 0;
end
if (Pulse_down == 1)
White_Line = round(White_Line/2);
Angle_detection_centers(i+1-White_Line) = 1;
White_Line = 0;
Pulse_down = 0;
Pulse_rise = 0;
end
end
% Detection of all line centers

Points_detected = 6;
138
x_c_det = 1000*ones(1,Points_detected);
y_c_det = 1000*ones(1,Points_detected);
Actual_pount_found = 1;
if (Angle_detection_centers(i+1) == 1)
x_c_det(Actual_pount_found) = Mag_vec*cos((i + Off_set)*pi/180);
y_c_det(Actual_pount_found) = Mag_vec*sin((i + Off_set)*pi/180);
Actual_pount_found = Actual_pount_found + 1;
end
if (Actual_pount_found > Points_detected)
break;
end
end
% Distance detection with respect to previous position

Dist_centers_wrt_previous = 1000*ones(1,Points_detected);
for i=1:Points_detected
Dist_centers_wrt_previous(i) = sqrt((xp - x_c_det(i))^2 +...
(yp - y_c_det(i))^2);
end
% minimum distance
[Min_dist,Indx_min] = min(Dist_centers_wrt_previous);
if (y_c_det(Indx_min) < 100) && (x_c_det(Indx_min) < 100)

y_incr = y_c_det(Indx_min);
x_incr = x_c_det(Indx_min);
else
y_incr = 0;
x_incr = 0;
end
if sum(Dist_centers_wrt_previous <= 1000) > 1

Start_End_Flag = 1;
else
Start_End_Flag = 0;
end
end
A.1.4 “Borderless” function
function BW_B_LESS = Borderless(BW_F)
% Figures that touch the edge of the image are removed

Border_size = 5;
%%
BW_B_LESS = BW_F;
% Image dimensions
Width = 158;
Height = 118;
% Top edge
BW_F(1:Border_size,1:Width) = BW_F(1:Border_size,1:Width).*-1;
% Bottom edge
139
BW_F(Height-Border_size+1:Height,1:Width) = BW_F(Height-Border_size+1:Height,1:Width).*-
1;
% Left edge
BW_F(Border_size+1:(Height-Border_size),1:Border_size) =...
BW_F(Border_size+1:(Height-Border_size),1:Border_size).*-1;
% Right edge
BW_F(Border_size+1:(Height-Border_size),Width-Border_size+1:Width) =...
BW_F(Border_size+1:(Height-Border_size),Width-Border_size+1:Width).*-1;
% Left - Right Search

for i = Border_size+1:(Height-Border_size)
for j = Border_size+1:(Width-Border_size)
if (BW_F(i,j) == 1) && ...
((BW_F(i-1,j) == -1) || ...
(BW_F(i+1,j) == -1) || ...
(BW_F(i,j-1) == -1) || ...
(BW_F(i,j+1) == -1) || ...
(BW_F(i-1,j-1) == -1) || ...
(BW_F(i-1,j+1) == -1) || ...
(BW_F(i+1,j-1) == -1) || ...
(BW_F(i+1,j+1) == -1))
BW_F(i,j) = -1;
end
end
end
% Right - Left Search

for i = (Height-Border_size):-1:Border_size+1 %decreasing
for j = (Width-Border_size):-1:Border_size+1 %decreasing
if (BW_F(i,j) == 1) && ...
((BW_F(i-1,j) == -1) || ...
(BW_F(i+1,j) == -1) || ...
(BW_F(i,j-1) == -1) || ...
(BW_F(i,j+1) == -1) || ...
(BW_F(i-1,j-1) == -1) || ...
(BW_F(i-1,j+1) == -1) || ...
(BW_F(i+1,j-1) == -1) || ...
(BW_F(i+1,j+1) == -1))
BW_F(i,j) = -1;
end
end
end
% Down - Top search

for j = (Width-Border_size):-1:Border_size+1 %decreasing
for i = (Height-Border_size):-1:Border_size+1 %decreasing
if (BW_F(i,j) == 1) && ...
((BW_F(i-1,j) == -1) || ...
(BW_F(i+1,j) == -1) || ...
(BW_F(i,j-1) == -1) || ...
(BW_F(i,j+1) == -1) || ...
(BW_F(i-1,j-1) == -1) || ...
(BW_F(i-1,j+1) == -1) || ...
(BW_F(i+1,j-1) == -1) || ...
(BW_F(i+1,j+1) == -1))
140
BW_F(i,j) = -1;
end
end
end
% Top - Down search

for j = Border_size+1:(Width-Border_size)
for i = Border_size+1:(Height-Border_size)
if (BW_F(i,j) == 1) && ...
((BW_F(i-1,j) == -1) || ...
(BW_F(i+1,j) == -1) || ...
(BW_F(i,j-1) == -1) || ...
(BW_F(i,j+1) == -1) || ...
(BW_F(i-1,j-1) == -1) || ...
(BW_F(i-1,j+1) == -1) || ...
(BW_F(i+1,j-1) == -1) || ...
(BW_F(i+1,j+1) == -1))
BW_F(i,j) = -1;
end
end
end
for i = 1:Height
for j = 1:Width
if (BW_F(i,j) == -1)
BW_B_LESS(i,j) = 0;
end
end
end
end
A.1.5 “Filter_C_Det” function
function [C_DET, BW_FF] = Filter_C_Det(BW_B_LESS)

% The lower is the threshold, the more filtered is the image
Filter_Threshold = 3;
% Pixel threshold to detect circle

C_pixels = 500;
%%
% Image dimensions
Width = 156;
Height = 116;
BW_FF = zeros(Height,Width);
for i = 2:(Height + 1)
for j = 2:(Width + 1)
BW_FF(i-1,j-1) = sum(BW_B_LESS(i-1:i+1,j-1:j+1),'all');
141
if BW_FF(i-1,j-1) > Filter_Threshold

BW_FF(i-1,j-1) = 1;
else
BW_FF(i-1,j-1) = 0;
end
end
end
% Circle detection in the center

if sum(BW_FF,'all') > C_pixels
C_DET = 1;
else
C_DET = 0;
end
end
A.1.6 “Ctrl_Circle” function
function [x_Last, y_Last, Sum_end] = Ctrl_Circle(BW_FF)
%Trying to center the circle to land in the center

% Image dimensions
Width = 156;
Height = 116;
y_inc = sum(BW_FF(:,1:(Width/2)),'all'); %upper half sum

y_dec = sum(BW_FF(:,((Width/2) + 1):Width),'all'); %lower half sum
x_inc = sum(BW_FF(((Height/2)+1):Height,:),'all'); %right half sum
x_dec = sum(BW_FF(1:(Height/2),:),'all'); %left half sum
y_Last = y_dec - y_inc;

x_Last = x_dec - x_inc;
Sum_end = y_inc + y_dec; %whole circle image sum
End
A.2 Path-planning algorithm for the 2nd software implementation
A.2.1 “Z_Ctrl” function
function [z_sp, Enab_Go] = Z_Ctrl(t_in)

% SP altitude
Zsp = -0.9;
%%
if (t_in == 1)
Enab_Go = 1;
142
else
Enab_Go = 0;
end
z_sp = Zsp;
end
A.2.2 “XY_Ctrl” function (for Tracking)
function [x_inc, y_inc] = XY_Ctrl(Trck_ena,u)

Gain_x_dir = 0.001;
Gain_y_dir = 0.001;
%%
if Trck_ena == 1
x_inc = u(1)*Gain_x_dir;
y_inc = u(2)*Gain_y_dir;
else
x_inc = 0;
y_inc = 0;
end
end
A.2.3 “XY_Ctrl” function (for precision slow movements)
function [x_inc_F, y_inc_F] = XY_Ctrl(Trig_End,U_prev,Circle_Detected,Land_End)

Gain_x_dir = 0.0005;
Gain_y_dir = 0.0005;
%%
if (Trig_End == 0) && (Circle_Detected == 0) && (Land_End == 0)
x_inc_F = U_prev(1)*Gain_x_dir;
y_inc_F = U_prev(2)*Gain_y_dir;
else
x_inc_F = 0;
y_inc_F = 0;
end
end
A.2.4 “XY_Ctrl” function (for Landing)
function [z_final, x_inc, y_inc, Landing_Kill] = XY_Ctrl(u,C_R,u_F,C_B)

143
Last_Gain = 0.00035;
Z_gain = 0.0005;
Landing_Final_Size_Circle = 3000;
%%
if C_R == 1 && (u_F(1) == 0 && u_F(2) == 0)

if C_B <= Landing_Final_Size_Circle
x_inc = u(1)*Last_Gain;
y_inc = u(2)*Last_Gain;
else
x_inc = 0;
y_inc = 0;
end
Landing_Kill = 1;
z_final = Z_gain;
else
Landing_Kill = 0;
x_inc = 0;
y_inc = 0;
z_final = 0;
end
end
A.3 Image processing for the 1st software implementation
function DegAngle = FindAngle(BWimage)
edgedBW = edge(BWimage,'canny'); %Canny;

[H, T] = hough(edgedBW); %Hough Transform;
P = houghpeaks(H,2); %peak values;
DegAngle = mean(T(P(:,2))); %found angle value;
end
CITED LITERATURE
1. PaperCraftSquare: Leonardo da Vinci's Aerial Screw Invention. http://www.papercraftsquare.com/

leonardo-da-vincis-aerial -screw-invention-free-paper-model-download.html [Online;
accessed 02/15/20].
2. Pantalone, M.: Introduzione. In Modellazione e Simulazione Di Un Quadricottero Multirotore.

Università di Bologna. 2015. https://amslaurea.unibo.it/9038/1/TESI_CORRETTA.pdf
[Online; accessed 02/17/20].
3. Wikipedia: Unmanned Aerial Vehicle. https://en.wikipedia.org/wiki/Unmanned_aerial_vehicle

4. International Business Times: Drones Aid Archaeologists in Exploring Ancient Sites, UAVs Like A
‘New Set of Eyes’ in Remote Areas. In Science. https://www.ibtimes.com/drones-aid-
archaeologists-exploring -ancient-sites-uavs-new-set-eyes-remote-areas-1575023 [Online;
accessed 02/23/20].
5. Microdrones: fast, Reliable, safe, cost-effective, uav/drone inspection tools from microdrones.
https://www.microdrones.com/en/industry-experts/inspection/ [Online; accessed 02/22/20].
6. Resonon: Airborne Remote Sensing System. https://resonon.com/airborne-remote-system [Online;

accessed 02/27/20].
7. Residence Style: The Use of Drones in Architecture Soars To New Heights. In Architect Design.
https://www.residencestyle.com/the-use-of-drones-in-architecture-soars-to-new-heights/
8. Wikipedia: Global Hawk. https://en.wikipedia.org/wiki/Northrop_Grumman_RQ4_Global_Hawk#/

media /File: Global_Hawk_1.jpg [Online; accessed 03/01/20].
9. Swissinfo: Using Swiss AI and drones to count African wildlife. https://www.swissinfo.ch/eng/

conservation_using-swiss-ai-and-drones-to-count-african-wildlife/44686308 [Online;
accessed 03/01/20].
10. Youtube: DJI - M200 Series – Search and Rescue in Extreme Environments.
https://www.youtube.com /watch?v=GkIJ2NJQHys [Online; accessed 03/02/20].
11. Vmire: Смотреть видео MIL SIL Tutorial: Part 1 на ВМире бесплатно. https://vmire.city/video/
dzgMYGWXpUc [Online; accessed 02/12/20].
12. Mathworks: Run on Target Hardware. Simulink Support package for Parrot Minidrones.
https://www.mathworks.com/help/supportpkg/parrot/run-on-target-hardware-main.html
144
145
CITED LITERATURE (continued)
13. MathWorks: MathWorks Minidrone Competitions. https://www.mathworks.com/academia/student-

competitions/minidrones.html [Online; accessed 02/10/20].
14. MathWorks: Rules and Guidelines. In MathWorks Minidrone Competitions.

https://www.mathworks.com/content/dam/mathworks/mathworks-dot-com/academia/
student-competitions/minidrone-competition/mathworks-minidrone-competition-
guidelines.pdf [Online; accessed 02/12/20].
15. Totu, L., C.: Introduction to Quadrotors and Control Theory. Udemy. https://www.udemy.com/
course/quadrotors/ [Online; accessed 02/15/20].
16. Orsag, M., Korpela, C., Oh P., Bogdan S.: Coordinate Systems and Transformations. In Aerial
Manipulation. Advances in Industrial Control. 08 2018. Springer, Cham. ISSN 1430-
9491 doi:10.1007/978-3-319-61022-1.
17. Aerodynamics for students: Blade Element Propeller Theory. Propulsion.

http://www.aerodynamics4students.com/propulsion/blade-element-propeller-theory.php
18. Drayer, A., G.: Programming Drones with Simulink. Videos and Webinars. Mathworks.
https://www.mathworks.com/videos/programming-drones-with-simulink-
1513024653640.html [Online; accessed 02/10/20].
19. Douglas, B.: Drone Simulation and Control, Part 1: Setting Up the Control Problem. Videos and
Webinars. Mathworks. https://www.mathworks.com/videos/drone-simulation-and-control-
part-1-setting-up-the-control-problem-1539323440930.html7 [Online; accessed 02/12/20].
20. Parrot: Parrot Mambo Fly. https://www.parrot.com/us/drones/parrot-mambo-fly [Online; accessed

01/25/20].
21. Mathworks: Install RNDIS for Parrot Minidrones. Simulink Support package for Parrot Minidrones.
https://it.mathworks.com/help/supportpkg/parrot/ug/intro.html [Online; accessed 02/06/20].
22. Karnavas, Y., L., and Chasiotis, I., D.: "PMDC coreless micro-motor parameters estimation
through Grey Wolf Optimizer," 2016, XXII International Conference on Electrical Machines
(ICEM), Lausanne, 2016, pp. 865-870, doi: 10.1109/ICELMACH.2016.7732627.
23. Parrot: Parrot Mambo Fly Quick Start Guide. https://www.parrot.com/files/s3fs- public/firmware/
mambo_quick-start-guide_uk-fr-sp-de-it-nl-pt-ar1.pdf [Online; accessed 02/09/20].
146
24. Educational Innovations: Deluxe Safety Glasses. https://www.teachersource.com/product/deluxe-

safety-glasses/lab-equipment-goggles [Online; accessed 02/10/20].
25. Banggod: Mini Wireless Dongle CSR 4.0 bluetooth Adapter V4.0 USB 2.0/3.0 For Win 7/8/10/XP
For Vista 32/64. https://www.banggood.com/Mini-Wireless-Dongle-CSR-4_0-Bluetooth-
Adapter-V4_0-USB-2_03_0-For-Win-7810XP-For-Vista-3264-p-1132661.html [Online;
accessed 03/14/20].
26. CDW: C2G 2m USB Cable - USB 2.0 A to Mini USB B Cable (6.6ft) - USB Cable.
https://m.cdw.com/product/c2g-2m-usb-cable-usb-2.0-a-to-mini-usb-b-cable-6.6ft-usb-
cable/328953 [Online; accessed 25/03/20].
27. Amazon: Parrot Mambo – PCB + Viti. https://www.amazon.it/Parrot-PF070237-Mambo-PCB-

viti/dp/B01MRL7UMA [Online; accessed 03/15/20].
28. Alibaba: MPU6050 IC 3 Axis Gyroscope. https://www.alibaba.com/product-detail/High-quality-

MPU6050-IC-3Axis-gyroscope_60826768961.html [Online; accessed 02/05/20].
29. InvenSense: MPU-6000 and MPU-6050 Product Specification Revision 3.4. p. 21.
https://www.invensense.com/wp-content/uploads/2015/02/MPU-6000-Datasheet1.pdf
30. Collins, D.: What are coreless DC motors? https://www.motioncontroltips.com/what-are-coreless-

dc-motors/ [Online; accessed 03/02/20].
31. Mathworks: Quadcopter Project. Aerospace Blockset. https://www.mathworks.com/help/aeroblks/

quadcopter-project.html?s_tid=srchtitle [Online; accessed 16/03/20].
32. Prouty, R.,W.: Helicopter performance, stability, and control. Malabar, Fla: Krieger Pub, 2005,
ISBN 1575242095 9781575242095.
33. Pounds, P., Mahony, R. and Corke, P.: Modelling and control of a large quad-rotor robot. In:
Control Engineering practice, pp. 691–699, Elsevier Ltd., 02 2010, ISSN: 0967-0661.
34. Douglas, B.: Drone Simulation and Control, Part 2: How Do You Get a Drone to Hover? Videos and
part-2-how-do-you-get-a-drone-to-hover--1539323448303.html [Online; accessed 02/15/20].
35. Douglas, B.: Drone Simulation and Control, Part 5: Tuning the PID Controller. Videos and
part-5-tuning-the-pid-controller-1540450868204.html [Online; accessed 02/21/20].
147
36. Douglas, B.: Drone Simulation and Control, Part 3: How to Build the Flight Code. Videos and
part-3-how-to-build-the-flight-code-1539323453258.html [Online; accessed 03/12/20].
37. Van de Maele, P., J.: Reading an IMU Without Kalman: The Complementary Filter. Pieter-Jan,
04 2013. http://www.pieter-jan.com/ [Online; accessed 05/09/20].
38. Jouybari, A., Ardalan, A., A., and Rezvani, M.-H.: experimental comparison between mahoney
and complementary sensor fusion algorithm for attitude determination by raw sensor data of
xsens imu on buoy, Int. Arch. Photogramm. Remote Sens. Spatial Inf. Sci., XLII-4/W4, 497–
502, 09 2017. https://doi.org/10.5194/isprs-archives-XLII-4-W4-497-2017.
39. Zhang, J. and Chen, G. X.: Research on the Geometric Distortion Auto-Correction Algorithm for
Image Scanned. In: Applied Mechanics and Materials pp.644–650. 09 2014. 4477–8.1
https://doi.org/10.4028/www.scientific.net/amm.644-650.4477.
40. Mathworks: Optical Flow with Parrot Minidrones. Simulink Support package for Parrot Minidrones.
Help Center. https://www.mathworks.com/help/supportpkg/parrot/ug/optical-flow-with-
parrot-minidrones.html [Online; accessed 05/22/20].
41. Douglas, B.: Drone Simulation and Control, Part 4: How to Build a Model for Simulation. Videos
and Webinars. Mathworks. https://www.mathworks.com/videos/drone-simulation-and
control-part-4-how-to-build-a-model-for-simulation-1539585112546.html [Online; accessed
02/24/20].
42. Amazon: Issuntex 10X16 ft Gray Background Muslin Backdrop,Photo Studio, Collapsible High
Density Screen for Video Photography and Television. https://www.amazon.com/gp/
product/B07NQ7VFJT/ref=ppx_yo_dt_b_asin_title_o04_s00?ie=UTF8&psc=1 [Online;
accessed 04/07/20].
43. Amazon: VATIN 4" Wide Double Faced Polyester White Satin Ribbon- 5 Yard/Spool, Perfect for
Chair Sash, Making Bow, Sewing and Wedding Bouquet. https://www.amazon.com/gp/
product/B079K8GN9B/ref=ppx_yo_dt_b_asin_title_o02_s00?ie=UTF8&psc=1 [Online;
accessed 03/12/20].
44. Mathworks: Connect a Parrot Mambo Minidrone to a Windows System Using Bluetooth.
Simulink Support package for Parrot Minidrones. https://www.mathworks.com/help/
supportpkg/parrot/ug/connect-parrot-mambo-minidrone-to-computer-using-bluetooth.html
45. Mathworks: Hough. Image processing Toolbox. Help Center. https://www.mathworks.com/help/

images/ref/hough.html [Online; accessed 04/18/20].
VITA
NAME: Paolo Ceppi
EDUCATION: Undergraduate exchange program (“SEC-U”), B. S. in Mechanical Engineering,

Vanderbilt University, Nashville, TN, USA, 2017
B.S. in Mechanical Engineering, Politecnico di Torino, Turin, Italy, 2018.
M.S. in Mechatronic Engineering, Politecnico di Torino, Turin, Italy, 2020.
M.S. in Electrical and Computer Engineering, University of Illinois at Chicago,

Chicago, IL, USA, 2020.
HONORS: “SEC-U” Program Scholarship, Politecnico di Torino and Vanderbilt University

of Nashville
“TOP-UIC” Program Scholarship, Politecnico di Torino and the University of

Illinois at Chicago
148

Ceppi Thesis 2020

Uploaded by

Copyright:

Available Formats

Ceppi Thesis 2020

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Ceppi Thesis 2020

Uploaded by

Copyright:

Available Formats

Model-based Design of a Line-tracking Algorithm for a Low-cost Mini

Drone through Vision-based Control

Igor Paprotny, Chair and Advisor

Happiness is to love an uncle like a big brother.

through which I applied the knowledge acquired during my academic studies.

useful in my future career.

optimistic and encouraged me not to give up.

tight-knit group and for supporting me during these months.

2 MODEL-BASED DESIGN AND SPECIFICATIONS…………………. 7

3 MODELING AND CONTROL OF QUADCOPTERS………………… 13

4 HARDWARE DESCRIPTION AND FIRMWARE SETUP…………... 31

5 SOFTWARE TOOLS DESCRIPTION AND MODEL

6 SIMULATION MODEL DESCRIPTION AND CONTROL

6.3.1 Simulation model…………………………………….................. 63

7 1st LINE-TRACKING ALGORITHM DESIGN…………….................. 96

8 2nd LINE-TRACKING ALGORITHM DESIGN……………................. 110

9 HARDWARE-IN-THE-LOOP TEST AND VALIDATION…………... 120

10 CONCLUSION AND RESULTS COMPARISON…………………….. 133

APPENDIX A…………………………………………………………….. 135

CITED LITERATURE…………………………………………………... 144

I TUNED GAINS OF CONTROLLERS ………………………………………. 92

1.1 “aerial screw” prototype as designed by Leonardo da Vinci [1] ……........……… 1

3.22 external force on the quadrotor [15]……………………………………………….. 27

5.14 FCS subsystem for the actuators test……………………………………………… 52

6.68 Control system simplification for Altitude controller tuning [34]………………… 89

UAV Unmanned Aerial Vehicle

RPA Remote Piloted Aircraft

DOF Degrees of Freedom

RPY Roll, Pitch, and Yaw

MIL Model in the loop

HIL Hardware in the loop

MMA Motor Mixing Algorithm

PWM Pulse Width Modulation

SONAR Sound Navigation and Ranging

IMU Inertial Measurement Unit

DMP Digital Motion Processor

RNDIS Remote Network Driver Interface Specification

PAN Personal Area network

FCS Flight Control System

RGB Red Green and Blue (Color space)

SHT Standard Hough Transform

hardware, instead of writing it manually.

made the drone model of this project.

suitable control system architecture and PID controllers tuning.

1.1. Brief History of UAV and Advantages

man's desires: flying.

Figure 1.1: “aerial screw” prototype as designed by Leonardo da Vinci [1]

or more generally a flying device, which is not piloted on board.

The considerable ease of piloting appears fundamental due to technological developments in

carried out by helicopters with human pilots.

manned helicopters were employed.

civil applications for RPAs:

Figure 1.2: An archaeologist monitoring a historical site [4]

cameras, multispectral cameras, etc.).

Figure 1.3: A drone monitoring an industrial power plant [5]

 Remote sensing. It is the technical-scientific discipline or applied science with diagnostic-investigative

Figure 1.4: Airborne Hyperspectral sensor [6]

buildings to create 3D models.

Figure 1.5: A drone monitoring the construction of a building [7]