1

Anti-collision Technologies for Unmanned Aerial

Vehicles: Recent Advances and Future Trends
Zhiqing Wei, Zeyang Meng, Meichen Lai, Huici Wu, Jiarong Han, Zhiyong Feng

Abstract
arXiv:2109.12832v3 [cs.RO] 1 Mar 2022

Unmanned aerial vehicles (UAVs) are widely applied in civil applications, such as disaster relief, agriculture
and cargo transportation, and so on. With the massive number of UAV flight activities, the anti-collision technologies
aiming to avoid the collisions between UAVs and other objects have attracted much attention. The anti-collision
technologies are of vital importance to guarantee the survivability and safety of UAVs. In this article, a compre-
hensive survey on UAV anti-collision technologies is presented. We firstly introduce laws and regulations on UAV
safety which prevent collision at the policy level. Then, the process of anti-collision technologies is reviewed from
three aspects, i.e., obstacle sensing, collision prediction, and collision avoidance. We provide detailed survey and
comparison of the methods of each aspect and analyze their pros and cons. Besides, the future trends on UAV
anti-collision technologies are presented from the perspective of fast obstacle sensing and fast wireless networking.
Finally, we summarize this article.

Index Terms

unmanned aerial vehicle, survivability and safety, anti-collision, laws and regulations, obstacle sensing, collision
prediction, collision avoidance, joint sensing and communication, AI chip, survey, review

G LOSSARY

2-D Two-Dimensional.
3-D Three-Dimensional.
6G 6th Generation Mobile Networks.
ES-A2C Experience-shared Advantaged Actor - Critic.
ACA Ant Colony Algorithm.
ACTS Airman Certificate Testing Service.
Automatic Dependent Surveillance -
ADS-B
Broadcast.
ANN Artificial Neural Network.
ARE Adaptive and Random Exploration.
ATCT Air Traffic Control Tower.

This work was supported in part by the Beijing Natural Science Foundation under Grant L192031, in part by the National Natural Science
Foundation of China under Grant 61901051, and in part by Young Elite Scientists Sponsorship Program by CAST (YESS20200283).
Zhiqing Wei, Zeyang Meng, Meichen Lai, Jiarong Han and Zhiyong Feng are with Key Laboratory of Universal Wireless Communications,
Ministry of Education, School of Information and Communication Engineering, Beijing University of Posts and Telecommunications, Beijing,
100876, China (e-mail: {weizhiqing, mengzeyang, laimeichen, hjr19991020, fengzy}@bupt.edu.cn).
H. Wu is with the National Engineering Lab for Mobile Network Technologies, Beijing University of Posts and Telecommunications
(BUPT), Beijing 100876, China, and also with Peng Cheng Laboratory, Shenzhen, China. (e-mail: dailywu@bupt.edu.cn)
Correspondence authors: Zhiqing Wei, Zhiyong Feng.
 2

CAAC Civil Aviation Administration of China.

MAC Protocol under Multi-channel Opportu-
CogMOR-MAC
nistic Reservation based on Cognitive Radio.
CNN Convolutional Neural Network.
DQN Deep Q Network.
DRL Deep Reinforcement Learning.
EM Expectation-Maximization.
FAA Federal Aviation Administration.
FLARM Flight Alarm.
Generalized Fuzzy Competitive Neural
G-FCNN
Network.
HVS Human Vision System.
IR Infrared Radiation.
ISAC Integrated Sensing and Communication.
JRC Joint Radar and Communication.
JSC Joint Sensing and Communication.
LSTM Long Short-Term Memory.
LWIR Long Wavelength Infrared.
MAC Multiple Access Control.
MIMO Multiple-Input and Multiple-Output.
Multi-Channel MAC Protocol with
MMAC-DA
Directional Antennas.
NAS United States National Airspace System.
NIR Near Infrared.
OS Operating System.
PH Pythagorean-Hodograph.
RCIS Radar-Communication Integrated System.
RNN Recurrent Neural Network.
RTT Rapidly-exploring Random Tree.
SAA Simulated Annealing Algorithm.
SAR Synthetic Aperture Radar.
STC Space-Time Coded.
TCAS Traffic Collision Avoidance System.
TSP Traveling Salesman Problem.
UAV Unmanned Aerial Vehicles.
UMSF UAV Flight Mission Scheduling Framework.

I. I NTRODUCTION
Unmanned aerial vehicles (UAVs) have wide applications in recent years. For example, UAVs are widely
applied in disaster relief, agriculture and cargo transportation, and so on. The number of UAV flight
activities has increased significantly worldwide. According to Federal Aviation Administration (FAA), 1.7
million UAVs have been registered up to November 2020 [1]. Besides, according to China Civil Aviation
Working Conference in 2020, there were more than 392,000 registered UAVs and 1.25 million hours of
 3

commercial UAV flight activities in 2019 [2]. It is also shown that the global UAV market size has reached
to 9 billion dollars in 2019. With the massive applications of UAVs, the anti-collision technologies of
UAVs have attracted much attention aiming to avoid the collisions between UAVs and other objects.
The anti-collision technologies of UAVs are of vital importance to maintain the survivability of UAVs.
The anti-collision technologies of UAVs generally consist of three procedures, namely, obstacle sensing,
collision prediction, and collision avoidance. By sensing the surrounding environment, UAV obtains the
identity and location information of obstacles, which can be used for collision prediction. During obstacle
sensing, the sensors in UAV obtain environmental information such as position and speed of obstacles. For
UAV swarm, environmental information can be shared among UAVs via the wireless network connecting
them. In the procedure of collision prediction, UAV will determine whether it will collide with the
obstacles. If the UAV predicts that the collision will occur, in the procedure of collision avoidance,
UAV will schedule the flight path to avoid the collision quickly. The anti-collision technologies of UAVs
face several challenges.
Firstly, UAVs fly in the three-dimensional (3-D) space. The complex trajectory of UAVs and the complex
positional relationship between obstacles and UAVs bring great difficulty to the design of anti-collision
algorithms. Secondly, UAVs will face more kinds of obstacles compared with ground vehicles, including
not only the objects on ground such as trees, mountains, and buildings, but also flying objects such as birds
and aircraft. Finally, most of the obstacles faced by UAVs are dynamic. The obstacle avoidance system
needs to respond quickly according to the current situation and predict the trajectories of obstacles.
The existing anti-collision technologies are mostly for ground vehicles [3]–[5]. Compared with UAVs,
the anti-collision techniques for ground vehicles have the following features. Firstly, ground vehicles
move in the two-dimensional (2-D) space, which are usually in the same plane as obstacles. Secondly,
the obstacles faced by ground vehicles are mainly trees, cars, and buildings. The kinds of obstacles of
ground vehicles are smaller than that of UAVs. Finally, for ground vehicles, the prediction of collision is
relatively simple. The existing anti-collision technologies for ground vehicles are difficult to satisfy the
requirements of the anti-collision system for UAVs. Nevertheless, some of the anti-collision technologies
for ground vehicles such as obstacle sensing methods can be enhanced and applied to UAVs. Academia
and industry have paid much attention on the anti-collision technologies of UAVs. In this article, we
summarize the existing anti-collision technologies to provide guidelines for related researches.
Regarding the research on anti-collision technologies of UAVs, there are several survey articles [6]–
[11]. Aswini et al. in [6] reviewed UAV obstacle sensing methods applied in obstacle avoidance. Ryan et
al. in [7] provided an overview of the cooperative UAV control, i.e., collision avoidance of UAV swarm.
Jenie et al. in [8] proposed a classification of collision detection and avoidance approaches for UAVs in
an integrated airspace. Chand et al. in [9] briefly introduced key technologies of sensing and collision
avoidance for UAV. Shakhatreh et al. in [10] introduced applications and challenges of UAVs in the civil
field, and studied collision avoidance methods of UAVs. Yasin et al. in [11] provided a survey of collision
avoidance systems and approaches on sensing and collision avoidance techniques for UAVs. However, the
existing literature has some limitations. None of the existing surveys have reviewed the comprehensive
anti-collision technologies of UAVs regarding procedures of sensing, prediction, and collision avoidance
methods. Besides, there are still no surveys discussing laws, regulations, and applications related to anti-
collision technologies of UAVs. Finally, the future trends of anti-collision technologies of UAVs are not
 4

analyzed in details. All these issues will be addressed in this article, which aims to provide a comprehensive
survey on anti-collision technologies of UAVs and provide the guidelines for the future trends. The
contributions of this article are as follows.
• The anti-collision technologies are classified into three procedures, i.e., obstacle sensing, collision
prediction, and collision avoidance, which are reviewed respectively in details. The related laws and
regulations to prevent anti-collision of UAVs are introduced.
• The future trends of anti-collision technology of UAV, such as fast obstacle sensing and fast wireless
networking are analyzed. We discover that joint sensing and communication (JSC), namely, integrated
sensing and communication (ISAC) technique is promising to improve the speed of sensing and
networking. Thus JSC technique is promising to improve the anti-collision capability of UAVs.
As shown in Fig. 1, the rest of this article is organized as follows. In Section II, the laws and
regulations related with aviation safety of UAVs are introduced. In Section III to Section V, the anti-
collision technologies of UAVs are reviewed. Specifically, in Section III, the obstacle sensing methods
are introduced, which consists of cooperative and non-cooperative obstacle sensing methods. Then, the
collision prediction methods are reviewed in Section IV. In Section V, the collision avoidance algorithms
are reviewed. In Section VI, the future trends on UAV anti-collision technologies are presented. Finally,
in Section VII, we summarize this article.

II. L AWS AND R EGULATIONS

Due to the severe environment and unstable control signal, UAVs have a large probability to collide
with obstacles. UAV aviation safety laws and regulations are thus established to guarantee the safety
of UAVs and objects that might be hit by UAVs. The UAV aviation safety laws and regulations can
be divided into three categories, namely, UAV flight control regulations, UAV airworthiness standards,
and pilot qualification certification regulations, where UAV flight control regulations regulate the flight
activities of UAVs, airworthiness standards provide standards for the design and production of UAVs, and
pilot qualification certificate regulations imposes the requirements for UAV operators, ensuring the flight
safety. The summary of laws and regulations of UAV is shown in Table I.

A. UAV Flight Control Regulations

In order to regulate the flight activities of UAVs, UAV flight control regulations have been established to
regulate the flight area, flight speed, and flight duration of UAVs. It reduces the probability of path conflict
between UAVs, forbids the flight behavior of UAVs in dangerous areas, thus fundamentally reduces the
probability of collision.
For the management of the flight area, the Civil Aviation Administration of China (CAAC) plans the
flight airspace of UAVs in [12], and stipulates the horizontal and vertical ranges of the flight airspace
based on the characteristics of small and light UAVs. Isolated airspace is also established to isolate UAV
flight activities from planes. However, for medium or large UAVs that have passed safety certification,
light UAVs not exceeding the safe altitude, and small UAVs with reliable monitoring and maintenance
capabilities, isolated airspace will not be reserved. Regulation 107 in the “Federal Aviation Regulations”,
which was issued by FAA, stipulates the flight range of civil small UAVs in the United States of America
(USA), and proposes that the flight activities of small UAVs must be kept within the sight of the operator.
 5

Paper Title

Sectionĉ: Introduction

Section Ċ: Laws and Regulations

UAV Flight Control Regulations
Airworthiness Standards For UAVs
Regulations For The Certification of
Pilot Qualifications

Section ċ: Obstacle Sensing Methods

Cooperative Obstacle Sensing Methods
Non-cooperative Obstacle Sensing Methods
Visual Detection
Nonvisual Detection
The Combination of Visual Detection
and Nonvisual Detection

Section Č: Collision Prediction Methods

Trajectory Fitting for Collision Prediction

Machine Learning for Collision Prediction
Section č: Collision Avoidance Methods

Classic Collision Avoidance Algorithms

Heuristic Collision Avoidance Algorithms

Section Ď: Future Trends

Fast Obstacle Sensing

Fast Wireless Networking

Section ď: Conclusion

Fig. 1. The organization of this paper

In addition, the United States National Airspace System (NAS) divides airspace into Class A, Class B,
Class C, Class D, Class E, Class G, and some special airspace, in which tower license is required for
operations in airspace Class B, C, D, and E, but not in Class G, as shown in Table II.
For countries of the European Union (EU), international flight of cargo UAVs in airspace Class A, B,
and C should take off and land at aerodromes under European Aviation Safety Agency (EASA)’s scope
[13]. In the United Kingdom (UK), Civil Aviation Authority (CAA) stipulates that the airspace division
of airspace Class A-E is not applicable to the UAVs flying at low altitude. However, prohibited areas,
restricted areas, dangerous areas and protected areas near the airport are forbidden to be entered by UAVs
[14].
 6

TABLE I
S UMMARY OF LAWS AND REGULATIONS

Category Reference One-sentence summary

Interim Regulations on Flight Management of Un-
[12]
manned Aircraft (Draft for Comment).
Administrative Measures for Air Traffic Management
UAV Flight Control [15]
of Civilian UAV Systems.
Regulations
Regulation on Dynamic Data Management of Light
[16]
and Small Civil UAV flight.
[17] Draft of UAV System Remote Identification Rules.
Airworthiness Standards for Medium and High-risk
[19]
Unmanned Helicopter Systems (Trial).
Airworthiness Standards for High-risk Fixed-wing
[20]
Cargo UAV Systems (Trial).
Airworthiness Standards [21] Large Cargo UAV airworthiness Standards.
For UAVs
Guidance on UAV Airworthiness Certification Based
[22]
on Operational Risks.
Administrative Procedures for Airworthiness Certifi-
[23]
cation of Civil Unmanned Aerial Vehicle Systems.
Regulations on the Administration of civilian UAV
Regulations For The [24]
pilots.
Certification of
Airman Certificate Testing Service (ACTS) Award
Pilot Qualifications [25]
Notice.

TABLE II
D EFINITION OF THE AIRSPACE CLASSES [27]

Classes Definition
Class A includes the en route and high altitude environment of aircraft from
Class A
one area to another in the identical country.
Class B airspace is defined at 29 high-density airports in the United States,
Class B
aiming to manage the air traffic activities around the airfield.
Class C airspace is defined around the airport with airport traffic control
Class C
tower and radar approach control.
Class D airspace is under the jurisdiction of a local Air Traffic Control Tower
Class D
(ATCT).
Class E completely separates aircraft from other aircraft, aiming to provide
Class E
air traffic services.
Class G airspace is defined as airspace not designated as class A, class B,
Class G
class C, class D or class E.

Regarding the supervision of UAVs, [15] promulgated by CAAC in 2016 puts forward requirements
on the flight space and flight conditions of UAVs to prevent them from affecting civil aviation. The
“Regulation on Dynamic Data Management of Light and Small Civil UAV Flight” promulgated in 2019
points out that light UAVs, small UAVs, and plant protection UAVs in the airspace should be monitored.
 7

The regulation requires them to transmit dynamic data including flight code, duration of flight, position,
speed, trajectory, and so on., in a certain format in real time for UAV management and research [16].
FAA issued “Draft of UAV System Remote Identification Rules” in December 2019 [17]. The rules
stipulate that UAVs need to provide a third-party system with the order number, position, and other
information during the flight, so as to ensure the flight safety. CAA stipulates that UAVs in the UK
must have remote ID system, which can broadcast status and identity data including the verification code
provided by CAA, the unique serial number of UAV, geographical location, flight route, and so on [18].

B. Airworthiness Standards For UAVs

The airworthiness standard of UAV specifies the performance requirements of UAV flight system, which
prevents the UAV from collision due to the substandard performance. CAAC has issued various types
of UAV airworthiness standards in “The Airworthiness Standard of Medium and High Risk Unmanned
Helicopter System (Trial)” [19], “The Airworthiness Standard of High Risk Cargo Fixed Wing UAVs
(Trial)” [20] and “The Airworthiness Standard of Large Cargo UAVs” [21]. They stipulate the relevant
UAV system performance requirements, including take-off, climb and glide performance, maneuverability,
and so on. In addition, CAAC promulgated the “Guidance on UAV Airworthiness Certification Based on
Operational Risks” [22] , which puts forward different airworthiness management modes according to the
risk level of UAV operation scenarios. The “Risk Assessment Guide for Civil UAV System Airworthiness
Certification Project (Draft for Comments)” issued in 2020 proposes the authorization principles of risk
system and product risk assessment. The product risk is based on the energy level of civil UAVs and the
collision possibility level in the operating environment to form the risk matrix, which is applied to reveal
the risk level of civil UAVs in different scenarios [23].
FAA does not require small UAVs to meet current airworthiness standards or obtain aircraft certification.
To ensure the normal operation of the safety system and the communication link between control station
and UAV, the operator needs to carry out relevant inspections before flight, but the UAV must be registered.
CAA suggests that large UAVs must have a valid certificate of airworthiness and small UAVs only need
to have a permit to fly. Meanwhile, it is necessary to conduct safety assessment on UAV to ensure the
reliability of UAV mission. Relevant laws are still being formulated [18].

C. Regulations for The Certification of Pilot Qualifications

In order to ensure the flight safety of UAVs, some laws and regulations stipulate that the UAV operators
must acquire a license. CAAC passed the “Regulations on the Management of Civil UAV Pilots” in 2019.
For the civil UAVs with weight of more than 7kg, the pilot must obtain the driving license when satisfying
the relevant standards [24]. FAA has also regulated that if a pilot wants to operate a small UAV, the pilot
needs to have a license or be directly supervised by a person with license, and the pilot must be at least 16
years old to acquire a pilot license [25]. EASA stipulates that for recreational or low-risk UAV activities,
the operator only needs to comply with safe operation requirements. For civil UAV activities with high
risks, operators need to obtain an operational authorization from the national competent authority before
starting the operation. For future high-risk activities such as carrying passengers by UAV, the safety
certification of the UAV operator and its UAV, as well as the licensing of the remote pilots are required
to ensure safety [26].
 8

TABLE III
S UMMARY OF COOPERATIVE OBSTACLE SENSING METHODS

Methods Reference One-sentence summary

Traffic Collision Avoidance System, a set of anti-collision systems
TCAS [28]
installed in medium and large aircraft.
Automatic Dependent Surveillance – Broadcast, a system transforming
ADS-B [29] segments of aviation, which can effectively improve the cooperation
between aircraft and enhance the performance of TCAS.
Flight Alarm, a traffic awareness and anti-collision technology for small
FLARM [30]
UAVs with human operators on the ground.

In the anti-collision techniques for UAVs, the sensing for obstacles, the collision prediction and collision
avoidance methods are essential. In the following sections, the three techniques are introduced respectively.

III. O BSTACLE S ENSING M ETHODS

Obstacle sensing is the first step of the anti-collision process for UAVs, through which UAV obtains
the awareness of the surrounding environment, estimates the locations of obstacles, and provides prior
information for anti-collision techniques of UAVs.
There are two kinds of obstacles. The first kind of obstacles are environmental objects on the ground,
such as trees, buildings, and mountains, which are usually static. The second kind of obstacles are flying
objects such as birds and other aircraft, which are usually dynamic. Due to the complexity of the UAV
trajectory, there is a high risk of collision between multiple UAVs if there are no anti-collision systems
equipped on UAVs [9].
There are cooperative and non-cooperative obstacle sensing methods. For cooperative sensing methods,
UAVs share their state information or the information of surrounding obstacles. For non-cooperative
sensing methods, UAVs relies on their own sensors to detect the obstacles without the participation of the
communication process.

A. Cooperative Obstacle Sensing Methods

Cooperative obstacle sensing system has been widely used in aircraft, and is gradually applied to UAV
in recent years. For example, the Matrice 200 series and Mavic 2 Enterprise UAVs released by DJI have
installed ADS-B system and can detect the ADS-B signals of aircraft miles away, then warns the aircraft
if there is a premonition of collision. In this section, we introduce and classify the existing cooperative
obstacle sensing methods and discuss the main challenges of them when applying to UAVs. There are
mainly three types of cooperative anti-collision systems: Traffic Collision Avoidance System (TCAS),
Automatic Dependent Surveillance – Broadcast system (ADS-B) and Flight Alarm (FLARM). TCAS and
ADS-B are generally used for large and medium UAVs, while FLARM is used for small UAVs. The
summary of cooperative obstacle sensing methods is shown in Table III.
TCAS is a set of anti-collision system installed in medium and large aircraft. At present, TCAS has
become the standard equipment on the newly produced medium and large aircraft. The function of this
 9

system is to send inquiry signals to neighboring aircraft, and obtain the altitude, heading, and other data
of the invading aircraft through the response of the onboard transponder system of the invading aircraft
[28]. Through data analysis, the TCAS system determines the threat level of the invading aircraft. If there
is a potential threat, the TCAS system will issue advisory prompts or vertical maneuver instructions to
the pilot, which guide the pilot to avoid conflict with the invading aircraft.
ADS-B technology is widely applied in civil aviation. The shared information in ADS-B is the position
information of aircraft and the information of conflict warning, track angle, route inflection point, and
so on, as well as aircraft classification and identification information [29]. Compared with TCAS, the
location report of ADS-B is spontaneously broadcast. Thus, the position report of the approaching UAV
can be received and processed without inquiries between UAVs.
FLARM is a traffic awareness and anti-collision technology for small UAVs with human operators
on the ground. The FLARM system obtains its own real-time position and altitude information through
a built-in high-sensitivity GPS receiver and altimeter. Combined with speed and position information,
FLARM can calculate an accurate predicted flight path. The path information is then broadcast to nearby
aircraft. Meanwhile, the FLARM system receives the flight path information sent by the FLARM systems
of nearby aircraft. If a collision is predicted, the FLARM system will send the warning information, as
well as the direction and altitude difference information of the invading aircraft, to the connected FLARM
information display of operator’s device. When receiving the warning information, the operators can take
corresponding actions to avoid possible collisions [30].
The existing cooperative anti-collision system can accurately detect the distance, altitude, and other
information of the aircraft, and the detection distance is long. However, it can only detect aircraft
installed with the same anti-collision system, and cannot detect obstacles such as mountains or trees
in the environment.

B. Non-cooperative Obstacle Sensing Methods

The non-cooperative sensing system obtains the location information of obstacles in the surrounding
environment through sensors. According to the types of sensors, the non-cooperative obstacle sensing
methods are classified into visual detection, non-visual detection, and the combination of them. The visual
detection system includes monocular vision system and stereo vision system. The non-visual detection
system consists of the systems using active sensors including radar, laser, and the systems using passive
sensors including thermal imaging, optoelectronics, and infrared. The fusion of visual and non-visual
detection methods is also a promising approach to improve the accuracy and widen the scenarios of
obstacle sensing. The summary of non-cooperative obstacle sensing methods is shown in Table IV.
1) Visual Detection: There are generally two kinds of visual detection system, namely, monocular
vision system and stereo vision system.
Monocular vision system employs a single camera to collect the obstacle information, which consists
of three kinds of methods according to the application scenario.
The first kind of methods can only detect specific targets, which use the prior knowledge of targets
to recognize obstacles from the background image. When the number of obstacles is huge or there are
obstacles that are not consistent with the prior information, the accuracy of detection is poor. Therefore
this kind of method is not widely used [31].
 10

TABLE IV
S UMMARY OF METHODS OF NON - COOPERATIVE OBSTACLE SENSING METHODS

Category Methods Reference One-sentence summary

Research progress of obstacle detection based on
[31]
monocular vision.
Monocular Proposed a special attention mechanism to distin-
[32]
vision guish salient obstacles.
Applied Gunnar-Farneback method to estimate
[33] the optical flow between two consecutive image
Visual detection frames.
Designed a system based on binocular vision sen-
[34] sors for the detection and obstacle avoidance of
Binocular stereo micro-UAV in indoor environment.
vision Applied binocular stereo vision combined with
[35]
IFDS path planning.
[36] Applied six 4K cameras to build 3-D maps.
Proposed a method to detect objects using ultra-
[38]
sonic sensors onboard.
Proposed a low cost radar solution based on the
Active type [39]
coherent MIMO principles.
Proposed a small-sized radar design according to
Non-visual [40]
the low-altitude high-speed environment of UAVs.
detection
Proposed a system using IR and ultrasonic sensors
[43] which is capable of avoiding obstacles like walls
Passive type and people.
Applied LWIR cameras combined with NIR laser
[44]
detection.
Applied monocular camera to estimate speed and
[45]
laser radar to help avoid obstacles.
Combination Combination Applied infrared light source to assist vision sys-
[46]
tem.
[47] Combined visual cameras with ultrasonic sensors.

The second kind of methods are inspired by the automatic focusing technology, which detects obstacles
by focus/defocus or by changing the focal length. For example, [32] proposed an obstacle sensing method
inspired by human vision system (HVS). [32] adopted discrete cosine transform to recognize areas of
prominent objects, so that UAVs can detect the obstacles in sight. This kind of method can simultaneously
recognize and localize the obstacles. However, it will make mistakes when multiple obstacles co-exist,
and the zooming ability of cameras should be high.
The third kind of methods detect obstacles by image difference or optical flow caused by the mobility
of targets, which can only detect moving targets. This kind of methods are also called motion parallax
methods. Motion parallax refers to the phenomenon that the moving velocity of the target seems to be
increasing with the decrease of the distance between observer and the target. As illustrated in Fig. 2, when
 11

v2

v1

v0

Fig. 2. The velocity of nearby trees v1 seems to be larger than that of distant mountains v2 in the scenery outside the windows

Fig. 3. The disadvantage of the monocular vision

taking train, the trees near the window seem to move faster than the mountains far away. One of the motion
parallax methods is to take advantage of the optical flow [48]. [33] applied Gunnar-Farneback method
to estimate the optical flow between two consecutive image frames. If the magnitude of the optical flow
vector is smaller than a threshold, there are no obstacles in front of UAV. Meanwhile, utilizing the state
of UAV’s quadrotor, the misjudgment caused by camera movement can be eliminated and the detection
accuracy is improved. However, the disadvantage of motion parallax methods is that the processing time
is long because multiple images of different times need to be processed.
Generally, the accuracy of monocular vision method is low, which is difficult to be overcome by
algorithms. As illustrated in Fig. 3, the volume and distance of the objects with the same area in two-
dimensional (2-D) image of monocular vision system may be different. Thus, monocular vision systems
are difficult to provide accurate information about distance and volume of the obstacle for collision
avoidance. Therefore, there are more and more literatures on stereo vision system. The stereo vision
system applies multiple cameras to get pictures from different angles, where the system with two cameras
is called binocular vision system. With the assistance of the pictures from two cameras, the size and
location of obstacles can be estimated. [34] designed a binocular vision system in a small UAV. With
distortion correction and binocular calibration of the pictures from two cameras, block matching algorithm
is applied to obtain the disparity map, which reveals the 3-D depth information of target. The target was
then identified from the disparity map.
In [35], the UAV with binocular vision system was applied in the power lines inspection scenario. They
calculated the parallax value matrix through left and right images from two cameras and used the matrix
along with camera parameters to obtain relative 3-D coordinates of targets.
Besides the binocular vision system, multiple cameras are installed in UAV to improve the accuracy
of obstacle detection. Skydio UAV [36] installs six 4K cameras to construct a 3-D map of surrounding
environment including trees, people, buildings, and so on. Pictures from multiple angles enable UAVs to
 12

accurately obtain the location and distance of obstacles in all directions.

In addition to the advantage in accuracy of obstacle detection, stereo vision system has a large field
of vision because of the application of multiple cameras, which gather comprehensive environment
information [49]. However, the complexity of hardware and algorithms of stereo vision system brings
high cost and large power consumption, which is the main challenge for the application of stereo vision
system in low-cost civilian UAVs.
Because of the mature algorithms and relatively low cost, visual detection systems are widely adopted
by UAV manufacturers [50]. However, visual detection requires a good sight. Darkness, bad weather, or
reflection from water surface like peaceful water will bring great challenges for the information acquisition
in visual detection systems.
2) Non-visual Detection: Non-visual detection sensors consist of active sensors and passive sensors.
The active sensors, including radar, laser, and ultrasonic sensors, detect obstacles by transmitting signals
to targets, receiving and analyzing echo signals. Laser has a large detect range, and can obtain not only
distance, but also azimuth information. Besides, it has a great penetrating power when facing smoke. UAVs
of Walkera [37] apply laser for sensing, which realized a maximum sensing range of 40 m. However, laser
is greatly affected by the bad weather. On the contrary, ultrasonic sensors are not easy to be disturbed
by smoke or gas. [38] proposed a method for UAVs to detect obstacles on the ground using ultrasonic
sensors. The authors confirmed that it is feasible to detect wall by ultrasonic sensors, with error of about
40 mm.
Radar provides the sensing information of ranging, direction, and closing speed [39], and the sensing
performance is stable regardless of the surrounding environment. Therefore, among the active sensors,
radar is widely applied in UAVs for obstacle detection. [39] designed a radar system based on multiple-
input and multiple-output (MIMO), which achieves accurate angle estimation and large detection range.
[40] proposed a small-sized radar design according to the low-altitude high-speed environment of UAVs,
whose successful detection probability is more than 90% on the condition that the maximum speed of
UAV is 440 km/h and the minimum detection range is 2.8 km. In the industry, DJI proposed agriculture
solution package containing new flight controller and radar sensing system [41]. The system equips high-
precision radar which is able to detect obstacles behind or in front of the UAV at a range of 1.5 to 30
m. The radar cannot detect objects on the left or the right, because plant protection UAVs don’t need to
make a turn frequently. The agricultural UAVs of DJI [41] and XAG [42] apply radar for obstacle sensing,
which can sense obstacles in the range of 1.5 30 m, whose measurement accuracy is smaller than 10
cm.
The passive sensors consist of infrared radiation (IR) sensors, thermal imaging cameras, electro optical
sensors, and so on. Unlike active sensors, passive sensors don’t transmit signals by themselves, but
collecting signals emitted from the objects. For example, the infrared radiation sensors can detect targets
by the emitted heat. The advantage of passive sensors is that they are usually much cheaper than active
sensors. However, the signals gathered by passive sensors usually contain much noise [43]. [43] proposed
a system using IR and ultrasonic sensors which is capable of avoiding obstacles like walls and people,
and proved that cheap sensors can also achieve efficient detection. In fact, IR sensors perform even better
compared with ultrasonic sensors because the ultrasonic sensors cannot detect human bodies reliably [43].
However, for practical applications, passive sensors are not usually applied alone in the anti-collision
system. On the contrary, they cooperate with other sensors like radars to detect obstacles. For example,
 13

TABLE V
P ROS AND CONS OF OBSTACLE SENSING METHODS

Name of the methods Pros Cons Common features

Cooperative
TCAS • Can only be used for
Obstacle
ADS-B • High accuracy. UAVs equipped with the -
Sensing
FLARM same system.
Methods
Monocular
• Affected by darkness or
Vision • Low cost. • The accuracy is
strong light.
System related to the
Visual performance of the
Detection Stereo • Improved accuracy • Affected by darkness or image processing
Vision compared with strong light. algorithm.
System monocular vision • High cost.
system.

• Radar: Sensitive to
• Radar: Great detection
electromagnetic interfer-
range.
Active ence.
Non-visual Sensors • Laser: High resolution. • Laser: High cost. • The accuracy is
Detection • Ultrasonic related to the
• Ultrasonic Sensors: Sensors: Sensitive quality of sensors.
Low cost. to the meteorologic
environment.
Passive
• Low cost. • Low accuracy.
Sensors

[44] applied a long wavelength infrared (LWIR) camera to achieve a fast detection of a thermal object,
and more precise detection was implemented by the Near Infrared (NIR) laser detector.
To sum up, compared with visual detection, non-visual detection can also be effective of sensing when
the field of vision is poor. However, they have lower accuracy than the visual detection. Therefore, plenty
of UAVs nowadays choose to apply the combination of visual detection and non-visual detection.
3) The Combination of Visual Detection and Non-visual Detection: The combination of visual and
non-visual sensors could exploit both of their advantages. When designing the UAV indoor navigation
system, [45] implemented a system with monocular camera and laser rangefinder radar. Experiment results
show that the estimation error of location is about 0.5 m, and the error of speed estimation is 0.2 m/s.
Companies like XAG apply binocular vision system to sense the environment information and apply
infrared light sensors to detect obstacles at night [46]. The Guidance system of DJI [47] combines stereo
vision system and ultrasonic sensors for sensing, which can detect obstacles with a speed of 0 ∼ 16 m/s
at 0.2 ∼ 20 m. It has a speed measurement accuracy of 0.04 m/s and a positioning accuracy of 0.05 m.
The pros and cons for collision avoidance methods are shown in Table V.
 14

IV. C OLLISION P REDICTION M ETHODS

Collision prediction is the second step of anti-collision process. UAV processes the information obtained
from obstacle sensing, predicts whether there will be collision, and judges whether it needs to avoid
collision according to the prediction results. The obstacles studied in collision prediction are mostly
dynamic obstacles such as birds or other UAVs [51]–[58], because the collision probabilities of stationary
obstacles are easy to estimate.
There are many ways for the classification of UAV collision prediction methods. According to the
types of obstacles, the collision prediction methods can be classified into collision prediction between
UAV and obstacles, and collision prediction between UAVs. According to the priori information required
for prediction, the collision prediction methods can be classified into the trajectory prediction of obstacles,
the trajectory prediction of UAVs, the prediction of collision probability, and so on.
In this article, according to the prediction algorithms, the collision prediction methods are classified
into trajectory fitting methods and machine learning methods.
The trajectory fitting methods apply criteria such as the least-square criterion to fit trajectory points
into a predetermined form of expression to predict the future trajectory [51]–[54]. Since the form of the
fitting function is fixed, such methods are difficult to predict precisely in complex environment of UAV.
However, this method has low computational overhead and is suitable to low-cost UAVs.
The machine learning methods apply machine learning algorithms such as Recurrent Neural Network
(RNN) and reinforcement learning to learn the previous motion information and output parameters such as
future trajectory or collision probability [55]–[57]. Generally speaking, the machine learning methods can
better adapt to the complex flight environment of UAV since nonlinear problems can be better solved by
machine learning. However, machine learning methods will increase the energy consumption on calculation
and reduce the endurance of UAV.

A. Trajectory Fitting for Collision Prediction

The trajectory fitting can be applied to predict the flight path of UAV according to the kinematics and
dynamic features of UAV.
[51] proposed a UAV collision prediction method based on sliding window polynomial fitting algorithm.
Firstly, the sliding window polynomial fitting method is applied to predict the obstacles’ trajectory. Then,
flight information of UAVs and obstacles is used to predict the collision. The algorithm needs a small
number of historical data to construct the polynomial fitting equation, so as to realize the real-time collision
prediction. This algorithm has been widely applied in the field of trajectory prediction [52].
[53] proposed the Pythagorean-Hodograph (PH) and eight Bernstein-Bezier polynomials based method
to predict the trajectory of UAV swarm for collision prediction. Each UAV established its trajectory based
on the five supporting points of the leading UAV, predicted the trajectories of the UAV swarm, and
avoided collision by fitting the trajectory curves of the UAV swarm. [54] proposed a UAV path planning
method based on an efficient quadratic Bezier curve. The algorithm used the starting point, end point and
intermediate control points to form a quadratic Bezier curve to generate the trajectory prediction of the
UAV, which can properly change the derived intermediate points to control the trajectory of UAV to avoid
collisions with identified obstacles.
 15

B. Machine Learning for Collision Prediction

UAVs estimate the movement of dynamic obstacles based on machine learning, so that UAVs can make
collision prediction.
[55] proposed a Long Short-Term Memory (LSTM) RNN based collision prediction method. This
algorithm predicts the movement state of obstacles based on the most recent observations, which can
greatly improve the prediction accuracy.
[56] proposed a new model-based reinforcement learning algorithm TEXPLORE, which can predict
the trajectory of UAVs in unknown or uncertain environments to reduce the probability of collisions.
Combining Deep Reinforcement Learning (DRL), [57] proposed a DRL assisted method to train a deep
Q network (DQN) to predict the trajectory of obstacles and thus fulfill collision prediction.
The above collision prediction methods using machine learning are based on simulation environments, in
which the dynamic obstacles are often assumed to be movable points, which is not appropriate for practical
scenarios. Therefore, [58] proposed a collision avoidance algorithm where neural network pipeline (NNP)
is applied to predict the possible collision. Besides, an object motion estimator (OME) using optical flow
is proposed to identify dynamic objects and estimate their trajectories in video streams. The algorithm
has high accuracy (about 2%) and fast processing speed.

V. C OLLISION AVOIDANCE M ETHODS

Collision avoidance is the core step of UAV anti-collision technology. UAVs process the prior infor-
mation obtained by obstacle sensing and collision prediction to guarantee the safe flight path without
collision. Generally speaking, collision avoidance problem is equivalent to path planning problem with
obstacles in the flight path, because both of their purposes are to make UAV avoid obstacles with certain
prior knowledge. Generally speaking, collision avoidance methods can be classified into two categories:
classic methods and heuristic methods.

A. Classic Collision Avoidance Algorithms

Classic collision avoidance algorithms of UAVs are developed from the robot obstacle avoidance
algorithms, which is relatively mature. This kind of algorithm consists of geometric methods, graph
theory methods, artificial potential field methods, and methods based on path selection. The summary of
classic collision avoidance algorithms is shown in Table VI.
1) Geometric Methods: Geometric method is an effective method to avoid collision, which takes
advantage of relative kinematic geometric relationships between UAVs and obstacles.
(1) Collision Cone
As illustrated in Fig. 4, the collision cone is formed by a set of tangent lines from the UAV to the
risk area, where the risk area is a sphere area around the obstacle. If the velocity vector of the UAV is
detected within the collision cone, the UAV will fly out of the collision cone as fast as possible to prevent
collision according to the collision cone algorithm.
Specifically, for stationary obstacles, [59] applied collision cone algorithm and established a collision
cone when obstacles appear in the flight path of UAV. When a collision is about to occur, the destination
of flight path is temporally set as the closest tangent point. Then, the UAV will fly along the tangent line
through this tangent point. When the UAV reaches this point, the destination of the UAV will be reset to
 16

TABLE VI
S UMMARY OF CLASSIC COLLISION AVOIDANCE ALGORITHMS

Category Methods Reference One-sentence summary

Chose the direction to make a turn by the
[59] angle between the direction vector and the
Collision cone tangential lines of the obstacle.
Found the best aiming point of the colli-
[60]
sion cone.
Geometric Extended the collision cone to 3-D sce-
[61]
methods nario.
Applied the Dubins path algorithm to get
[62]
all the paths and find the shortest one.
Applied the Dubins path algorithm on
Dubins path [63]
ground obstacle avoidance.
Extended the Dubins path algorithm on
[64]
avoiding dynamic obstacles.
Presented the construction of the Voronoi
[65] graph search algorithm and estimated its
Voronoi graph performance.
Graph theory search algorithm
An improvement and application of the
methods [66]
Voronoi graph search algorithm.
Applied the Laguerre graph search algo-
Laguerre graph [67] rithm to plan the flight path before a flight
search algorithm mission.
Applied the potential field on the UAV
[68] collision avoidance and analyzed its per-
formance
Improved the algorithm by changing the
[69] repulsive potential function according to
Artificial Potential Artificial Potential the detection of the field.
Field Field Added weights on different obstacles and
[70] imported virtual targets in points of local
minimum problems.
Introduced a destination switching scheme
[71] in the traditional artificial potential field
method.
Proposed an improved RRT algorithm
[72]
Path Selection Path Selection which shorten the time of path selection.
based Methods based Methods Proposed an optimization based online
[73] collision avoidance algorithm for Internet
of Drones (IOD) formation.

the original destination. The UAV will not enter the collision cone again due to the inertia, so that it can
bypass the obstacle.
 17

Velocity
Velo
locity vector
vecto

Collision cone Obstacle

UAV

Fig. 4. Collision cone

Dubins
path

Turning
radius

Fig. 5. An example of Dubins path

For dynamic obstacles, [60] constructed a 2-D collision cone and searches for the minimum angle θd
between the relative velocity vector of the UAV and the hypotenuses of the cone. The direction of the
UAV will be rotated immediately with an angle greater than θd to ensure that the UAV flies out of the
collision cone. [61] extended the collision cone method to 3-D scenario. A pyramid collision cone is
established to avoid cube obstacles. The flight test is taken indoors with static obstacles in the flight path,
which shows that the algorithm can successfully avoid static obstacles in 3-D space.
Collision cone is usually applied to deal with the obstacles nearby and can be applied in the scenarios
without too many obstacles. Besides, due to the low computational complexity and high real-time perfor-
mance, the collision cone methods have good performance on collision avoidance of dynamic obstacles.
(2) Dubins Path
The Dubins path is the shortest path between two positions considering the turning radius and the speed
direction of vehicles, which was proposed by Dubins [74]. Because of the turning radius, the Dubins path
is usually constructed by a series of arcs and the tangents between the arcs, as illustrated in Fig. 5. The
Dubins path is widely used in collision avoidance algorithms of UAVs. Considering 2-D path planning
problem of UAV, [62] constructed Dubins paths between the vertices of the polygon envelopes of the
obstacles in the environment. By flying along the Dubins paths, UAV can reach the destination without
collision. The path planning problem without collision is thus transformed into the shortest path problem
of an undirected graph, which can be solved using Dijkstra’s algorithm.
Instead of concentrating on the targets in the air, [63] applied Dubins path algorithm on ground obstacle
avoidance and carried out the software-in-the-loop experiment. The results show that the flight path
deviation of UAV at 10 m/s is lower than that at 15 m/s, because Dubins paths have a large number
of arc turning paths and UAV is more difficult to control accurately when passing these paths at high
 18

0.9

0.8

0.7

0.6

0.5

0.4

0.3

0.2

0.1

0
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9

Fig. 6. A Voronoi graph, where points represent obstacles and lines represent possible flight paths

speed.
One of the disadvantages of the Dubins path algorithm is that it is difficult to be applied in the
scenarios with dynamic obstacles. Therefore, [64] applied a variation of Rapidly-exploring Random Tree
(RRT) in the collision avoidance algorithm using 3-D Dubins path. When the current path is predicted to
be collided with the updated obstacle position, the algorithm will update the path according to the new
obstacle position, so as to avoid dynamic obstacles.
2) Graph Theory Methods: The graph theory based collision avoidance methods consist of Voronoi
graph search algorithm, Laguerre graph search algorithm, and so on. This kind of methods divide the
space into a graph and traverse the graph to find the best path from source to destination. The graph
theory methods are difficult to be applied to avoid collision with dynamic obstacles, because the global
graph needs to be reconstructed after the change of the positions of obstacles, which brings huge time
consumption.
(1) Voronoi Graph Search Algorithm
As illustrated in Fig. 6, each edge in the Voronoi graph is a vertical bisector of two points. The Voronoi
graph search algorithm for collision avoidance regards obstacles as points, regards edges of the Voronoi
graph as flight paths, and applies algorithms such as Dijkstra algorithm to search for the shortest path from
source to destination. The algorithm was widely used by robots and is applied in UAVs’ path planning
recently. [65] designed the Voronoi graph search algorithm for UAVs and estimated its performance. It
shows that the time consumption of the algorithm mainly comes from the establishment and smoothing
of the Voronoi graph, and will increase sharply when the number of obstacles increases.
Furthermore, [66] improved the Dijkstra algorithm after the Voronoi graph was constructed. When the
destination changes, the algorithm can update the flight path in time, which improves the applicability of
the Voronoi graph search algorithm.
(2) Laguerre Graph Search Algorithm
The Laguerre graph is a variation of Voronoi graph. The distance from each edge to the adjacent points
is no longer the same, but proportional to the weight of the adjacent points. When applied for collision
avoidance, the weight of the points is set by the radius of circumscribed circle of obstacles, which avoids
the disadvantage of Voronoi graph that the flight path may cross the obstacles when the obstacles are too
large or too close.
 19

Similar to the Voronoi graph search algorithm, the Laguerre graph search algorithm first constructs the
Laguerre graph, and then applies shortest path algorithms on the graph. For example, [67] applied this
algorithm in path planning before a flight mission. The irregular obstacles near the UAV is simplified as
circles, which reduces the complexity of the algorithm. Meanwhile, a path optimization algorithm was
proposed to make the flight path smooth. [67] showed that the Laguerre search graph algorithm applied in
path planning is both fast and accurate. However, simplifying the obstacles into circles may make some
of them intersect, which will leave no space for a potential flight path, so that a longer flight path will
be generated.
3) Artificial Potential Field: Using the potential function, the artificial potential field algorithm estab-
lishes a repulsive force field for each obstacle. With the guidance of force fields, the UAVs avoid the
obstacles and reach the destinations.
[68] applied the Gaussian mixture model to build the physical model of obstacles, and applied
expectation-maximization (EM) algorithm to modify the potential fields by obstacles’ information. By
updating the Gaussian mixture model of the obstacles, the method proposed in [68] can be used with
dynamic obstacles. However, the potential field algorithm easily generates the local minimum solution
in complex environment. Facing such dilemma, [69] proposed a new repulsive potential function, which
added an adjustment factor m, so that the optimal flight path can be obtained by adjusting m. [70] added
weights on different obstacles to represent different levels of threat. To overcome local minimum problem,
[70] placed virtual obstacles in the points with possible local minimum solutions. Considering obstacle
avoidance problem in a UAV formation, [71] introduced a destination switching scheme in the traditional
artificial potential field method. The destinations of UAVs are switched when the distance between UAVs
is too small, which avoids local optimization points caused by nearby UAVs. The flight test is taken
outdoors, which verifies the rationality of the proposed algorithm.
4) Path Selection based Methods: For path selection based methods, the paths are firstly randomly gen-
erated between the UAV and the destination. The obstacle-free path is then selected by RRT, optimization,
or other algorithms. [72] proposed an improved RRT algorithm, which shortens the time of path selection
of the traditional RRT algorithm. Both stationary and dynamic obstacles are considered, where dynamic
obstacles are represented by reachable sets, i.e. the areas that may be reached in a period of time. The
problem of avoiding dynamic obstacles is transformed into avoiding stationary reachable sets. However,
only obstacles moving with constant velocity can be represented by the reachable sets, so the proposed
method lacks the ability to deal with emergencies. [73] proposed an online collision avoidance algorithm
for Internet of Drones (IOD) formation. The total flight time is divided into small time slots. In each time
slot, multiple paths are generated randomly and the optimal path is selected by gradient optimization and
various constrains. The proposed algorithm can deal with dynamic obstacles with changeable velocity,
because it only relies on instantaneous locations of obstacles in each time slot.

B. Heuristic Collision Avoidance Algorithms

Compared with the classical methods searching the optimal solution, heuristic algorithms aim to find
the appropriate solution according to historical experience. The solution of heuristic algorithm is probably
approximately optimal. Heuristic collision avoidance algorithms consist of neural network algorithm,
reinforcement learning, genetic algorithm, simulated annealing algorithm, colony algorithm, and so on.
The summary of heuristic collision avoidance algorithms is shown in Table VII and VIII.
 20

TABLE VII
S UMMARY OF HEURISTIC COLLISION AVOIDANCE ALGORITHMS (1)

Category Reference One-sentence summary

[75] Proposed a path planning algorithm based on ANN.
[76] Applied G-FCNN algorithm to evaluate the path.
Applied neural network and Q-learning algorithm to solve the 3-D
[77]
collision avoidance problem.
Neural network Applied CGAN network to build the depth map used for further
algorithm [78]
collision avoidance.
[79] The depth map was predicted by using CNN network.
An algorithm combining the depth neural network and the epipolar
[80]
geometry was proposed.
[81] Applied CNN to estimate the optical flow of the pictures.
Introduced the ARE method on navigation and obstacle avoidance
[82]
problems.
[83] Proposed Advantaged Actor - Critic algorithm innovatively.
Reinforcement A multi-UAV anti-collision framework based on reinforcement learn-
[84]
learning ing was proposed.
[85] Proposed the two-stage reinforcement learning method.
Proposed a distributed DRL framework on collision avoidance of
[86]
dynamic obstacles.
Applied Bellman Ford algorithm to solve the problem of overflying
[87]
multiple interest points in a given area.
The adaptive differential multi-objective optimization algorithm was
[88]
used to find the optimal solution to avoid obstacles.
Genetic [89] Solved the problem of premature stagnation in genetic algorithm.
algorithm Applied genetic algorithm and evolutionary robot to evolve neural
[90]
network controller.
Combined genetic algorithm and TSP algorithm to minimize the path
[91] length and adopted a new cellular decomposition to avoid concave
obstacles.
[92] Applied Bayesian method to improve the genetic algorithm.

1) Neural Network Algorithm: Artificial neural network (ANN) can process different types of input
data, such as position information from GPS, image information, and then output the next movement
decision of UAV. [75] proposed a path planning algorithm based on ANN, which directly processed the
image of the visual system on the UAV, and applied the back-propagation neural network to get the best
control command of the next movement (forward, right, left, and so on).
In 3-D obstacle avoidance for UAV, the UAV’s movements are constrained by a series of conditions.
How to smooth the flight path of UAV to satisfy the flight restrictions of UAV, and how to evaluate the
flight path and finally get the relatively optimal path are the main challenges in the 3-D obstacle avoidance.
[76] proposed a path planning algorithm, where a Generalized Fuzzy Competitive Neural Network (G-
 21

FCNN) was applied to design the path. The irregular obstacles are distributed in 3-D environment, which
brings great difficulty for the design of obstacle avoidance algorithms. Considering avoiding random
moving obstacles, DRL algorithm was applied in [77], where neural network and Q-learning were used to
learn from the experience. Combined with dual network and priority sampling, the accuracy of collision
avoidance will reach to 97.5% after 13000 times of training.
To solve obstacle avoidance problems indoors with dynamic obstacles, an obstacle avoidance algorithm
based on DRL was proposed to learn surrounding environment from monocular vision in [78], which
used Conditional Generative Adversarial Network (CGAN) to build the depth map and finally used the
depth map to make collision avoidance decisions. This method retained the key information about the
environment during long observation, which can be applied to predict the motion information of objects,
so as to better avoid dynamic obstacles such as human. In [79], the depth map of monocular RGB image
was predicted by using convolutional neural network (CNN), then the depth map is used to predict and
avoid obstacles. The size of image input by CNN network was shrunk to ensure real-time processing, and
the mean filtering and histogram equalization were used to ensure the reliability of depth information. An
algorithm combining the depth neural network and epipolar geometry was proposed to extract the depth
value information of the image through deep learning in [80]. Epipolar geometry was used to reduce
the dependence on the training model, which improved the performance of the system when there are
differences between the real data and the training data.
CNN was applied in [81] to estimate the optical flow of the pictures obtained by the camera. The
decision of turning left or right is made by comparing the sum of the optical flow of left half-plane
and right half-plane of the pictures. Besides, the expansion of the optical flow is calculated to recognize
wall-like frontal obstacles. The experiment with a DJI F550 drone was carried out near a bike path, which
verified the proposed obstacle avoidance algorithm.
2) Reinforcement Learning: Reinforcement learning is usually applied in the problem with continuous
decision, which is often used in obstacle avoidance and path planing of UAV.
To enable the UAV to avoid obstacles when it is trapped, [82] introduced the adaptive and random
exploration (ARE) method in obstacle avoidance problem. This algorithm combines self-learning and
random search, so that the UAV can avoid obstacles and simultaneously conquer learning mistakes.
As to the obstacle avoidance algorithm of multiple UAVs, a multi-UAV anti-collision framework based
on reinforcement learning was proposed in [84], so as to avoid collision with other UAVs. Four elements
of environment control, action space, environment model, and return value in the decision-making process
of UAV were analyzed in details, so as to effectively solve multi-UAV anti-collision problem. To prevent
collision with dynamic obstacles in the environment, [83] proposed Experience-shared Advantaged Actor
- Critic (ES-A2C) algorithm innovatively, where the UAV can share the learned experience with other
UAVs in the swarm. Compared with traditional algorithm such as Q learning algorithm and Actor-Critic
algorithm, the ES-A2C algorithm has higher success rate and shorter training period.
In order to solve the collision avoidance problem between multiple UAVs, [85] proposed a multi-
UAV obstacle avoidance method based on two-stage reinforcement learning, which firstly adopted the
monitoring method with loss function, and then adopted the traditional reinforcement learning method in
the second stage to overcome the shortcomings of the traditional reinforcement learning methods, such
as high variance and poor repeatability.
In order to ensure safety of UAV in high-dynamic multi-obstacle environment, [86] proposed a dis-
 22

tributed DRL framework, which applied LSTM to help UAVs capture more information from the dynamic
environment. A clipped DRL loss function derived from UAV’s exploration in the environment was
proposed, which is more suitable for high-dynamic environment. Compared with DQN, Double DQN
(DDQN), and other advanced DRL algorithms, this algorithm has better convergence property.
3) Genetic Algorithm: Genetic algorithm is one of the intelligent optimization algorithms, which has
been widely applied in UAV path planning. The genetic algorithm encodes the solutions of the optimization
model into the chromosomes, evaluates the fitness of the chromosome, and retains the chromosome with
high fitness to the next generation. This process is repeated until the optimized solution is found.
The collision avoidance problem of UAV can be converted into a path planing problem, which is further
equivalent to a Traveling Salesman Problem (TSP) [93]. When solving TSP problem, genetic algorithm
was applied for path planning and avoiding obstacles. [87] first solved the problem of overflying multiple
points of interest in a given area by Bellman Ford algorithm, and then used genetic algorithm to find the
optimal path. [88] defined the path planing problem of UAV as a multi-objective optimization problem,
and proposed a new multi-gene structure to describe the path, in which the adaptive adjustment, crossover,
and mutation strategies were adopted, and the adaptive differential multi-objective optimization algorithm
was applied to obtain the optimal solution to avoid obstacles and meet the flight restrictions of UAV. [89]
used the minimum spanning tree and adaptive tournament selection to quantify and control the genetic
diversity, which solved the problem of premature stagnation. For the obstacle avoidance problem in multi-
UAV scenario, [90] used genetic algorithm and evolutionary robot to evolve neural network controller,
solving the obstacle avoidance problem of multi-UAV.
In agricultural scenarios, UAV path planning algorithm usually needs to maximize the coverage area
to complete specific tasks such as pesticide spraying. Besides, the agricultural environment often consists
of concave obstacles, which brings difficulties for collision avoidance algorithms. [91] combined genetic
algorithm and TSP algorithm to minimize the path length and adopted a new cellular decomposition to
avoid concave obstacles. In order to avoid concave obstacles, [92] applied Kalman filter to predict the
state of obstacles, and then used Bayesian method to improve the genetic algorithm so that the UAVs will
be easier to bypass concave obstacles.
4) Simulated Annealing Algorithm: Simulated annealing algorithm (SAA) has the characteristic of
probabilistic jumping out of local optimum, which can accelerate the path planing. [94] adopted SAA to
improve the global search efficiency and introduced random mutation operation on the basis of genetic
algorithm to improve the population diversity. [95] defined the detection search target problem as a TSP
problem. SAA was used to obtain the approximate optimal solution in a short time, which can effectively
guide UAV escaping from threat areas. For the path planning problem of multiple UAVs, a new path
planning method for UAV was proposed in [96], which considered three constraints including the number
of UAVs, obstacle avoidance and coverage area. The flight area was decomposed into small areas, then
K-means algorithm was applied to gather the target points. Finally, SAA was adopted to get the optimal
path under the largest coverage area, solving the path planning problem under multiple constraints.
5) Ant Colony Algorithm: As shown in Fig. 7, ant colony algorithm (ACA) is an algorithm that
simulates the optimal path for ants to find food. When a better path is obtained, a positive feedback signal
will be sent to guide the ants to the optimal path. In addition, the diversity of ant colony algorithm can
avoid local optimization.
ACA was used to solve the trajectory planning problem of UAV in [97]. The trajectory model of
 23

TABLE VIII
S UMMARY OF CLASSIC METHODS (2)

Category Reference One-sentence summary

Adopted SAA to improve the global search efficiency of collision
[94]
avoidance.
Applied SAA to obtain the approximate optimal solution in a
Simulated annealing [95]
short period of time.
algorithm (SAA)
Applied simulated annealing algorithm to get the optimal path
[96] under the largest area coverage in path planning problem under
multiple constraints.
Applied ACA to deal with the trajectory planning problem of
[97]
UAV.
Introduced a guidance factor of target node pheromone and re-
Ant Colony [98] excitation learning mechanism to improve the ACA in collision
Algorithm (ACA) avoidance.
Applied ACA to search for the shortest path for UAV in mountain
[99]
environment.
[100] Proposed a new multiple ant colony algorithm.

Fig. 7. Ant colony algorithm

UAV is derived, and the simulation results showed that UAVs can accurately complete the path planning
mission in a 3-D environment within 16 times of iteration. [98] introduced a guidance factor of target
node pheromone in the process of the state transition of traditional ACA. Besides, re-excitation learning
mechanism was applied to update pheromone on the path. Those two improvements greatly increased the
convergence speed.
[99] proposed an ant colony based obstacle avoidance method for UAV in mountain environment.
Firstly, Voronoi polygon was used to get the initial feasible solution of obstacle avoidance path in mountain
environment, and then ACA was used to search for the shortest path. Finally, unnecessary obstacles were
eliminated, which improved the searching speed and shortened the flight path.
For multiple UAVs, the multiple ant colony algorithm was proposed in [100]. The concept of antibody
similarity in immune optimization algorithm is introduced to measure the similarity of feasible solutions.
For multiple similar individuals, only the representative individuals are selected for calculation, so as to
 24

improve the searching speed.

The pros and cons for collision avoidance methods are shown in Table IX.

VI. F UTURE T RENDS

In order to enhance the anti-collision capability of UAVs, most of current researches tend to improve the
performance of collision avoidance algorithm. On the contrary, the performance of sensing has an impact
on the performance of anti-collision capability of UAVs. Fast obstacle sensing techniques can shorten
the sensing time and provide more response time during collision prediction and collision avoidance
procedures. Fast wireless networking techniques enable UAV swarm to spread the sensing information
quickly, and improve the cooperative sensing performance of UAV network. In this section, future trends
of anti-collision techniques for UAVs are summarized from two aspects, i.e, fast obstacle sensing and fast
wireless networking.

A. Fast Obstacle Sensing

1) Joint Sensing and Communication Based Obstacle Sensing: The obstacle sensing methods consists
of cooperative and non-cooperative obstacle sensing methods. In the cooperative obstacle sensing methods,
the communication between UAV and obstacle is applied to recognize and localize the obstacle. In the non-
cooperative obstacle sensing methods, the active or passive sensors are applied to recognize and localize
the obstacles. However, with joint sensing and communication (JSC) technology, the cooperative and non-
cooperative obstacle sensing methods can work simultaneously for obstacle sensing. The communication
signal is used for cooperative obstacle sensing, meanwhile, the communication signal can be applied
for radar sensing, positioning, imaging, and so on, which belong to non-cooperative sensing. Hence, the
JSC technology can realize fast and comprehensive sensing for obstacles, which is crucial for obstacle
avoidance. The JSC technology actually attracts wide attention in the era of 6th generation mobile networks
(6G) [101].
In the JSC system, the performance of sensing and communication can be benefited from the joint
design. The fuse of communication and sensing results improves the sensing accuracy [102]. And the prior
information of sensing can be applied to improve the performance of wireless networking [103]. Moreover,
with JSC technology, the spectrum and hardware can be reused by sensing and communication, which
improves the resource efficiency [104]. The application of JSC in UAV networks realizes the simultaneous
cooperative and non-cooperative obstacle sensing, which has the potential to realize fast sensing and
networking for UAV swarm and finally improve the obstacle avoidance capability. The integration of
positioning, communication, and radar is a promising technology for anti-collision system of UAVs,
because it can realize comprehensive sensing and communication simultaneously. With the combination
of FDA and OFDM technology, a FDA-OFDM scheme with broad application prospect was proposed in
[105], which realizes radar, communication, and positioning. In recent years, the communications over
mmWave and THz spectrum bands have attracted wide attention. With the wide bandwidth on mmWave
and THz band, the radar imaging is promising to be deployed. Since the radar image can provide detailed
information of the obstacles, it can be applied to recognize the obstacles. In [106], the airborne MIMO radar
and space-time coded (STC) waveform was designed, which is used to realize the joint communication and
synthetic aperture radar (SAR) system. [107] proposed a joint waveform for simultaneous communicate
and synthetic aperture radar imaging.
 25

TABLE IX
P ROS AND CONS OF COLLISION AVOIDANCE METHODS

Name of the methods Pros Cons Common features

• Low accuracy when the
• Low computation over-
Geometric number of obstacles is
head.
Methods large.
• Low cost. • Large turning radius.
Classic • High requirements for • No iteration.
Collision sensing since the global • Fast
Avoidance Graph obstacle distribution is calculation.
Algorithms Theory • Global optimization. required. • Low accuracy.
Methods
• Large construction time
of the graph.
• Low computation over- • Easy to be trapped
Artificial head. into local minimum
Potential when the number of
Field • Low cost.
obstacles is large.
• Able to process pure
Neural image, which decrease
• Prone to over fitting.
the calculation overhead
Network
Algorithm of sensing.
• Good performance to • Easy to fall into local
solve nonlinear problems. optimization.
Reinforce- • High variance and low
ment • Global optimization.
reproducibility.
Learning • Good
• Premature convergence generality.
Heuristic • Global optimization
and stagnation problem. • Low
Collision Genetic • Good performance on • High requirements for calculation speed.
Avoidance Algorithm collision avoidance of sequences coding. • Easy to fall
Algorithms
UAV swarm • Low efficiency. into local
optimization.
Simulated • Greatly affected by the
Annealing • Probabilistic jumping initial value.
Algorithm out of local optimization. • Slow convergence.

• Greatly affected by the

• Reduce the probability
Ant initial value.
of local optimization
Colony • A contradiction be-
because of population
Algorithm tween population diver-
diversity.
sity and convergence rate.

2) AI Chip Based Obstacle Sensing: The machine learning algorithms are widely applied in obstacle
sensing. With the chip-level obstacle sensing, the speed and accuracy of machine learning based obstacle
 26

sensing are further improved. The machine learning algorithms can be realized by central processing unit
(CPU), graphics processing unit (GPU), field programmable gate array (FPGA) and application specific
integrated circuit (ASIC). Generally, the FPGA has better performance/Watt compared with CPU and GPU
[108], [109], and CPU is more versatile. However, the high overhead suffered from the operating system
(OS) will degrade the speed of computation [110]. The GPU has high speed of computation. However,
the power consumption is high. The FPGA and ASIC have low power consumption compared with CPU
and GPU, which are suitable to be implemented in UAVs. However, the versatility of FPGA and ASIC
is low. Besides, the ASIC is customized, providing higher energy efficiency and speed of computation
compared with FPGA, which has potential to be applied in small-size UAVs.
In [111], the UAV is installed with the AI chip, whose structure is like the nervous system of fruit fly.
And [111] verified that the UAV can avoid obstacles with relatively low energy consumption. In [112],
with NVIDIA Jetson TX2 implemented in UAV, the data of six 4K cameras is processed for obstacle
sensing. A chip named “Navion” is developed in [113] for UAV, with tiny energy consumption, the chip
can process 171 frames per second for positioning. Overall, for the small UAVs, the ASIC or FPGA is
an optimal choice. However, for the large UAVs, the GPU can be applied. In academia and industry,
the implementation of AI chip on UAVs for obstacle sensing has attracted wide attention to enhance the
capability of obstacle sensing for UAVs.

B. Fast Wireless Networking

The fast wireless networking is of vital importance to enhance the anti-collision performance of UAVs.
The sensing information of obstacles with fast wireless networking can be spread to other UAVs in the
network quickly, such that UAVs can make anti-collision decision in advance and improve the anti-collision
performance. The research of wireless networking of UAVs mainly focuses on neighbor discovery, multiple
access control and the routing scheme.
1) Neighbor Discovery: The prior information of the distribution of neighbors can accelerate the speed
of neighbor discovery. Compared the neighbor discovery algorithms without the prior information of radar
sensing, the time of neighbor discovery algorithm is significantly reduced using radar sensing information
[114]. [115] proposed a three handshake neighbor discovery algorithm for radar-communication integrated
system (RCIS) network. The algorithm applies radar sensing to obtain the location information of neigh-
bors, so as to increase the speed of neighbor discovery. In [116], the efficiency of neighbor discovery
was improved using dual phased array radar. [117] proposed a neighbor discovery algorithm based on
joint radar and communication (JRC) technique. The algorithm takes advantage of the neighbors’ location
information obtained from radar sensing, and avoids repeated attempts on directions without neighbors.
The analysis and simulation shows that the algorithm can greatly improve the neighbor discovery speed.
2) Multiple Access Control: In order to design the fast multiple access control (MAC) scheme, UAVs
can estimate the number of competing nodes in the network with sensing information, thereby reducing
the congestion probability of the MAC protocol. As the number of nodes increases, the access competition
of nodes in a single channel will become more and more intense. By applying multi-channel mechanism,
the load and the delay of the network will be reduced. [118] proposed a MAC protocol applying integrated
sensing and communication (ISAC) technique, where radar and communication can cooperate with each
other. The simulation result shows that MAC protocol with ISAC can achieve high throughput. Feng
 27

et al. designed a MAC protocol under multi-channel opportunistic reservation based on cognitive radio
(CogMOR-MAC) [119]. The CogMOR-MAC applies multi-channel opportunistic reservation mechanism
to reserve resources, which reduces the access conflict and waiting time of UAV nodes. [120] proposed
a multi-channel MAC protocol with directional antennas (MMAC-DA) for UAV nodes to transmit and
receive directionally. MMAC-DA applies common control channel to negotiate control information, which
increases the throughput of UAV network.
3) Routing: The assistance of sensing information is promising to improve the efficiency of the routing
scheme. [121] proposed a fountain-code based greedy queue and location information assisted routing
protocol, which applies velocity and distance information to improve routing efficiency. Aiming at solving
the performance degradation caused by flooding in routing protocol, [122] proposed a location information
and cluster-based routing protocol, where each node selects the node with the largest forwarding distance
as the next hop to reduce the number of hops. In order to reduce the routing overhead, [123] proposed a
location-assisted zone routing protocol, which has better performance on routing overhead and end-to-end
delay compared with traditional routing schemes.

VII. C ONCLUSION
Due to the complex and changeable flight environment, flexible and reliable anti-collision technologies
are urgently needed to ensure UAV’s safety. To meet such requirement, we provide an overview for UAV
anti-collision technologies in this article. Firstly, we introduced laws and regulations of governments on
UAV flight safety. The laws and regulations establish a bridge between academia and application. Secondly,
anti-collision algorithms were introduced according to three aspects, namely, obstacle sensing, collision
prediction, and collision avoidance. We select the representative articles of anti-collision algorithms and
classified them in details, so as to provide a clear structure for readers. Finally, the future trends of
UAV’s anti-collision techniques are revealed from two aspects, i.e, fast obstacle sensing and fast wireless
networking. Combined with state-of-art technologies of sensing, computation and communication, we
provide a new perspective for UAV anti-collision technologies. This article may provide guidelines for
the design of anti-collision mechanisms for UAVs.

R EFERENCES
[1] FAA. “UAS by the Numbers.” FAA.gov. https://www.faa.gov/uas/resources/by the numbers/ (accessed Nov. 20, 2020).
[2] CAAC. “China Civil Aviation Working Conference held in 2020.” CAAC.gov.cn. http://www.caac.gov.cn/ZTZL/RDZT/2020QGMHGZHY/2020TPXW
(accessed Nov. 20, 2020).
[3] J. N. Yasin, S. A. S. Mohamed, M. Haghbayan, et al. “Night vision obstacle detection and avoidance based on Bio-Inspired Vision
Sensors,” 2020 IEEE Sensors, 2020, pp.1–4.
[4] J. N. Yasin, H. Mahboob, M. H. Haghbayan, et al. “Energy-Efficient Navigation of an Autonomous Swarm with Adaptive
Consciousness,” Remote Sensing, vol. 13, no. 6, 2021, pp. 1059.
[5] A. Horst, E. Kaba, R. Dornberger, et al. “Improving Collision Avoidance Behavior of a Target-Searching Algorithm for Kilobots,”
Journal of Robotics and Control (JRC), vol. 2, no. 6, 2021, pp. 588–593.
[6] N. Aswini, E. Krishna Kumar, S. V. Uma. “UAV and obstacle sensing techniques–a perspective,” International Journal of Intelligent
Unmanned Systems, Vol. 6, No. 1, pp. 32–46, 2018.
[7] A. Ryan, M. Zennaro, A. Howell, R. Sengupta and J. K. Hedrick, “An overview of emerging results in cooperative UAV control,”
2004 43rd IEEE Conference on Decision and Control (CDC), Nassau, Bahamas, 2004, pp. 602–607.
[8] Y. I. Jenie et al., “Taxonomy of Conflict Detection and Resolution Approaches for Unmanned Aerial Vehicle in an Integrated Airspace,”
IEEE Transactions on Intelligent Transportation Systems, vol. 18, no. 3, pp. 558–567, March 2017.
 28

[9] B. N. Chand, P. Mahalakshmi and V. P. S. Naidu, “Sense and avoid technology in unmanned aerial vehicles: A review,” 2017
International Conference on Electrical, Electronics, Communication, Computer, and Optimization Techniques (ICEECCOT), Mysuru,
2017, pp. 512–517.
[10] H. Shakhatreh et al., “Unmanned Aerial Vehicles (UAVs): A Survey on Civil Applications and Key Research Challenges,” IEEE
Access, vol. 7, pp. 48572–48634, 2019.
[11] J. N. Yasin, S. A. S. Mohamed, M. Haghbayan, J. Heikkonen, H. Tenhunen and J. Plosila, “Unmanned Aerial Vehicles (UAVs):
Collision Avoidance Systems and Approaches,” IEEE Access, vol. 8, pp. 105139–105155, 2020.
[12] The State Council of the P.R.C. “Interim Regulations on Flight Management of Unmanned Aircraft (Draft for Comment).” gov.cn.
http://www.gov.cn/xinwen/2018-04/10/content 5281308.htm (accessed Nov. 20, 2020).
[13] EASA. “Drones - regulatory framework background.” EASA.europa.eu. https://www.easa.europa.eu/domains/civil-drones-rpas/certified-category-civil
(accessed Aug. 25, 2021).
[14] CAA. “Airspace restrictions for unmanned aircraft and drones.” EASA.europa.eu. https://www.caa.co.uk/Consumers/Unmanned-aircraft/Our-role/Airs
(accessed Aug. 25, 2021).
[15] CAAC. “Administrative Measures for Air Traffic Management of Civilian UAV Systems.” CAA.co.uk.
http://www.caac.gov.cn/XXGK/XXGK/GFXWJ/201610/t20161008 40016.html (accessed Nov. 21, 2020).
[16] CAAC. “Regulation on Dynamic Data Management of Light and Small Civil UAV flight.” CAAC.gov.cn.
http://www.caac.gov.cn/XXGK/XXGK/TZTG/201911/P020191120352957148255.pdf (accessed Nov. 21, 2020).
[17] FAA. “Suggested rules for UAV remote identification.” FAA.gov. https://www.faa.gov/uas/getting started/remote id/ (accessed Nov.
21, 2020).
[18] CAA. “CAP 722: Unmanned Aircraft System Operations in UK Airspace - Guidance .” CAA.co.uk.
https://publicapps.caa.co.uk/docs/33/CAP722%20Edition8(p).pdf (accessed Aug. 25, 2021).
[19] CAAC. “Airworthiness Standards for Medium and High-risk Unmanned Helicopter Systems (Trial).” CAAC.gov.cn.
www.caac.gov.cn/HDJL/YJZJ/202002/P020200225622960426433.pdf (accessed Nov. 21, 2020).
[20] CAAC. “Airworthiness Standards for High-risk Fixed - wing Cargo UAV Systems (Trial).” CAAC.gov.cn.
http://www.caac.gov.cn/XXGK/XXGK/ZFGW/202005/t20200511 202468.html (accessed Nov. 21, 2020).
[21] CAAC. “Large Cargo UAV airworthiness Standards.” CAAC.gov.cn. http://www.caac.gov.cn/HDJL/YJZJ/201908/t20190807 197964.html
(accessed Nov. 21, 2020).
[22] CAAC. “Guidance on UAV Airworthiness Certification Based on Operational Risks.” CAAC.gov.cn.
http://www.caac.gov.cn/XXGK/XXGK/ZFGW/201901/P020190125609946283900.pdf (accessed Nov. 21, 2020).
[23] CAAC. “Risk Assessment Guide for Civil UAV System Airworthiness Certification Project (Draft for Comments).” CAAC.gov.cn.
http://www.caac.gov.cn/HDJL/YJZJ/202003/P020200330336461465630.pdf (accessed Nov. 21, 2020).
[24] CAAC. “Regulations on the Administration of civilian UAV pilots.” CAAC.gov.cn http://www.caac.gov.cn/XXGK/XXGK/JYTNDF/201906/t2019061
(accessed Nov. 21, 2020).
[25] FAA. “Airman Certificate Testing Service (ACTS) Award Notice.” FAA.gov. https://www.faa.gov/training testing/testing/acts/media/acts notification
(accessed Nov. 21, 2020).
[26] EASA. “Civil drones (unmanned aircraft).” EASA.europa.eu. https://www.easa.europa.eu/domains/civil-drones-rpas (accessed Aug.
25, 2021).
[27] FAA. “Appendix A-National Airspace System Overview.” FAA.gov. https://www.faa.gov/air traffic/nas/nynjphl redesign/documentation/feis/media/A
(accessed Nov. 21, 2020).
[28] T. Williamson and N. A. Spencer, “Development and operation of the Traffic Alert and Collision Avoidance System (TCAS),”
Proceedings of the IEEE, vol. 77, no. 11, pp. 1735–1744, Nov. 1989.
[29] E. Valovage, “Enhanced ADS-B Research,” 2006 IEEE 25TH Digital Avionics Systems Conference, Portland, OR, 2006, pp. 1–7.
[30] M. Marques, A. Brum, S. Antunes and J. G. Mota, “Sense and Avoid Implementation in a Small Unmanned Aerial Vehicle,” 2018
13th APCA International Conference on Automatic Control and Soft Computing (CONTROLO), Ponta Delgada, 2018, pp. 395–400.
[31] Q. Wu, J. Wei and X. Li, “Research Progress of Obstacle Detection Based on Monocular Vision,” 2014 Tenth International Conference
on Intelligent Information Hiding and Multimedia Signal Processing, Kitakyushu, 2014, pp. 195–198.
[32] Z. Ma, T. Hu, L. Shen, W. Kong and B. Zhao, “A detection and relative direction estimation method for UAV in sense-and-avoid,”
2015 IEEE International Conference on Information and Automation, Lijiang, 2015, pp. 2677–2682.
[33] X. Peng, H. Lin and J. Dai, “Path planning and obstacle avoidance for vision guided quadrotor UAV navigation,” 2016 12th IEEE
International Conference on Control and Automation (ICCA), Kathmandu, 2016, pp. 984–989.
[34] Z. Chen, X. Luo and B. Dai, “Design of Obstacle Avoidance System for Micro-UAV Based on Binocular Vision,” 2017 International
Conference on Industrial Informatics - Computing Technology, Intelligent Technology, Industrial Information Integration (ICIICII),
Wuhan, 2017, pp. 67–70.
[35] C. Shuai, H. Wang, W. Zhang, P. Yao and Y. Qin, “Binocular vision perception and obstacle avoidance of visual simulation system
for power lines inspection with UAV,” 2017 36th Chinese Control Conference (CCC), Dalian, 2017, pp. 10480–10485.
 29

[36] Skydio. “How does Skydio 2 work.” Skydio.com https://support.skydio.com/hc/en-us/articles/360036116834-How-does-Skydio-2-work-

(accessed Jan. 5, 2021).
[37] Walkera. “Specs of R1000 RTK.” Walkera.com. https://www.walkera.com/index.php/Goods/info/id/85.html/ (accessed Aug. 25, 2021).
[38] K. Niwa, K. Watanabe and I. Nagai, “A detection method using ultrasonic sensors for avoiding a wall collision of Quadrotors,” 2017
IEEE International Conference on Mechatronics and Automation (ICMA), Takamatsu, 2017, pp. 1438–1443.
[39] S. Kemkemian, M. Nouvel-Fiani, P. Cornic and P. Garrec, “A MIMO radar for Sense and Avoid function: A fully static solution for
UAV,” 11-th INTERNATIONAL RADAR SYMPOSIUM, Vilnius, 2010, pp. 1–4.
[40] Y. K. Kwag, M. S. Choi, C. H. Jung and K. Y. Hwang, “Collision Avoidance Radar for UAV,” 2006 CIE International Conference
on Radar, Shanghai, 2006, pp. 1–4.
[41] DJI. “AGRICULTURE SOLUTION PACKAGE 2.0.” DJI.com. https://www.dji.com/uk/agriculture-solution-v2/ (accessed Jan. 12,
2021).
[42] XAG. “Parameters of XAG V40 2021 agricultural UAV.” XA.com.https://www.xa.com/v40-2021/specs/ (accessed Aug. 25, 2021).
[43] N. Gageik, P. Benz and S. Montenegro, “Obstacle Detection and Collision Avoidance for a UAV With Complementary Low-Cost
Sensors,” IEEE Access, vol. 3, pp. 599–609, 2015.
[44] V. M. Polyakov, I. N. Kaliteevsky, K. S. Amelin, V. A. Smyslov and M. A. Permyakov, “Complexed NIR laser detector and LWIR
camera optical system with neural network management for UAV collision avoidance system,”2018 International Conference Laser
Optics (ICLO). IEEE, 2018, pp. 280–280.
[45] W. Zheng, J. Xiao and T. Xin, “Integrated navigation system with monocular vision and LIDAR for indoor UAVs,” 2017 12th IEEE
Conference on Industrial Electronics and Applications (ICIEA), Siem Reap, 2017, pp. 924–929.
[46] XAG. “Description of the performance and usage of XCope obstacle avoidance system.” Xa.com
https://www.xa.com/news/official/xai/327/ (accessed Jan. 12, 2021).
[47] DJI. “Features of the Guidance System.” DJI.com. https://www.dji.com/uk/guidance?site=brandsite (accessed Jan. 12, 2021).
[48] B. K. P. Horn, B. G. Schunck . “Determining optical flow,” Artificial intelligence, vol. 17, pp. 185–203, Aug. 1981.
[49] R. Giubilato, M. Pertile and S. Debei, “A comparison of monocular and stereo visual FastSLAM implementations,” 2016 IEEE
Metrology for Aerospace (MetroAeroSpace), Florence, 2016, pp. 227–232.
[50] I. Alzugaray, L. Teixeira and M. Chli, “Short-term UAV path-planning with monocular-inertial SLAM in the loop,” 2017 IEEE
International Conference on Robotics and Automation (ICRA), Singapore, 2017, pp. 2739–2746.
[51] X. Gan, H. Qu, P. Liu and Q. Wang, “Detection of Threat Situation for UAV Collision Avoidance Based on Sliding Window Polynomial
Fitting Prediction Method,” 2019 International Conference on Intelligent Computing, Automation and Systems (ICICAS), Chongqing,
China, 2019, pp. 230–233.
[52] Y. Li. “Real-time position prediction of aerial target based on polynomial fitting,” Computer and Digital Engineeringvol. 43, no. 3,
pp. 404–407, 2015.
[53] D. L. Vinokursky, O. S. Mezentceva and P. V. Samoylov, “Trajectory Planning of UAV Group: Pythagorean Hodograph and Bernstein-
Bezier Composite Curves in the Plane,” 2020 International Russian Automation Conference (RusAutoCon), Sochi, Russia, 2020, pp.
704–707.
[54] J. A. S. Jayasinghe and A. M. B. G. D. A. Athauda, “Smooth trajectory generation algorithm for an unmanned aerial vehicle (UAV)
under dynamic constraints: Using a quadratic Bezier curve for collision avoidance,” 2016 Manufacturing and Industrial Engineering
Symposium (MIES), Colombo, Sri Lanka, 2016, pp. 1–6.
[55] V. Kurtz and H. Lin, “Toward Verifiable Real-Time Obstacle Motion Prediction for Dynamic Collision Avoidance,” 2019 American
Control Conference (ACC), Philadelphia, PA, USA, 2019, pp. 2633–2638.
[56] N. Imanberdiyev, C. Fu, E. Kayacan and I. Chen, “Autonomous navigation of UAV by using real-time model-based reinforcement
learning,” 2016 14th International Conference on Control, Automation, Robotics and Vision (ICARCV), Phuket, Thailand, 2016, pp.
1–6.
[57] Y. Chen, N. González-Prelcic and R. W. Heath, “Collision-Free UAV Navigation with a Monocular Camera Using Deep Reinforcement
Learning,” 2020 IEEE 30th International Workshop on Machine Learning for Signal Processing (MLSP), Espoo, Finland, 2020, pp.
1–6.
[58] D. Pedro, J. P. Matos-Carvalho, J. M. Fonseca, et al. “Collision Avoidance on Unmanned Aerial Vehicles Using Neural Network
Pipelines and Flow Clustering Techniques,” Remote Sensing, vol. 13, no. 13, 2021, pp. 2643.
[59] L. N. N. T. Ha, D. H. P. Bui, and S. K. Hong. “Nonlinear control for autonomous trajectory tracking while considering collision
avoidance of UAVs based on geometric relations,” Energies, vol. 12, no. 8, pp: 1551, April 2019.
[60] L. R. Salazar et al., “A novel system for non-cooperative UAV sense-and-avoid,” Proceedings of european navigation conference,
April 2013.
[61] C. Y. Tan, S. Huang, K. K. Tan, et al. “Three dimensional collision avoidance for multi unmanned aerial vehicles using velocity
obstacle,” Journal of Intelligent & Robotic Systems, vol. 97, no. 1, 2020, pp. 227–248.
 30

[62] C. Yang, L. Liu and J. Wu, “Path planning algorithm for small UAV based on dubins path,” 2016 IEEE International Conference on
Aircraft Utility Systems (AUS), Beijing, 2016, pp. 1144–1148.
[63] R. Kikutis, J. Stankūnas, D. Rudinskas, T. Masiulionis, “Adaptation of Dubins paths for UAV ground obstacle avoidance when using
a low cost on-board GNSS sensor,” Sensors, vol. 17, no. 10, pp. 2223, Sep. 2017.
[64] Y. Lin and S. Saripalli, “Path planning using 3D Dubins Curve for Unmanned Aerial Vehicles,” 2014 International Conference on
Unmanned Aircraft Systems (ICUAS), Orlando, FL, USA, 2014, pp. 296–304.
[65] C. Zhang, H. Liu and Y. Tang, “Quantitative Evaluation of Voronoi Graph Search Algorithm in UAV Path Planning,” 2018 IEEE 9th
International Conference on Software Engineering and Service Science (ICSESS), Beijing, China, 2018, pp. 563–567.
[66] X. Chen and X. Chen, “The UAV dynamic path planning algorithm research based on Voronoi diagram,” The 26th Chinese Control
and Decision Conference (2014 CCDC), Changsha, 2014, pp. 1069–1071.
[67] Z. Xu, R. Wei, K. Zhou, Q. Zhang and R. He, “Laguerre Graph self-optimize path planning algorithm for UAVs in irregular obstacle
environment,” 2016 IEEE Chinese Guidance, Navigation and Control Conference (CGNCC), Nanjing, 2016, pp. 1076–1079.
[68] J. Mok, Y. Lee, S. Ko, I. Choi and H. S. Choi, “Gaussian-mixture based potential field approach for UAV collision avoidance,” 2017
56th Annual Conference of the Society of Instrument and Control Engineers of Japan (SICE), Kanazawa, 2017, pp. 1316–1319.
[69] L. Liu, R. Shi, S. Li and J. Wu, “Path planning for UAVS based on improved artificial potential field method through changing
the repulsive potential function,” 2016 IEEE Chinese Guidance, Navigation and Control Conference (CGNCC), Nanjing, 2016, pp.
2011–2015.
[70] Y. Zhang, “Flight path planning of agriculture UAV based on improved artificial potential field method,” 2018 Chinese Control And
Decision Conference (CCDC), Shenyang, 2018, pp. 1526–1530.
[71] J. W. Hu, M. Wang, C. H. Zhao, et al. “Formation control and collision avoidance for multi-UAV systems based on Voronoi partition,”
Science China Technological Sciences, vol. 63, no. 1, 2020, pp.65–72.
[72] A. Saravanakumar, A. Kaviyarasu, R. A. Jasmine. “Sampling based path planning algorithm for UAV collision avoidance,” Sadhana,vol.
46, no. 3, 2021, pp. 1–8.
[73] G. Ahmed, T. Sheltami, M. Deriche, et al. “An energy efficient IoD static and dynamic collision avoidance approach based on gradient
optimization,” Ad Hoc Networks, vol. 118, 2021, pp. 102519.
[74] L.E. Dubins, “On Curves of Minimal Length with a Constraint on Average Curvature, and with Prescribed Initial and Terminal
Positions and Tangents,” American Journal of Mathematics 79, no. 3 (1957): 497–516
[75] M. M. Kurdi, A. K. Dadykin, I. Elzein and I. S. Ahmad, “Proposed system of artificial Neural Network for positioning and navigation
of UAV-UGV,” 2018 Electric Electronics, Computer Science, Biomedical Engineerings’ Meeting (EBBT), Istanbul, 2018, pp. 1–6.
[76] X. Liang, H. Wang, D. Li and C. Liu, “Three-dimensional path planning for unmanned aerial vehicles based on fluid flow,” 2014
IEEE Aerospace Conference, Big Sky, MT, 2014, pp. 1–13.
[77] X. Han, J. Wang, J. Xue and Q. Zhang, “Intelligent Decision-Making for 3-Dimensional Dynamic Obstacle Avoidance of UAV Based
on Deep Reinforcement Learning,” 2019 11th International Conference on Wireless Communications and Signal Processing (WCSP),
Xi’an, China, 2019, pp. 1–6.
[78] A. Singla, S. Padakandla and S. Bhatnagar, “Memory-Based Deep Reinforcement Learning for Obstacle Avoidance in UAV With
Limited Environment Knowledge,” IEEE Transactions on Intelligent Transportation Systems, vol. 22, no. 1, pp. 107–118, Jan. 2021.
[79] Z. Zhang, M. Xiong and H. Xiong, “Monocular Depth Estimation for UAV Obstacle Avoidance,” 2019 4th International Conference
on Cloud Computing and Internet of Things (CCIOT), Changchun, China, 2019, pp. 43–47.
[80] C. Wang, X. Liang, S. Zhang and C. Shen, “Motion Parallax Estimation for Ultra Low Altitude Obstacle Avoidance,” 2019 IEEE
International Conference on Unmanned Systems (ICUS), Beijing, China, 2019, pp. 463–468.
[81] S. Back, G. Cho, J. Oh, et al. “Autonomous UAV trail navigation with obstacle avoidance using deep neural networks,” Journal of
Intelligent & Robotic Systems, vol. 100, no. 3, 2020, pp. 1195–1211.
[82] Y. Zhao, Z. Zheng, X. Zhang and Y. Liu, “Q learning algorithm based UAV path learning and obstacle avoidence approach,” 2017
36th Chinese Control Conference (CCC), Dalian, 2017, pp. 3397–3402.
[83] X. Han, J. Wang, Q. Zhang, X. Qin and M. Sun, “Multi-UAV Automatic Dynamic Obstacle Avoidance with Experience-shared A2C,”
2019 International Conference on Wireless and Mobile Computing, Networking and Communications (WiMob), Barcelona, Spain,
2019, pp. 330–335.
[84] B. Fnag, X. Feng, S. Xu, “Research on Cooperative Collision Avoidance Problem of Multiple UAV Based on Reinforcement
Learning,”2017 10th International Conference on Intelligent Computation Technology and Automation (ICICTA), Changsha, 2017,
pp. 103–109.
[85] D. Wang, T. Fan, T. Han and J. Pan, “A Two-Stage Reinforcement Learning Approach for Multi-UAV Collision Avoidance Under
Imperfect Sensing,”IEEE Robotics and Automation Letters, vol. 5, no. 2, pp. 3098–3105, April 2020.
[86] G. U. O. Tong, N. Jiang, L. I. Biyue, et al. “UAV navigation in high dynamic environments: A deep reinforcement learning approach,”
Chinese Journal of Aeronautics, vol. 34, no. 2, 2021, pp. 479–489.
 31

[87] S. Mufti, V. Roberge and M. Tarbouchi, “A GPU Accelerated Path Planner for Multiple Unmanned Aerial Vehicles,” 2019 IEEE
Canadian Conference of Electrical and Computer Engineering (CCECE), Edmonton, AB, Canada, 2019, pp. 1–4.
[88] Y. Liu, R. Lv, X. Guan and J. Zeng, “Path planning for unmanned aerial vehicle under geo-fencing and minimum safe separation
constraints,” 2016 12th World Congress on Intelligent Control and Automation (WCICA), Guilin, 2016, pp. 28–31.
[89] A. Behjat, S. Chidambaran and S. Chowdhury, “Adaptive Genomic Evolution of Neural Network Topologies (AGENT) for State-to-
Action Mapping in Autonomous Agents,” 2019 International Conference on Robotics and Automation (ICRA), Montreal, QC, Canada,
2019, pp. 9638–9644.
[90] S. Oh and J. Suk, “Evolutionary design of the controller for the search of area with obstacles using multiple UAVs,” ICCAS 2010,
Gyeonggi-do, 2010, pp. 2541–2546.
[91] T. H. Pham, Y. Bestaoui and S. Mammar, “Aerial robot coverage path planning approach with concave obstacles in precision
agriculture,” 2017 Workshop on Research, Education and Development of Unmanned Aerial Systems (RED-UAS), Linkoping, 2017,
pp. 43–48.
[92] Z. Zheng and H. Yao, “A Method for UAV Tracking Target in Obstacle Environment,” 2019 Chinese Automation Congress (CAC),
Hangzhou, China, 2019, pp. 4639–4644.
[93] M. Cakir, “2D path planning of UAVs with genetic algorithm in a constrained environment,” 2015 6th International Conference on
Modeling, Simulation, and Applied Optimization (ICMSAO), Istanbul, 2015, pp. 1–5.
[94] Z. Cheng and D. Li, “Improved GASA Algorithm for Mutation Strategy UAV Path Planning,” 2018 10th International Conference
on Communication Software and Networks (ICCSN), Chengdu, 2018, pp. 506–510.
[95] B. Basbous, “2D UAV Path Planning with Radar Threatening Areas using Simulated Annealing Algorithm for Event Detection,” 2018
International Conference on Artificial Intelligence and Data Processing (IDAP), Malatya, Turkey, 2018, pp. 1–7.
[96] X. Yue and W. Zhang, “UAV Path Planning Based on K-Means Algorithm and Simulated Annealing Algorithm,” 2018 37th Chinese
Control Conference (CCC), Wuhan, 2018, pp. 2290–2295.
[97] Z. Li and R. Han, “Unmanned Aerial Vehicle Three-dimensional Trajectory Planning Based on Ant Colony Algorithm,” 2018 37th
Chinese Control Conference (CCC), Wuhan, 2018, pp. 9992–9995.
[98] W. Zhang and B. Zhang, “Improvement of UAV Track Trajectory Algorithm Based on Ant Colony Algorithm,” 2019 International
Conference on Intelligent Transportation, Big Data Smart City (ICITBS), hangsha, China, 2019, pp. 28–31.
[99] Z. Ma, H. Gong and X. Wang, “An UAV Path Planning Method in Complex Mountainous Area Based on a New Improved Ant Colony
Algorithm,”2019 International Conference on Artificial Intelligence and Advanced Manufacturing (AIAM), Dublin, Ireland, 2019, pp.
125–129.
[100] W. Lin, Y. Zhu, W. Zeng and S. Wang,“Track Planning Model for Multi-UAV Based on New Multiple Ant Colony Algorithm,”2018
Chinese Automation Congress (CAC), Xi’an, China, 2018, pp. 3862–3867.
[101] D. K. Pin Tan et al., “Integrated Sensing and Communication in 6G: Motivations, Use Cases, Requirements, Challenges and Future
Directions,” 2021 1st IEEE International Online Symposium on Joint Communications and Sensing (JCAS), Dresden, Germany, 2021,
pp. 1–6.
[102] C. Sturm and W. Wiesbeck, “Waveform Design and Signal Processing Aspects for Fusion of Wireless Communications and Radar
Sensing,” in Proceedings of the IEEE, vol. 99, no. 7, pp. 1236–1259, July 2011.
[103] Z. Feng et al., “Joint radar and communication: A survey,” China Communications, vol. 17, no. 1, pp. 1–27, Jan. 2020.
[104] T. Wild, V. Braun and H. Viswanathan, “Joint Design of Communication and Sensing for Beyond 5G and 6G Systems,” in IEEE
Access, vol. 9, pp. 30845–30857, 2021.
[105] H. Huang and W. Wang, “FDA-OFDM for integrated navigation, sensing, and communication systems,” in IEEE Aerospace and
Electronic Systems Magazine, vol. 33, no. 5-6, pp. 34–42, May-June 2018.
[106] J. Wang, X. Liang, L. Chen, L. Wang and S. Shi, “Joint Wireless Communication and High Resolution SAR Imaging Using Airborne
Mimo Radar System,” IGARSS 2019 - 2019 IEEE International Geoscience and Remote Sensing Symposium, Yokohama, Japan, 2019,
pp. 2511–2514.
[107] A. Herschfelt and D. W. Bliss, “Joint radar-communications waveform multiple access and synthetic aperture radar receiver,” 2017
51st Asilomar Conference on Signals, Systems, and Computers, Pacific Grove, CA, USA, 2017, pp. 69–74.
[108] E. Nurvitadhi et al., “Accelerating recurrent neural networks in analytics servers: Comparison of FPGA, CPU, GPU, and ASIC,” 2016
26th International Conference on Field Programmable Logic and Applications (FPL), Lausanne, Switzerland, 2016, pp. 1–4.
[109] E. Nurvitadhi, D. Sheffield, Jaewoong Sim, A. Mishra, G. Venkatesh and D. Marr, “Accelerating binarized neural networks: Comparison
of FPGA, CPU, GPU, and ASIC,” 2016 International Conference on Field-Programmable Technology (FPT). IEEE, 2016: 77–84.
[110] Arrow Electronics. “FPGA vs CPU vs GPU vs Microcontroller: How Do They Fit into the Processing Jigsaw Puzzle?” ARROW.com.
https://www.arrow.com/en/research-and-events/articles/fpga-vs-cpu-vs-gpu-vs-microcontroller (accessed Apr. 5, 2021).
[111] H. Hsueh. “NTHU Research Team Teaches Drone to Fly Like an Insect.” Businesswire.com.
https://www.businesswire.com/news/home/20200304005029/en/NTHU-Research-Team-Teaches-Drone-to-Fly-Like-an-Insect
(accessed Apr. 5, 2021).
 32

[112] N. Alarcon. “Inception Spotlight: New Skydio 2 Drone Powered by NVIDIA Jetson.” NVIDIA.com
https://developer.nvidia.com/blog/skydio-2-jetson-tx2-drone/ (accessed Apr. 5, 2021).
[113] J. Chu. “Chip upgrade helps miniature drones navigate.” MIT News. https://news.mit.edu/2018/novel-chip-upgrade-helps-miniature-drones-navigate-0
(accessed Apr. 5, 2021).
[114] Z. Wei, C. Han, C. Qiu, Z. Feng and H. Wu, “Radar assisted fast neighbor discovery for wireless Ad Hoc networks,” IEEE Access,
vol. 7, pp. 176514–176524, 2019.
[115] J. Li, L. Peng, Y. Ye, R. Xu, W. Zhao and C. Tian, “A neighbor discovery algorithm in network of radar and communication integrated
system,” 2014 IEEE 17th International Conference on Computational Science and Engineering. Chengdu, China, 2014, pp. 1142–1149.
[116] N. Liu, L. Peng, R. Xu, J. Zhang, W. Zhao and J. Zhu, “Neighbor discovery in wireless network with double-face phased array radar,”
2016 12th International Conference on Mobile Ad-Hoc and Sensor Networks (MSN), Hefei, China, 2016, pp. 434–439.
[117] D. Ji et al., “Radar-communication integrated neighbor discovery for wireless ad hoc networks,” 2019 11th International Conference
on Wireless Communications and Signal Processing (WCSP), Xi’an, China, 2019, pp. 1–5.
[118] R. M. Gutierrez, H. Yu, A. R. Chiriyath, G. Gubash, A. Herschfelt and D. W. Bliss, “Joint sensing and communications multiple-access
system design and experimental characterization,” 2019 IEEE Aerospace Conference, Big Sky, MT, USA, 2019, pp. 1–8.
[119] P. Feng et al., “CogMOR-MAC: A cognitive multi-channel opportunistic reservation MAC for multi-UAVs ad hoc networks,” Computer
Communications, vol.136, pp. 30–42, 2019.
[120] D. N. M. Dang, H. T. Le, H. S. Kang, C. S. Hong and J. Choe, “Multi-channel MAC protocol with Directional Antennas in wireless
ad hoc networks,” 2015 International Conference on Information Networking (ICOIN), Cambodia, 2015, pp. 81–86.
[121] C. Yin, Z. Xiao, X. Cao, X. Xi, P. Yang and D. Wu, “Enhanced routing protocol for fast flying UAV network,” 2016 IEEE International
Conference on Communication Systems (ICCS), Shenzhen, China, 2016, pp. 1–6.
[122] S. Liu, K. Liu, J. Zhang, T. Zhang, Z. Xu and F. Liu, “A Location and Cluster-Based MAC and Routing Protocol for Wireless
Mesh Networks,” 2018 2nd IEEE Advanced Information Management, Communicates, Electronic and Automation Control Conference
(IMCEC), Xi’an, China, 2018, pp. 1963–1969.
[123] T. P. Thi Minh, T. T. Nguyen and D. Kim, “Location Aided Zone Routing Protocol in Mobile Ad Hoc Networks,” 2015 IEEE 20th
Conference on Emerging Technologies & Factory Automation (ETFA), Luxembourg, Luxembourg, 2015, pp. 1–4.