You are on page 1of 9

This article has been accepted for publication in a future issue of this journal, but has not been

fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TII.2019.2913683, IEEE
Transactions on Industrial Informatics
1

An Algorithm of Reactive Collision Free 3D


Deployment of Networked Unmanned Aerial
Vehicles for Surveillance and Monitoring
Hailong Huang and Andrey V. Savkin

Abstract—This paper focuses on the application of surveillance UAV

and monitoring using Unmanned Aerial Vehicles (UAVs). A UAV-GN link UAV-UAV link
novel coverage model is proposed to characterize the quality 𝛼 UAV
of coverage (QoC) of a target by a UAV. Based on this model,
a reactive collision free 3D deployment algorithm is proposed, 𝑈2
𝑈1 𝛼
with the goal of maximizing the overall QoC of targets by a
network of UAVs. The algorithm consists of two navigation laws GN
for the horizontal movement and the vertical movement, both
𝒟
of which are easily implementable in real time. The convergence
of the algorithm is proved, and the computational complexity is
analysed. Computer simulations are conducted to demonstrate
the performance of the proposed method. Fig. 1: Using UAVs for target surveillance.
Index Terms—Unmanned aerial vehicles, network of drones,
reactive deployment, surveillance and monitoring, quality of
coverage.
both U1 and U2 are within the vision cone of the UAV, they are
monitored with different QoC. However, to the best knowledge
I. I NTRODUCTION of the authors, the existing publications simply regard them as
being covered, while have not studied the QoC.
U NMANNED aerial vehicles (UAVs) or drones, which
were initially used in military applications, have now
been applied widely for civilian probems [1], such as infras-
In this paper, a novel coverage model characterizing the
QoC of a targets by a UAV is proposed. Beyond the basic
tructure inspection [2], wireless communication support [3], vision cone model, this model distinguishes targets at different
surveillance and monitoring [4]. In surveillance applications, positions inside the cone. Based on this model, a reactive al-
UAVs are equipped with some specific sensors, such as gorithm is developed to deploy UAVs in continuous 3D space.
cameras. They fly into the sky and monitor ground targets For a given fleet of UAVs, the objective is to maximize the
of interest, such as humans, vehicles, landmarks, animals for overall QoC of a given set of targets. It is assumed that each
the purpose of surveillance, security, etc [5]. UAV can estimate the locations of targets which are seen by
There is a number of publications focusing on the surveil- the on-board camera, using some available image processing
lance application using UAVs [6], [7], [8]. Though they techniques. Also, each UAV can measure the positions of
consider different problems, one similarity lies in the coverage other nearby UAVs. With these two pieces of information as
model. As shown in Fig. 1, a UAV is equipped with a ground input, each UAV runs the proposed algorithm independently
facing camera with the visibility angle α. The camera can see a to decide its movement. The movement of each UAV is
circular area on the ground and the radius of this area depends separated into the horizontal sub-movement and the vertical
on the altitude and α. If a target is within this circle, it is sub-movement. The proposed algorithm consists of two navi-
considered to be covered by the UAV. For several targets inside gation laws for these two sub-movements, respectively. Several
the circle, though they are all covered, they may correspond constraints have to be satisfied when UAVs move. There is a
to various quality of coverage (QoC). For example, the target set of ground nodes (GNs), referring to some data centres, that
U1 in Fig. 1 is covered with a lower QoC than U2 . The reason need to collect some required information from UAVs timely.
is that since U1 is on the boundary of the vision cone, a small The UAVs are required to maintain valid links with GNs all
movement of either the target or the UAV may lead to the loss the time, so that any required information can be delivered to
of such a target. In contrast, U2 is far from the boundary and the data centres without significant delays. Also, some pairs
the coverage of this target is not sensitive to a small movement. of UAVs are required to always keep communication links
Moreover, if a target is closer to the boundary of the vision between each other. Furthermore, the UAVs need to avoid
cone, the measurement noise is bigger. Therefore, although collisions with each other [9], and they are not allowed to
enter some no-flight areas.
This work is supported by Australian Research Council. The main contributions of this paper are the coverage model
H. Huang and A.V. Savkin are with School of Electrical Engineering and
Telecommunications, University of New South Wales, Sydney 2052, Australia. and the reactive collision free 3D deployment algorithm. The
(E-mail: {hailong.huang, a.savkin}@unsw.edu.au). coverage model not only tells whether a target is covered,

1551-3203 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TII.2019.2913683, IEEE
Transactions on Industrial Informatics
2

but also specifies how well it is covered. The algorithm is of UAVs, while maximizing the fulfilment of their data rate
distributed and based on local measurements. It consists of two requirements.
navigation laws, both of which are easily implementable. We Besides the surveillance coverage and wireless communi-
give a mathematically rigorous proof of the convergence of the cation coverage, another category is sensing coverage, which
algorithm, and analyse its complexity. Computer simulations is well studied in mobile sensor networks [21]. Three main
are conducted to show the performance of the developed types of sensing coverage have been comprehensively consid-
algorithm. ered: barrier coverage [22], sweep coverage [23] and blanket
The rest of the paper is organized as follows. Section II coverage [24]. In barrier coverage, an array of sensing nodes
discusses some relevant publications. Section III introduces is arranged to form a sensing barrier so that any intrusion
the system model, formally states the studied optimization through the barrier is detected. In sweep coverage, a number
problem, and presents the proposed deployment algorithm. of sensing nodes move across the area to detect and search for
Section IV presents simulation results and finally, Section V objects of interest in the field. The idea of blanket coverage
concludes the paper. is to deploy a network of sensing nodes to monitor an area
of interest such that any object appearing in the area can be
detected.
II. R ELATED W ORKS
There are several key differences between this paper and
In this section, some related publications are reviewed first. the aforementioned ones. Firstly, the coverage in surveillance
Then, the similarities and differences between this paper and applications using UAVs focuses on a set of targets, rather than
the reviewed publications are discussed to show the novelty the area of interest in sensing coverage. Secondly, in terms of
of this paper. the coverage model, most of these publications simply follow
In the surveillance applications using UAVs, the cover- the vision cone model, while the model proposed by this
age performance is a key system metric. Several interesting paper distinguishes the QoC of targets at different positions.
problems have been studied. To track a single target, the Thirdly, beyond the existing approaches for the 2D case like
reference [10] proposes a simple control strategy to reach a [3], [15], [17], this paper considers deploying UAVs in 3D
given target view angle; and the reference [11] focuses on space. Although some publications have already considered
optimal routing for two UAVs cooperatively tracking a moving the 3D case like [6], [7], [8], [12], they either discretize the
target. For multiple targets, in [12], the problem of minimizing space into grids or select positions for UAVs from a given set
the number of UAVs to fully cover a given set of targets of candidates. Thus, the mixed integer linear programming is
is considered. In [6], the authors study the joint problem of often used to formulate the optimization problems. In contrast,
coverage, connectivity, and charging UAVs. In [7], the authors the algorithm presented in this paper finds positions for UAVs
address the optimization problem of covering a set of targets in 3D space without discretization. Furthermore, most of the
with a fleet of UAVs. The goal is to deploy a connected set available approaches for UAVs deployment are proactive and
of UAVs continuously monitoring the targets, and collecting centralized. They compute the positions of UAVs before the
or sending data to them. As an extension, the authors of [8] real deployment and they require the global information as
consider the problem of monitoring a set of mobile sensors. input. The references like [3], [17], [18] have proposed some
The authors of [13] provide a power efficient and reliable decentralized algorithms, however, they are not suitable to
scheduling by adjusting the UAVs’ positions to ensure the implement on-line, due to their assumptions of knowing the
surveillance of all the targets. user density. This paper aims at developing a reactive and
The coverage problem related to UAVs has also been studied decentralized algorithm that navigates each UAV based on the
in the application of providing wireless communication to local measurements rather than the global information.
cellular users. The paper [14] focuses on the modelling of
air-to-ground wireless communication channel and studies the III. P ROBLEM S TATEMENT AND M AIN R ESULTS
optimal altitude of the UAV to maximize the user coverage. Consider a system consisting of n UAVs labelled i =
The reference [15] considers the problem of deploying the 1, 2, . . . , n and k ground nodes (GNs) labelled h = 1, 2, . . . , k.
minimum number of UAVs at the same altitude to cover all Some UAVs and GNs should be able to communicate to each
the users. The paper [16] uses a K-means clustering algorithm other according to a given connected graph G which is called
to partition the ground users into K clusters which are served the communication graph. The graph G has n + k vertices
by K UAVs to offload the demand at the base stations. The v1u , v2u , . . . , vnu , v1g , . . . , vkg that correspond to n UAVs and k
publication [17] aims at maximizing user coverage and mini- GNs, where the superscript u represents UAV and g represents
mizing the communication cost between the UAVs. The paper GN. Vertices viu and vhu are connected by an edge if the UAVs
[18] considers the issue maximizing the coverage of users and i and h should be able to communicate. Vertices viu and vhg are
proposes a centre of mass based distributed algorithm. With connected by an edge if the UAV i and the GN h should be
the objective of maximizing the number of covered targets, able to communicate. GNs are not required to communicate
the authors of [19] propose a reactive algorithm to navigate to each other. For instance, in the illustrative example shown
each flying robot based on virtual forces including hotspots in Fig. 1, the UAV on the right hand side is connected to the
attractive force, target attractive force, nearby robot repulsive UAV on the left, and the latter is also connected to a GN.
force, and obstacle repulsive force. The authors of [20] propose Let (x, y) be Cartesian coordinates on the ground plane and
a scheme to cover all ground users using a minimum number z be the coordinate axis perpendicular to the ground plane.

1551-3203 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TII.2019.2913683, IEEE
Transactions on Industrial Informatics
3

Moreover, let D be some given bounded closed area on the centred at (xi , yi ). Assume that if the distance between the
ground with a smooth boundary and the projections of UAVs target j and the projection (xi , yi ) of the UAV i on the ground
can be inside or on the boundary of D, rather than outside it. plane (x, y) is no greater than ri , then the UAV i knows the
Also, let Zmin and Zmax be given minimum and maximum coordinates Uj of the target j.
altitudes for UAV deployment, Zmax > Zmin > 0. Assume Let w(d) be a given differentiable weighting function such
that UAVs can be deployed only at some points P1 , P2 , . . . , Pn that w(d) and its derivative w0 (d) are defined for d ≥ 0 and
of the region D × [Zmin , Zmax ], i.e. if (xi , yi , zi ) are the co- satisfy the following conditions:
ordinates of the UAV i, then (xi , yi ) ∈ D, zi ∈ [Zmin , Zmax ]
w(d) > 0 ∀d ∈ [0, 1), w(d) = 0 ∀d ≥ 1;
for all i. Notice that D may have ”holes” that may corre-
0 0 0
spond, for example, to tall buildings or some other areas w (0) = w (1) = 0, w (d) < 0 ∀d ∈ (0, 1). (5)
over which UAVs are not allowed to be deployed. The region Furthermore, let g(z) be a given differentiable weighting
D × [Zmin , Zmax ] is called the UAV deployment region. function such that g(z) and its derivative g 0 (z) are defined
Considering the limitation of communication range, it is for all altitudes z ∈ [Zmin , Zmax ] and satisfy the following
required that the UAVs and GNs, which should be able to conditions:
communicate with each other according to G, are within cer-
tain ranges to have reliable wireless links. Consider the motion g(z) > 0, g 0 (z) < 0 ∀z ∈ [Zmin , Zmax ]. (6)
of UAVs at any time satisfying the following constraints:
Examples of w(d) and g(z) satisfying (5) and (6) are the
|Pi , Ph | ≤ R1 (1) following functions:
if the UAVs viu , vhu are connected by an edge in G; and w(d) = cos(πd) + 1 ∀d ∈ [0, 1); w(d) = 0 ∀d ≥ 1,
1
|Pi , Qh | ≤ R2 (2) g(z) = ∀z ∈ [Zmin , Zmax ]. (7)
z
if the UAV viu and the GN vhg are connected by an edge Definition III.1. QoC of target j from UAV i: For target j
in G. Here Qh is the fixed location of GN vhg , and |·, ·| de- (j = 1, 2, . . . , N ), the QoC from the UAV i is defined by the
notes the standard 3D Euclidean distance between two points. dj
function g(zi )w( rii ), where ri is defined by (4), and dji is the
R1 > 0 and R2 > 0 are some given constants describing the distance between the target j and the projection (xi , yi ) of the
communication ranges between UAVs and a UAV and a GN, UAV i to the ground plane.
beyond which the wireless links may be unreliable. Notice
dj
that the conditions (1) and (2) are under the assumption that It is clear that g(zi )w( rii ) = 0 if the target is outside of
the wireless signal is in a perfect communication situation dj
the visibility cone of the UAV i. Moreover, g(zi )w( rii ) is
where packets are always delivered as long as the two nodes dj
are within a certain distance (either R1 or R2 ). The radio increasing as dji is decreasing; and g(zi )w( rii ) is increasing
propagation model will be considered in the future study. as zi is increasing. Hence, the closer the UAV i is to the
Furthermore, the probability of physical collision between target j (the smaller the altitude of the UAV is), the bigger
dj
UAVs arises, as they move around the UAV deployment g(zi )w( rii ) is. On the other hand, if the altitude of the UAV
region. Let dsaf e > 0 be a given constant describing the is fixed, the closer the target is to the boundary of the visibility
minimum safety distance between UAVs. It is also assumed dj
cone of the UAV, the smaller g(zi )w( rii ) is.
that dsaf e < R1 . The safety requirement is that at any time
the following constraints should hold Definition III.2. QoC of target j: Since a target may be
monitored by more than one UAVs at the same time, the QoC
|Pi , Ph | ≥ dsaf e (3) of this target is defined as the largest QoC among all the UAVs.
For target j, introduce the pair (z j , dj ) such as
for all i 6= h. It is always assumed that at the initial time, the
safety condition (3) holds for all i 6= h. The requirement (3) dj dji
allows to avoid collisions of drones with each other, which is g(z j )w( ) = max g(zi )w( ), (8)
rj i=1,...,n ri
very important in navigation and deployment, see e.g. [25],
[26], [27], [28], [29]. In practice, the constant dsaf e may be i.e. (z j , dj ) corresponds to the UAV which monitor the target
dj
different for different types of UAVs. The value of this constant j with the best QoC in sense of the measure g(zi )w( rii ).
is selected based on properties of a specific UAV type and Definition III.3. QoC by the UAVs: For the targets’ loca-
should be sufficient to avoid any possible collisions between tions U1 , U2 , . . . , UN , the QoC by the UAVs placed at points
UAVs, see e.g. [30], [31], [32]. P1 , P2 , . . . , Pn is described by the following function:
There are N ground point-wise targets labelled 1, 2, . . . , N
with the coordinates U1 , U2 , . . . , UN where Uj ∈ R2 . Notice X dj
W (P1 , . . . , Pn ) := g(z j )w( j ). (9)
that the targets may be outside of the region D. r
j=1,...,N
Assume that all UAVs have a given observation angle 0 <
α < π, so that the UAV i with the coordinates (xi , yi , zi ) can It is also assumed that each UAV knows the coordinates of
only see ground targets that are inside of the circle of radius some other UAVs that are close enough to it. For this purpose,
α it is assumed that each UAV is equipped with a GPS and it
ri := tan( )zi (4) knows its own current position. Every a certain period of time,
2

1551-3203 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TII.2019.2913683, IEEE
Transactions on Industrial Informatics
4

it broadcasts its current position. A UAV can receive at most


ܷଶ
n − 1 messages from other UAVs. Furthermore, each UAV
ߠଶ ܷଵ
knows for each target that is inside its vision cone via some ߠଵ axes
ߠଷ
image processing techniques. With the positions of neighbour ܲଵ
UAVs, whether the minimum in (8) is achieved at this UAV ܷଷ
or not can be known by a UAV.
Now consider the following optimization problem:
Fig. 2: Construction of axes and θj .
W (P1 , . . . , Pn ) → max, (10)
where the maximum is taken for fixed targets’ locations
U1 , U2 , . . . , UN over all P1 , . . . , Pn ∈ D × [Zmin , Zmax ] hence, the maximum in the constrained optimization problem
satisfying the constraints (1), (2), (3). (10), (1), (2), (3) is achieved at some P10 , . . . , Pn0 ∈ D ×
Problem Statement: The optimization problem under con- [Zmin , Zmax ] [33]. Therefore, there exists a non-empty set of
sideration is as follows: for the given targets’ locations points of local maximum, and the global maximum is achieved
U1 , U2 , . . . , UN , the communication graph G, the UAV de- at one of them.
ployment region D × [Zmin , Zmax ], the GNs’ locations Before presenting the proposed algorithm, some properties
Q1 , . . . , Qk , the weighting functions w(d), g(z) and the con- of the local optimal solution to the problem (10), (1), (2), (3)
stants r1 , r2 , dsaf e , move UAVs to locations P1 , . . . , Pn ∈ D are discussed.
that maximize the function (10) so that the constraints (1), (2),
Definition III.4. Let (P10 , P20 . . . , Pn0 ) ∈ D × [Zmin , Zmax ]
(3) are always satisfied.
and (xi , yi , zi ) are the coordinates of Pi0 for any i. The point
The following notations are needed. The set of N targets is
(P10 , P20 . . . , Pn0 ) is called a stationary point of the constrained
divided into n non-intersecting subsets U1 , U2 , . . . , Un where
optimization problem (10), (1), (2), (3), if for any i = 1, . . . , n,
for any i = 1, 2, . . . , n, Ui is the set of targets j for which
at least one of the following five conditions holds:
the distance between the target j and the projection (xi , yi ) of
the UAV i on the ground plane (x, y) is less than ri defined S1: |Pi0 , Ph0 | = R1 for some i such that the corresponding
by (4) and the maximum in (8) is achieved at this UAV i. In UAVs vir , vhr are connected by an edge in G;
cases, when the maximum in (8) is achieved at more than one S2: |Pi0 , Qh | = R2 for some i such that the corresponding
UAV, the target j belongs to the corresponding set Ui with the UAV vir and the corresponding GN vhb are connected by an
smallest index i among all such UAVs. edge in G;
Furthermore, for any i, consider an axes in the ground plane S3: |Pi , Ph | = dsaf e for some i 6= h;
(xi , yi ) departing from the current location of the UAV i into S4: The projection (xi , yi ) of Pi0 to the ground plane (x, y)
some given direction. Let θ be the angle measured in the belongs to the boundary of D;
counter-clockwise direction from this axes. θ takes values in S5: β(θ) = 0 for all θ ∈ [0, 2π) where β(θ) is defined by
the interval [0, 2π). Now for all j ∈ Ui , θj denotes the angle (11), and at least one of the following three conditions holds:
between the axes and the direction to the location of the target S5A: zi = Zmin and γ(xi , yi , zi ) < 0 where γ(xi , yi , zi )
Uj , see Fig. 2. Let β(θ) be the function defined as is defined by (12);
S5B: zi = Zmax and γ(xi , yi , zi ) > 0;
X dj
β(θ) := − w0 ( ) cos(θj − θ). (11) S5C: γ(xi , yi , zi ) = 0.
r
j∈Ui
Proposition III.1. If (P10 , P20 . . . , Pn0 ) ∈ D × [Zmin , Zmax ]
Moreover, let γ(x, y, z) be the function defined as is a local maximum in the constrained optimization problem
(10), (1), (2), (3), then (P10 , P20 . . . , Pn0 ) is a stationary point.
X dj
γ(x, y, z) := g 0 (z) w( )
tan( α2 )z Proof. We prove that if (P10 , P20 . . . , Pn0 ) is a point of local
j∈Ui
maximum and for some i all the conditions S1 – S4 do
g(z) X j 0 dj not hold, then the condition S5 holds. First we notice that
− α 2 d w( ). (12)
tan( 2 )z tan( α2 )z if β(θ) < 0 for some θ ∈ [0, 2π), then β(θ̂) > 0 where θ̂ cor-
j∈Ui
responds to the direction opposite to θ. Therefore, if S5 does
Remark III.1. The functions (11) and (12) are defined to
not hold, then β(θ) > 0 for some θ ∈ [0, 2π). Consider now
navigate UAVs in horizontal direction and vertical direction,
the straight line departing from the point Pi0 in the direction θ
respectively. Specifically, β(θ) is proportional to the deriva-
in the plane (x, y) parallel to the ground plane. Moreover, con-
tive of the objective function W (P1 , . . . , Pn ) with respective
sider the movement of along this line with some constant speed
to the distance between the UAV and its covered targets
v > 0. Then the derivative Ẇ (P1 , . . . , Pn ) of the function (9)
in the horizontal plane; and γ(x, y, z) is the derivative of
along this trajectory satisfies Ẇ (P1 , . . . , Pn ) = vβ(θ) > 0.
W (P1 , . . . , Pn ) with respective to the z coordinate in the
Hence, for some small movement of the point Pi0 along this
vertical direction. We will show how to use them to navigate
straight line, the value W (P1 , . . . , Pn ) will increase, and since
UAVs in the proposed algorithm (A1) – (A3).
S1 – S4 do not hold, the constraints (1), (2) and (3) will still
The function W (P1 , . . . , Pn ) defined for all P1 , . . . , Pn ∈ hold, and the coordinates (xi , yi ) of Pi0 will still be in D. This
D × [Zmin , Zmax ] is a continuous function on a compact set, means that if the conditions S1 – S4 together with the first of

1551-3203 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TII.2019.2913683, IEEE
Transactions on Industrial Informatics
5

conditions S5 do not hold, we can make a small motion of Pi0 𝑑𝑠𝑎𝑓𝑒


𝑑𝑠𝑎𝑓𝑒
so that the function (10) increases. Hence, (P10 , P20 . . . , Pn0 ) 𝑃2
𝒫1
is not a local maximum. Now we prove that if all condition 𝒟 𝑃2
𝑃1 𝑃1 𝑄1
S1 – S4 do not hold, then at least one of three conditions 𝒟
𝒬1 𝒫1
S5A – S5C. Indeed, if all these three conditions and S1 – 𝑄1 𝒬1

S4 do not hold, we obviously can make a small motion of Pi (a) (b)


along the vertical axis zi so that the function (10) increases
and all the constraints of the constrained optimization problem 𝑃2 𝑑𝑠𝑎𝑓𝑒
still hold. Hence, (P10 , P20 . . . , Pn0 ) is not a local maximum.

𝛾(𝑥, 𝑦, 𝑧)
𝒟 𝑍𝑚𝑖𝑛 𝑧
The statement of Proposition III.1 immediately follows from 𝑃1 0
𝒫1 𝑍𝑚𝑎𝑥
this. 𝒬1 𝑄1

(c) (d)
To describe our algorithm, some more notations are needed.
Let Ph and Qh denote the set of points Pi satisfying the Fig. 3: Several typical situations a UAV may meet during
constraint (1) and (2) for a particular h where i and h are the movement. (a) When it is at P2 , it is within R2 to Q1
connected in the communication graph G. It is obvious that and outside dsaf e to P1 . The shaded area is the set P̂1 , i.e.,
such sets are balls and their boundaries are spheres of radius removing the circular area centred at P1 with the radius dsaf e
R1 and R2 , respectively. Moreover, let P̂h denote the set of from D. Thus, the coverage circle is the candidate set to find
points Pi satisfying the constraint (3) for a particular h. It is the best angle θ0 . (b) When it is on the boundary of P1 , the
obvious that such sets are exteriors of spheres of radius dsaf e . candidate set is the semi-circle Θ1 = [0, 0.45π] ∪ [1.45π, 2π).
The proposed algorithm is as follows. (c) The UAV is on the boundaries of Q1 and P̂1 . Then, the
A1: Start with some initial points P1 , P2 . . . , Pn ∈ D × candidate for selecting a future moving direction is the green
[Zmin , Zmax ] satisfying the constraints (1), (2) and (3). For sector: Θ2 = [0.83π, 1.2π]. (d) When the UAV moves in the
any i, continuously move the point Pi as follows: vertical direction, it keeps check the sign of γ(x, y, z) and
A2: For each current location of Pi = (xi , yi , zi ), move Pi it always moves towards the direction such that γ(x, y, z)
in the plane (x, y) keeping the constant altitude zi as follows: becomes closer to 0.
(i) if Pi is an interior point of all the sets D×zi , Ph , Qh , P̂h ,
find a value of θ0 ∈ [0, 2π) at which the maximum of the
function β(θ) defined by (11) is achieved. It is obvious that
A3, several typical situations a UAV may meet during the
β(θ0 ) ≥ 0. Continuously move the point Pi along a trajectory
movement are shown in Fig. 3.
we build so that the slope of the tangent to this trajectory is
always equal to θ0 . Stop this when Pi reaches a point beyond Proposition III.2. The algorithm (A1) – (A3) converges to a
which it is impossible to continue the trajectory of Pi so that stationary point of the constrained optimization problem (10),
(xi , yi ) remains in D, and the conditions (1), (2), (3), β(θ0 ) ≥ (1), (2), (3) and the constraints (1), (2), (3) hold at any point
0 hold; of the trajectory we build in (A1) – (A3).
(ii) if Pi is on the boundary of one of the sets D ×
zi , Ph , Qh , P̂h , then build a trajectory as in (i) with the circle Proof. The algorithm (A1) – (A3) generates a trajectory
of possible directions θ ∈ [0, 2π) replaced by the semi-circle (P1 , P2 , . . . , Pn ). Since this trajectory is in some compact set
cut by the tangent to the boundary of the corresponding set D×[Zmin , Zmax ], it has a subsequence that converges to some
(this guarantees that we will move Pi only in a direction that (P10 , P20 , . . . , Pn0 ) [33]. Furthermore, it follows from (A1) –
does not leave D, Ph , Qh , P̂h , see Fig. 3b). Such a feasible (A3) that when (P1 , P2 . . . , Pn ) in some small neighbourhood
semi-circle is denoted by Θ1 ; of (P10 , P20 . . . , Pn0 ), it will reach (P10 , P20 . . . , Pn0 ) and stays
(iii) if Pi is on the boundary of two or more of the sets there. Moreover, it is obvious from (A1) – (A3), that it the
D × zi , Ph , Qh , P̂h , then build a trajectory as in (i) with the trajectory cannot be continued beyond a certain point, then this
circle of possible directions θ ∈ [0, 2π) replaced by the sector point is stationary. When we move Pi , the condition β(θ0 ) ≥ 0
of this circle that is the intersection of the straight angles made always holds. Therefore, the function (9) is non-decreasing on
by the tangents to the boundaries of the corresponding sets, the trajectory (P1 , P2 , . . . , Pn ).
see Fig. 3c. Such a feasible sector is denoted by Θ2 .
A3: If Pi cannot move further in the plane (x, y), we move it
along the vertical axis zi keeping (xi , yi ) constant as follows:
if γ(xi , yi , zi ) < 0, we move Pi down, if γ(xi , yi , zi ) > 0, we
More details about the solution to the problem of maxi-
move Pi up. So we move Pi until either γ(xi , yi , zi ) = 0 or it
mizing (11) are presented below. According to (7), the term
is impossible to continue the trajectory so that the conditions
w0 (dmin
j ) can be expressed as follows:
(1), (2), (3), zi ∈ [Zmin , Zmax ] hold. Then go to A2.
To facilitate the understanding of this algorithm, the pseu-
docode running at each UAV is shown in Algorithm 1. Also, π π
corresponding to the cases of A2 (i), A2 (ii), A2 (iii), and w0 (dj ) = − sin( dj ), ∀dj ∈ [0, r). (13)
r r

1551-3203 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TII.2019.2913683, IEEE
Transactions on Industrial Informatics
6

Algorithm 1 The algorithm running at UAV i.


600
2 4
1: while (1), (2), (3), β(θ0 ) ≥ 0 hold do 6 10
5
11
2: Compute Ph , Qh , and P̂h . 400
7

Y (m)
12
3: if Pi is inside D × zi ∩ Ph ∩ Qh ∩ P̂h then 13 9

4: Find θ0 ∈ [0, 2π) that maximizes (11). 200 1


8
14 4

5: else if Pi is on the boundary of one of the sets D × 1


2 3
3
15
zi , Ph , Qh , P̂h then 0
0 200 400 600 800
6: Find θ0 ∈ Θ1 that maximizes (11). X (m)
7: else if Pi is on the boundary of two or more of the Fig. 4: The area of interest.
sets D × zi , Ph , Qh , P̂h then
8: Find θ0 ∈ Θ2 that maximizes (11).
9: end if
10: Pi moves in the horizontal plane along θ0 for a unit sets of Ph , Qh , and P̂h . Correspondingly, the complexities
distance. are O(n), O(k) and O(n) in the worst case. When the UAV
11: end while is in the case of (i) in A2, it can find the optimal moving
12: while γ(xi , yi , zi ) 6= 0 and (1), (2), (3), zi ∈ direction analytically by solving (14); while when it is in the
[Zmin , Zmax ] hold do cases of either (ii) or (iii) in A2, it needs to check m possible
13: Compute Ph , Qh , and P̂h . angles (if the maximum turning angle of the UAV is 2π m ). Then,
14: if γ(xi , yi , zi ) < 0 then the complexity is either O(1) or O(m). When the UAV is in
15: Pi moves up for a unit distance. A3, to determine the vertical moving direction, it only needs
16: else if γ(xi , yi , zi ) > 0 then to evaluate the sign of (12). Then, the complexity is O(1).
17: Pi moves down for a unit distance. Hence, the overall computational complexity for one UAV is
18: end if either O(3n + k + 1) or O(3n + k + m).
19: end while Although the computational complexity for one UAV is
linearly related to the number of UAVs in the network, it is
worth pointing out that such complexity is for the worst case
Substitute (13) into (11), one obtains: where a UAV receives the broadcast from all the other UAVs.
In practice, the UAVs that are too far away from a UAV will
π X π
β(θ) = sin( dj ) cos(θj − θ) not impact its movement decision; while only those which are
r r connected with this UAV in the communication graph G and
j∈Ui
π X π those which are within dsaf e influence its future movement.
= sin( dj )(cos(θ) cos(θj ) + sin(θ) sin(θj ))
r r In other words, a UAV makes movement decision only based
j∈Ui
π X on the measured locations of targets and some nearby UAVs’
= φj cos(θ) + ψj sin(θ) positions no matter how many UAVs the network has. In
r
j∈Ui practice, a threshold can be applied, such that if the received
= Φ cos(θ) + Ψ sin(θ) signal at a UAV is below this threshold, it will be ignored.
p
= Φ2 + Ψ2 cos(θ − ϑ). (14)
IV. S IMULATIONS
where φjP= sin( πr dj ) cos(θj ), ψP π j
j = sin( r d ) sin(θj ), The performance of the proposed algorithm is demonstrated
π π
Φ = r j∈Ui φj , and Ψ = r j∈Ui ψj . If Φ 6= 0, by computer simulations carried out with Matlab.
ϑ = tan−1 ( Ψ π
Φ ) + mπ; otherwise, ϑ = sign(Ψ) 2 + mπ, where We consider an area with the size of 800 m by 600 m as
sign(Ψ) = 1, if Ψ > 0; sign(Ψ) = −1, otherwise. Notice shown in Fig. 4. In this area, there are four GNs located at the
that the value of integer m depends on the signs of Φ and Ψ, corners (marked by black squares). There are totally 15 depots,
and should make ϑ ∈ [0, 2π) hold. each of which stores one UAV (marked by green circles). The
So far, how UAVs should move in the plane (x, y) is topology of a given communication graph G is also shown in
determined by the solution to a simple one variable equation Fig. 4, where the UAV-GN link is shown by the blue dash
(14). If the UAV i is in the case of (i) of A2, i.e., the possible lines and the UAV-UAV link is shown by the red solid lines.
directions θ ∈ [0, 2π), one can obtain the solution θ0 = ϑ; The UAV deployment region D is an ellipse represented by
otherwise, in the case of (ii) or (iii) of A2, UAV i needs to the black solid line and the area has two holes shaded by the
choose θ0 from the candidate sectors to maximize (14), which blue solid lines where UAVs cannot fly. The system parameters
is also easy to solve. Thus, the control input applied to the are specified as follows: Zmin = 50 m and Zmax = 250 m,
UAV i is u = [cos(θ0 ) sin(θ0 )]0 . R1 = 400 m, R2 = 350 m, dsf ae = 30 m, α = π2 and every
Finally, the complexity of the proposed algorithm is anal- movement in both horizontal and vertical directions is 1 m.
ysed. At each UAV, the proposed algorithm repeats every a Firstly, the simulation result with 15 UAVs is demonstrated.
certain period of time and such a period is consistent with There are 200 targets randomly deployed in the area and they
that of broadcasting its current position. In each period, it may be outside D, see Fig. 5a. The initial positions of the
receives at most n − 1 messages including the positions of UAVs are the locations of the depots as shown in Fig. 4 and
other UAVs. The complexity is O(n). Then, it computes the the initial altitudes are all 100 m. Fig. 5a shows how these 15

1551-3203 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TII.2019.2913683, IEEE
Transactions on Industrial Informatics
7

800 2 4 Zmax 600 2 4


250 1.35
6 6

Quality of coverage ( W)
10 10
5 5
11 11
7 200

Altitude (m)
7
600 400
Y (m)

1.3

Y (m)
12 13 9 12 13 9
150

400 14 4 4
1 200 1 14
8 100
8
1.25
2 3 Zmin
2 3
15 50 15
200 1 3
0 1 3
200 400 600 800 0 100 200 300 400 0 200 400 600 800 0 20 40 60
X (m) Step X (m) Step

(a) (b) (a) (b)


2 600 2 4
1.4
400
Quality of coverage (W)

Quality of coverage ( W)
10
1.8 5
11 1.35
300 7
Distance (m)

max linked UAV-GN 400


1.6 max linked UAV-UAV

Y (m)
12 13 9
200 min UAV-UAV 1.3
1.4 min UAV-boundary
200 14 4
100 1
8 1.25
1.2
2 3
0 15
1 0 1 3
1.2
0 100 200 300 400 0 100 200 300 400 0 200 400 600 800 0 200 400 600
Step Step X (m) Step

(c) (d) (c) (d)


Fig. 5: Simulation result with 15 UAVs. (a) The horizontal Fig. 7: Simulation results by the baseline method. (a) The
movement. (b) The vertical movement. (c) The QoC. (d) horizontal movement for the static target case. (b) The QoC
Maximum linked UAV-GN distance, maximum linked UAV- for the mobile target case. (c) The horizontal movement for
UAV distance, minimum UAV-UAV distance, and minimum the static target case. (d) The QoC for the mobile target case.
UAV-boundary distance.

records several distance metrics during the movement. It is


800 2 4 Z max
6
10
250 clear that the maximum distance between any pair of linked
5
7
11 UAVs does not exceed R1 , the maximum distance between
Altitude (m)

600 200
any pair of linked UAV and GN does not exceed R2 , the
Y (m)

12 13 9
150
400 1 14 4 minimum distance between any pair of UAVs is always larger
8 100
than dsaf e , and the minimum distance from UAVs to boundary
2 3 Z min
200 1
15
3 50 D is always no smaller than 0.
200 400 600 800 0 200 400 600
X (m) Step Furthermore, we demonstrate the simulation result for the
(a) (b)
above case but the targets are mobile. In particular, each target
1.8
randomly selects one moving direction and when it meets the
400 boundary of the area, it changes a direction rather than leaving
Quality of Coverage (W)

1.6 300 max linked UAV-GN the area. Any movement is bounded by 0.5 m. The simulation
Distance (m)

max linked UAV-UAV

200
min UAV-UAV result for this case is shown in Fig. 6. Different from the
min UAV-boundary
1.4
situation with static targets, the positions of UAVs may not
100
converge to some positions. Thus, we consider running the
1.2
0
simulation for 600 steps. Besides, the QoC is not guaranteed
0 150 300 450 600 0 150 300 450 600
Step Step
to increase all the time, see Fig. 6c. The reason lies in the
movement of targets. However, the proposed algorithm is
(c) (d)
able to lead the UAVs to some positions where better QoC
Fig. 6: Simulation results with 15 UAVs for the case with can be achieved. Similar to the above case, all the links are
mobile targets. (a) UAVs’ trajectories. (b) UAVs’ altitudes. (c) guaranteed, any pair of UAVs keep away from each other, and
The QoC. (d) Maximum linked UAV-GN distance, maximum no UAVs leave the deployment region D.
linked UAV-UAV distance, minimum UAV-UAV distance, and To better assess the performance of the proposed approach, a
minimum UAV-boundary distance. baseline method in [18] is considered for comparison. In [18],
UAVs are deployed at the same altitude and each UAV moves
towards the centre of mass of its covered targets. Here, the
UAVs move (indicated by the red solid lines). It can be seen altitude of UAVs is set as the middle of the allowed deployed
the UAVs neither enter the holes nor cross the boundary of D. space, i.e., 150 m. The simulation results for both the static
The altitudes of the UAVs during the movements are shown targets and mobile targets are shown in Fig. 7, where the
in Fig. 5b. The QoC of the network is shown in 5c. The QoC movements of UAVs are shown in Fig. 7a and 7c, and the
keeps increasing with step and the network converges in about QoCs are shown in Fig. 7b and 7d. For the static case, although
360 steps. The final positions of UAVs increase the QoC by the compared method converges faster, the proposed approach
58% compared to the initial positions. Furthermore, Fig. 5d outperforms of the compared one by about 35% in terms of the

1551-3203 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TII.2019.2913683, IEEE
Transactions on Industrial Informatics
8

2
Proposed, static targets
this algorithm consists of two navigation laws for horizontal
Compared, static targets sub-movement and vertical sub-movement. The convergence

Quality of coverage ( W)
Proposed, mobile targets
1.5
Compared, mobile targets of this algorithm was proved. Simulation results demonstrated
1 the performance of the algorithm.
Since the UAVs may operate at different altitudes, their
0.5
energy consumptions may be different as well. Considering
0
the fact that commercial UAVs are now usually powered
5 10 15
Number of UAVs
by batteries with limited capacity, combining some energy
consumption model with the proposed coverage model is a
Fig. 8: Average final QoC in cases with static targets and promising way to improve the UAV network operation and an
average approximate final QoC in cases with mobile targets. important direction for future work.
R EFERENCES
QoC. Since the compared method is not aiming at improving [1] H. Huang and A. V. Savkin, “Towards the internet of flying robots: A
QoC, instead its objective is to reduce the average UAV-target survey,” Sensors, vol. 18, no. 11, p. 4038, 2018.
[2] Z. Zhou, C. Zhang, C. Xu, F. Xiong, Y. Zhang, and T. Umer, “Energy-
distance, the resulting QoC does not increase with step like the efficient industrial internet of UAVs for power line inspection in smart
proposed algorithm shown in Fig. 5c. For the mobile case, the grid,” IEEE Transactions on Industrial Informatics, vol. 14, no. 6, p.
proposed algorithm outperforms the compared one by about 2705, 2018.
[3] H. Huang and A. V. Savkin, “A method for optimized deployment
25% in QoC. of unmanned aerial vehicles for maximum coverage and minimum
Moreover, the proposed algorithm and that of [18] are also interference in cellular networks,” IEEE Transactions on Industrial
applied to the cases with other numbers of UAVs. In each of Informatics, 2018.
[4] J. Gu, T. Su, Q. Wang, X. Du, and M. Guizani, “Multiple moving
these cases, the UAVs’ initial positions are selected from the targets surveillance based on a cooperative network for multi-UAV,”
depots as shown in Fig. 4. Also, the connectivity requirement IEEE Communications Magazine, vol. 56, no. 4, pp. 82–89, April 2018.
should be satisfied at any UAV. In other words, if one depot [5] M. Erdelj, E. Natalizio, K. R. Chowdhury, and I. F. Akyildiz, “Help from
the sky: Leveraging UAVs for disaster management,” IEEE Pervasive
is selected, the parent depot should be selected as well. For Computing, vol. 16, no. 1, pp. 24–32, Jan 2017.
instance, the depot with the index of 13 can only be used if and [6] A. Trotta, M. Di Felice, F. Montori, K. R. Chowdhury, and L. Bononi,
only if the one indexed by 11 is selected. For a certain number “Joint coverage, connectivity, and charging strategies for distributed
UAV networks,” IEEE Transactions on Robotics, vol. 34, no. 4, pp.
of UAVs, there are some different and feasible topologies. We 883–900, 2018.
randomly pick 10 topologies to conduct the below simulations [7] C. Caillouet and T. Razafindralambo, “Efficient deployment of connected
and the results shown below are the best one among these 10 unmanned aerial vehicles for optimal target coverage,” in Global Infor-
mation Infrastructure and Networking Symposium (GIIS). IEEE, 2017,
topologies. We mainly pay attention to the metrics of the final pp. 1–8.
QoC in the cases with static targets and the approximate final [8] C. Caillouet, F. Giroire, and T. Razafindralambo, “Optimization of
QoC in the cases with mobile targets. The second metric is the mobile sensor coverage with UAVs,” in Conference on Computer Com-
munications Workshops. IEEE, 2018, pp. 622–627.
average QoC of last 300 steps out of the overall 600 steps. In [9] I. Mahjri, A. Dhraief, A. Belghith, and A. S. AlMogren, “Slide:
the cases with static targets, the targets are randomly deployed. A straight line conflict detection and alerting algorithm for multiple
To mitigate the influence of the randomness, 10 simulations unmanned aerial vehicles,” IEEE Transactions on Mobile Computing,
vol. 17, no. 5, pp. 1190–1203, 2018.
are conducted and the final QoC in each case is the averaged [10] P. Theodorakopoulos and S. Lacroix, “A strategy for tracking a ground
value. In the cases with mobile targets, the mobility model is target with a UAV,” in IEEE/RSJ International Conference on Intelligent
consistent with the above case. Also 10 simulations are run Robots and Systems, Sep. 2008, pp. 1254–1259.
[11] S. A. P. Quintero, F. Papi, D. J. Klein, L. Chisci, and J. P. Hespanha,
and the approximate final QoC in each case is the averaged “Optimal UAV coordination for target tracking using dynamic program-
value. ming,” in the 49th IEEE Conference on Decision and Control (CDC),
The number of UAVs is set from 5 to 15 and the com- Dec 2010, pp. 4541–4546.
[12] L. D. P. Pugliese, F. Guerriero, D. Zorbas, and T. Razafindralambo,
parisons are displayed in Fig. 8. It is clear that the proposed “Modelling the mobile target covering problem using flying drones,”
algorithm achieves better final QoC in the cases with static Optimization Letters, vol. 10, no. 5, pp. 1021–1052, 2016.
targets and approximate final QoC in the cases with mobile [13] D. Zorbas, T. Razafindralambo, D. P. P. Luigi, and F. Guerriero, “Energy
efficient mobile target tracking using flying drones,” Procedia Computer
targets, and in both situations, the former outperforms the latter Science, vol. 19, pp. 80 – 87, 2013.
by about 20%-35%. Also we can see that with the increase [14] A. Al-Hourani, S. Kandeepan, and S. Lardner, “Optimal LAP altitude
of the number of UAVs, the gain achieved by the proposed for maximum coverage,” IEEE Wireless Communications Letters, vol. 3,
no. 6, pp. 569–572, Dec 2014.
algorithm against the compared tends to be larger. In other [15] J. Lyu, Y. Zeng, R. Zhang, and T. J. Lim, “Placement optimization
words, for the scenarios where more UAVs are available, the of UAV-mounted mobile base stations,” IEEE Communications Letters,
proposed method provides much better QoC. vol. 21, no. 3, pp. 604–607, March 2017.
[16] B. Galkin, J. Kibilda, and L. A. DaSilva, “Deployment of UAV-mounted
access points according to spatial user locations in two-tier cellular
V. C ONCLUSION networks,” in Wireless Days (WD), March 2016, pp. 1–6.
[17] H. Huang and A. V. Savkin, “An algorithm of efficient proactive
The paper considered to use a network of UAVs for surveil- placement of autonomous drones for maximum coverage in cellular
lance purpose. A novel coverage model was proposed, which networks,” IEEE Wireless Communications Letters, vol. 7, no. 6, pp.
characterizes the QoC of a target by the UAVs. Based on this 994–997, 2018.
[18] A. V. Savkin and H. Huang, “Deployment of unmanned aerial vehicle
model, a reactive collision free 3D deployment algorithm was base stations for optimal quality of coverage,” IEEE Wireless Commu-
developed. Aiming at maximizing the overall QoC of targets, nications Letters, vol. 8, no. 1, pp. 321–324, 2019.

1551-3203 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TII.2019.2913683, IEEE
Transactions on Industrial Informatics
9

[19] H. Zhao, H. Wang, W. Wu, and J. Wei, “Deployment algorithms for Andrey V. Savkin was born in 1965 in Norilsk,
UAV airborne networks towards on-demand coverage,” IEEE Journal Russia. He received the M.S. and Ph.D. degrees in
on Selected Areas in Communications, vol. 36, no. 9, pp. 2015–2031, mathematics from the Leningrad State University,
2018. Saint Petersburg, Russia, in 1987 and 1991, respec-
[20] S. Sabino and A. Grilo, “Topology control of unmanned aerial vehicle tively. From 1987 to 1992, he was with the Tele-
(UAV) mesh networks: A multi-objective evolutionary algorithm ap- vision Research Institute, Leningrad, Russia. From
proach,” in the 4th ACM Workshop on Micro Aerial Vehicle Networks, 1992 to 1994, he held a Postdoctoral position in
Systems, and Applications. ACM, 2018, pp. 45–50. the Department of Electrical Engineering, Australian
[21] H. Huang, A. V. Savkin, M. Ding, and C. Huang, “Mobile robots in Defence Force Academy, Canberra. From 1994 to
wireless sensor networks: A survey on tasks,” Computer Networks, vol. 1996, he was a Research Fellow in the Department
148, pp. 1–19, 2019. of Electrical and Electronic Engineering and the
[22] T. M. Cheng and A. V. Savkin, “A distributed self-deployment algorithm Cooperative Research Centre for Sensor Signal and Information Processing,
for the coverage of mobile wireless sensor networks,” IEEE Communi- University of Melbourne, Australia. From 1996 to 2000, he was a Senior
cations Letters, vol. 13, no. 11, 2009. Lecturer, and then an Associate Professor in the Department of Electrical
[23] T. M. Cheng, A. V. Savkin, and F. Javed, “Decentralized control of a and Electronic Engineering, University of Western Australia, Perth. Since
group of mobile robots for deployment in sweep coverage,” Robotics 2000, he has been a Professor in the School of Electrical Engineering
and Autonomous Systems, vol. 59, no. 7-8, pp. 497–507, 2011. and Telecommunications, University of New South Wales, Sydney, NSW,
[24] T. M. Cheng and A. V. Savkin, “Decentralized control of mobile sensor Australia. His current research interests include robust control and state esti-
networks for asymptotically optimal blanket coverage between two mation, hybrid dynamical systems, guidance, navigation and control of mobile
boundaries,” IEEE Transactions on Industrial Informatics, vol. 9, no. 1, robots, applications of control and signal processing in biomedical engineering
pp. 365–376, 2013. and medicine. He has authored/coauthored seven research monographs and
[25] H. X. Pham, H. M. La, D. Feil-Seifer, and M. C. Deans, “A distributed numerous journal and conference papers on these topics. Prof. Savkin has
control framework of multiple unmanned aerial vehicles for dynamic served as an Associate Editor for several international journals.
wildfire tracking,” IEEE Transactions on Systems, Man, and Cybernet-
ics: Systems, no. 99, pp. 1–12, 2018.
[26] P. Basu, J. Redi, and V. Shurbanov, “Coordinated flocking of uavs for
improved connectivity of mobile ground nodes,” in IEEE MILCOM
2004. Military Communications Conference, 2004., vol. 3. IEEE, 2004,
pp. 1628–1634.
[27] A. N. Brintaki and I. K. Nikolos, “Coordinated uav path planning using
differential evolution,” Operational Research, vol. 5, no. 3, pp. 487–502,
2005.
[28] B. Zhang, Z. Mao, W. Liu, and J. Liu, “Geometric reinforcement
learning for path planning of uavs,” Journal of Intelligent & Robotic
Systems, vol. 77, no. 2, pp. 391–409, 2015.
[29] G. Regula and B. Lantos, “Formation control of a large group of uavs
with safe path planning,” in 21st Mediterranean Conference on Control
and Automation. IEEE, 2013, pp. 987–993.
[30] A. V. Savkin, A. S. Matveev, M. Hoy, and C. Wang, Safe robot
navigation among moving and steady obstacles. Elsevier, 2015.
[31] M. Hoy, A. S. Matveev, and A. V. Savkin, “Algorithms for collision-
free navigation of mobile robots in complex cluttered environments: a
survey,” Robotica, vol. 33, no. 3, pp. 463–497, 2015.
[32] C. Wang, A. V. Savkin, and M. Garratt, “A strategy for safe 3D
navigation of non-holonomic robots among moving obstacles,” Robotica,
vol. 36, no. 2, pp. 275–297, 2018.
[33] A. N. Kolmogorov and S. V. Fomin, Introductory real analysis. New
York: Dover, 1975.

Hailong Huang was born in China in 1988. He


received the B.Sc. degree in automation, from China
University of Petroleum, Beijing, China, in 2012,
and received Ph.D degree in Systems and Control
from the University of New South Wales, Sydney,
Australia, in 2018. He is a research associate at
the School of Electrical Engineering and Telecom-
munications, University of New South Wales, Syd-
ney, Australia. His current research interests include
wireless sensor networks, networking protocols, and
guidance, navigation, and control of mobile robots.

1551-3203 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

You might also like