This article has been accepted for inclusion in a future issue of this journal.

Content is final as presented, with the exception of pagination.

IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS 1

A Clustering-Based Coverage Path Planning

Method for Autonomous Heterogeneous UAVs
Jinchao Chen , Member, IEEE, Chenglie Du, Member, IEEE, Ying Zhang ,
Pengcheng Han, and Wei Wei , Senior Member, IEEE

Abstract— Unmanned aerial vehicles (UAVs) have been widely

applied in civilian and military applications due to their high
autonomy and strong adaptability. Although UAVs can achieve
effective cost reduction and flexibility enhancement in the devel-
opment of large-scale systems, they result in a serious path
planning and task allocation problem. Coverage path planning,
which tries to seek flight paths to cover all of regions of interest,
is one of the key technologies in achieving autonomous driving
of UAVs and difficult to obtain optimal solutions because of its
NP-Hard computational complexity. In this paper, we study the
coverage path planning problem of autonomous heterogeneous
UAVs on a bounded number of regions. First, with models of
separated regions and heterogeneous UAVs, we propose an exact
formulation based on mixed integer linear programming to fully Fig. 1. Path planning in autonomous driving UAV systems.
search the solution space and produce optimal flight paths for
autonomous UAVs. Then, inspired from density-based clustering
methods, we design an original clustering-based algorithm to
classify regions into clusters and obtain approximate optimal strong robustness and high parallelism, have become one of the
point-to-point paths for UAVs such that coverage tasks would be hot research fields in recent years, by taking full advantages
carried out correctly and efficiently. Experiments with randomly of cooperation and coordination of UAVs [4].
generated regions are conducted to demonstrate the efficiency In large-scale real-time and distributed collaborative sys-
and effectiveness of the proposed approach.
tems, there is an increasing tend toward performing the detect-
Index Terms— Coverage path planning, unmanned aerial vehi- ing or searching tasks by multiple autonomous UAVs with
cle, clustering-base method, autonomous heterogeneous UAVs. different operation capabilities and kinematic constraints [5].
For instance, in the military domain, autonomous UAVs with
I. I NTRODUCTION heterogeneous sensors and detectors are adopted to search
the entire area of interest and locate enemy targets. In the
W ITH the development of electronic and automation
technologies, unmanned aerial vehicle (UAV) plays an
increasingly important role for national economy and human
agriculture domain, multiple UAVs cooperate with each other
in terraces and forests to monitor the growth states of veg-
etation and weeds. Autonomous heterogeneous UAVs adapt
social life in many fields, such as traffic monitoring, disaster
to different types of mission requirements, and dramatically
relief, anti-terror campaign and target acquisition [1]. Com-
improve system performance and conserve energy with respect
pared with manned vehicles, UAV could save human from dull,
to homogeneous or manned vehicles.
dirty and dangerous tasks [2] and carry out facilities efficiently.
Although autonomous heterogeneous UAVs can efficiently
Because of the limited sensing range, calculating speed and
enhance flexibility and adaptability in practical applications,
energy supply, lots of practical applications have gone beyond
they compound difficulty for system designers to provide
the capability of a single UAV [3]. Multiple UAV systems with
reasonable collaboration strategies in achieving functionalities
Manuscript received October 31, 2020; revised January 26, 2021, of path planning, task allocation and intelligent decision. Espe-
February 3, 2021, and February 18, 2021; accepted March 11, 2021. This cially in multiple region systems where geographical zones
work was supported in part by the National Key Research and Develop- are monitored and points of interest concern all information
ment Program under Grant 2017YFB1001900 and Grant 2018YFB2101304.
The Associate Editor for this article was H. Lu. (Corresponding author: of zones [6], coverage path planning which tries to seek valid
Jinchao Chen.) flight paths from start points to end points able to fully cover
Jinchao Chen, Chenglie Du, Ying Zhang, and Pengcheng Han are regions of interest [7], is NP-Hard in the strong sense [8] and
with the School of Computer Science, Northwestern Polytechnical Univer-
sity, Xi’an 710072, China (e-mail: [email protected]; [email protected]; difficult to directly obtain optimal solutions.
[email protected]; [email protected]). As an important part of mission planning in autonomous
Wei Wei is with the School of Computer Science and Engineer- driving UAV systems, path planning plays an ever more
ing, Xi’an University of Technology, Xi’an 710048, China (e-mail:
[email protected]). significant role to improve safety, reliability and adaptability
Digital Object Identifier 10.1109/TITS.2021.3066240 of UAVs [9]. As shown in Fig. 1, in order to efficiently
1558-0016 © 2021 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.

Authorized licensed use limited to: NORTHWESTERN POLYTECHNICAL UNIVERSITY. Downloaded on May 05,2021 at 15:05:35 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

2 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS

achieve the autopilot, path planning has to provide good II. R ELATED W ORK
enough trajectory solutions for the ground control station in Coverage path planning is of great significance to improve
an affordable time, while taking into account a large number the efficiency of UAVs in carrying out the reconnaissance,
of constraints, such as environment, UAV capabilities, task surveillance, search and other tasks by traversing all of regions
constraints and goals. With path planning, the ground control of interest. Basically, path planning of UAVs can be clas-
station can assign tasks to appropriate UAVs and generate sified into three types, i.e., offline, online and cooperative
valid commands to continuously control the moving of UAVs. planning [11], according to the time that path planning is
Path planning of UAVs is essential to ensure that autonomous realized. If the global information about the circumstance
driving missions would be carried out correctly and efficiently. is known in advance, path planning occurs before UAVs
With the vigorous growth of the number of regions and take off and is called offline planning [12]. On the contrary,
UAVs, system experts start to provide valid point-to-point if the environment is partially known or unknown, paths of
fight paths for UAVs with the help of decision-making tools. UAVs should be planned online during the flight [13]. In case
In previous works, many methods have been presented to missions are too many or too complex, cooperative planning
settle the coverage path planning issue with different objectives is adopted to control UAVs in dealing with unforeseeable
and requirements [10]. However, due to the extraordinary changes of the circumstance and achieving the overall goal [9].
computational complexity, most of the existing methods only Among these types, offline planning is extensively studied and
consider the geometrical constraints of a single region, without widely used due to its effectiveness and efficiency. In this
taking into consideration the peculiar features of separated work, we also focus on the offline coverage path planning.
regions and heterogeneous UAVs. Offline coverage path planning of multiple UAVs is more
In this study, we concentrate on the coverage path planning complex and harder to solve than that of a single UAV. It not
of autonomous heterogeneous UAVs on a bounded number of only needs to seek an optimal sequence of way-points for
regions. We aim to provide a convenient and clustering-based each UAV, but also provides a guarantee that the overall mis-
approach to produce reasonable flight paths for UAVs and sion would be performed effectively. Coverage path planning
guarantee that coverage tasks are completed as soon as possi- problem of multiple UAVs can be modelled as an optimization
ble. Main contributions of this paper are as follows: one that tries to find out a feasible flight path from the start
1. Separated regions and heterogeneous UAVs are modelled, point to the end point for each UAV while satisfying the
and an exact formulation based on the mixed integer linear system constraints and goals. There are two major problems
programming (MILP) is proposed to solve the coverage path that should be urgently resolved:
planning problem. This formulation can fully search the solu- 1. How to reasonably decompose a region into several sub-
tion space, and effectively guide the design of autonomous regions for UAVs, such that coverage tasks can be completed
driving UAV systems, by providing best flight paths for at the earliest time?
heterogeneous UAVs. 2. For a given UAV, how to seek an optimal flight path
2. Inspired from density-based clustering methods, we pro- that greatly reduces the time and energy consumption while
pose an original clustering-based algorithm to find the centers satisfying the coverage, resolution and energy constraints?
of regions and classify regions into different clusters according For the first problem, many approaches have been pro-
to their densities calculated via relative distances. It takes the posed to automatically decompose a search region for UAVs,
influence of both the flying features of UAVs and the geo- considering both capabilities of UAVs and sharps of regions.
graphical locations of regions into consideration, and produces Janchiv et al. [14] converted the path planning problem into
an approximate optimal flight path for each UAV with an a flow network, and developed a time-efficient region division
objective of minimizing the completion time of coverage tasks. algorithm by combining exact cell decomposition and template
The proposed approach adds time constraint into matching. Zhang et al. [15] modelled the path planning as
density-based clustering methods, changes the calculation a non-linear optimal problem with non-convex constraints,
strategy of key clustering parameters and would achieve and proposed an algorithm to approximate the non-convex
reasonable solutions in a relatively short time. Although parts by a series of sequential convex programming ones.
this approach is not optimal, it provides a guarantee that Although these approaches can efficiently decompose a region
all of regions of interest would be covered correctly if the into several subregions, most of them adopted the same sweep
deadlines of coverage tasks are not less than the obtained direction strategies into all subregions and optimal flight paths
task completion times. could not be obtained in most case situations.
The remainder is as follows. Related work about the As for the second problem, great attention has been paid
coverage path planning is presented in Section II. System and optimal path solutions have been proposed while con-
models and problem definition are presented in Section III. sidering the influence of various factors, such as vehicle
Section IV analyses the constraints of the coverage path plan- capabilities [16], network communications [17] and system
ning problem and proposes an exact formulation. Section V schedulability [18]. By taking flying speeds and power con-
introduces the clustering-based approach that classifies regions sumption of UAVs into account, Franco and Buttazzo [19]
into clusters and provides valid flight paths for autonomous derived an energy model from practical applications and
UAVs. Section VI builds the simulation experiments and planed flight paths for UAVs while minimizing the energy
shows the results. Section VII presents the conclusions of consumption of UAVs. Yang et al. [11] decomposed the con-
this work. straints and objective functions of the path planning problem

Authorized licensed use limited to: NORTHWESTERN POLYTECHNICAL UNIVERSITY. Downloaded on May 05,2021 at 15:05:35 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

CHEN et al.: CLUSTERING-BASED COVERAGE PATH PLANNING METHOD FOR AUTONOMOUS HETEROGENEOUS UAVs 3

into evaluation functions, and adopted separately evaluating TABLE I

and evolving way-points to settle the problem. Based on M AIN N OTATIONS U SED IN T HIS S TUDY
the extended Kalman filter, Shahidian and Soltanizadeh [20]
provided effective paths for two UAVs by selecting way-points
that can improve the estimation accuracy of location of source
targets. Although all of approaches mentioned above can
obtain reasonable solutions for the path planning problem of
UAVs efficiently, they couldn’t be directly adopted in search
systems with a large number of separated regions. In order
to solve this problem, this paper provides a clustering-based
approach to produce reasonable flight paths for UAVs to fully
cover all of separated regions, by classifying regions into
clusters and assigning clusters to UAVs.

III. S YSTEM M ODEL AND P ROBLEM D EFINITION

A. System Model
In this study, n heterogeneous UAVs U = {U1 , U2 , .., Un } U = {U1 , U2 , .., Un }, and a set of m separated regions R =
are adopted to efficiently carry out the task of fully search- {R1 , R2 , . . . , Rm }, the coverage path planning problem is to
ing m separated regions R = {R1 , R2 , . . . , Rm }. UAVs are find an optimal flight path for each autonomous UAV, such that
autonomous and have different flying speeds, onboard cameras all of regions would be covered in the shortest time, while the
and energy supplies. Each UAV can be characterized by a triple path and energy consumption constraints of UAVs are satisfied.
Ui = Vi , Ti , Wi  where Vi is its maximum flying speed, Since the coverage path planning problem of autonomous
Ti is the maximum flight time before its battery runs out, heterogeneous UAVs is NP-hard in the strong sense, it is
and Wi is the scanning width of its onboard camera. Usually, hard for systems designers to settle directly. In this paper,
the maximum flight time of UAVs depends on lots of influence we first analyze the path and energy constraints of UAVs, and
factors, such as flight environment, battery capacities, flying present an exact formulation based on linear programming to
speeds and payload weights. These factors may change by time fully search the solution space and provide optimal solutions.
and have important actual meanings in studying the intelligent Then, in order to improve the efficiency of problem-solving
control of UAVs. However, in this paper, we neglect these approaches, we design a clustering-based heuristic to allocate
factors since we only focus on the offline path planning of regions to UAVs and obtain approximate optimal flight paths
UAVs and the global information about the circumstance is for UAVs.
known in advance.
IV. E XACT F ORMULATION
We use A j and D j,k to respectively denote the scanning
area of a given region R j and the distance between each two In this section, constraints of the coverage path planning
regions R j and Rk . Due to the heterogeneity property, UAVs problem of multiple autonomous UAVs are analysed, and
may have different flying and scanning speeds even on the an exact formulation based on the Mixed Integer Linear
same region. T Si, j and T Fi, j,k are used to represent the time Programming (MILP) is built to achieve the best flight paths
consumption of Ui in scanning the region R j and in flying for UAVs, with an objective of minimizing the completion
from a region R j to another region Rk , respectively, then time of coverage tasks.
When all of regions of interest are correctly covered by
Aj
T Si, j = UAVs, the following path and energy constraints should be
Vi Wi satisfied:
D j,k
T Fi, j,k = (C1) Each UAV should stay on the training base unless it
Vi is chosen to carry out the coverage task.
In this study, we assume that autonomous UAVs are initially (C2) Once a UAV is chosen to scan the regions, it should
placed in the same training base. UAVs chosen to carry out take off from the training base and fly back to the base after
the coverage task, should take off from the base and fly back all regions assigned to it have been scanned.
to the base after tasks are finished. A virtual region R0 is (C3) In order to insure that missions are achieved efficiently
used to represent the training base. As UAVs do not need without any region being scanned repeatedly or not scanned
to scan the training base at all, the scanning area of R0 is at all, we assume that each real region is scanned by one and
zero, i.e., A0 = 0. Meanwhile, D0, j and D j,0 ( j ∈ [1, m]) only one UAV.
are respectively used to denote the flight distances from the (C4) Each UAV used in the coverage task, should visit and
training base to a given region R j and from R j to the training cover the regions one by one.
base. Main notations adopted in this study are listed in Table I. (C5) Each UAV should land at the training base before its
battery runs out.
B. Problem Definition We use a three-dimensional matrix X = {x i, j,k | i ∈
The problem studied in this study is defined as fol- [1, n], j, k ∈ [0, m]} to express the flight paths of UAVs.
lows: Given a set of n autonomous heterogeneous UAVs Each element x i, j,k is a global boolean variable, and represents

Authorized licensed use limited to: NORTHWESTERN POLYTECHNICAL UNIVERSITY. Downloaded on May 05,2021 at 15:05:35 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

4 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS

whether Ui flies from regions R j to Rk . If Ui flies from R j In order to exclude subtour solutions, Constraint (C4)
to Rk , the value of x i, j,k is 1; otherwise, x i, j,k is equal to 0. requires that all of regions assigned to UAVs should be covered
Thus, X can be calculated via, one by one. We use a two dimensional array S = {si, j | 1 ≤
 i ≤ n, 0 ≤ j ≤ m} to record the visiting orders of regions.
1, if the UAV Ui flies from R j to Rk , and j = k si, j is an integer variable and represents the visiting order of
x i, j,k =
0, otherwise a given region R j in regions allocated to a given UAV Ui .
Since all of UAVs are required to fly from the training base,
In the following, using the flight path matrix X, we formally
the visiting order of the training base R0 for each UAV is 1,
analyze the constraints and objective function of the coverage
i.e., ∀i ∈ [1, n], si,0 = 1. If Ui flies from R j to a real region
path planning problem. Constraint (C1) indicates that every
Rk , the visiting order of R j is one more than that of Rk , i.e.,
UAV can either stay on the training base, or fly from the
training base to carry out the scanning task. Since x i,0, j ∀i ∈ [1, n], ∀ j ∈ [0, m], ∀k ∈ [1, m]
represents whether the UAV Ui flies from the training base x i, j,k = 1 ⇒ si,k = si, j + 1
to a region R j , Constraint (C1) can be expressed by:
Given that the number of regions visited by a single UAV

m
cannot be larger than the total region number, the visiting order
∀i ∈ [1, n], x i,0, j ≤ 1 (1)
j =1
of any region is no more than m. Therefore Constraint (C4)
can be expressed as:
Constraint (C2) indicates that all of UAVs can task off from
the training base at most one time, and requires that UAVs ∀i ∈ [1, n], ∀ j ∈ [0, m], si, j ≤ m (4)
leaving the training base must return back to the training Constraint (C5) requires that UAVs must return back to
base after its scanning task have been finished, which can be the training base before its battery runs out. For each UAV
expressed via, that is chosen to carry out a scanning task, it should take off

m 
m from the training base, cover regions allocated to it, and finally
∀i ∈ [1, n], x i,0, j = x i, j,0 (2) land to the training base. Since T Si, j and T Fi, j,k respectively
j =1 j =1 represent the time consumption of UAV Ui in scanning the
Constraint (C3) demonstrates that each real region should region R j and in flying from R j to Rk , Constraint (C5) is
be visited and scanned by one and only one UAV. It contains written as:
two important aspects: (1) the numbers of UAVs both leaving 
m  m

from and flying to a real region are 1; (2) a UAV either does ∀i ∈ [1, n], x i, j,k (T Si, j + T Fi, j,k ) ≤ Ti (5)
j =0 k=0
not fly to a real region, or flies to and leaves the region after
this region is scanned. Now we compute the task completion time of the multiple
Since x i, j,k and x i,k, j respectively represent whether the UAV system, i.e., the maximum time cost of UAVs in perform-
UAV Ui flies from R j to Rk and from Rk to R j , the first ing the coverage tasks. We use T (Ui ) and T (U, R) to denote
aspect can be expressed via: the time spent by a given UAV Ui and the task completion time
of the whole system, respectively. Then, T (Ui ) and T (U, R)

n 
m 
n 
m
∀ j ∈ [1, m], x i, j,k = 1, x i,k, j = 1 can be calculated via,
i=1 k=0 i=1 k=0 
m 
m
T (Ui ) = x i, j,k (T Si, j + T Fi, j,k )
In the second aspect, if Ui does not fly to a given real region

m j =0 k=0
R j , it cannot fly from R j to any other region, i.e., x i,k, j = T (U, R) = max T (Ui )
k=0 1≤i≤n

m
0 ⇔ x i, j,l = 0. Correspondingly, when Ui flies to R j , The coverage path planning problem studied in this paper
l=0 can be defined as seeking an optimal flight path matrix X for

m 
m
it must leave this region, i.e., x i,k, j = 1 ⇔ x i, j,l = 1. autonomous UAVs such that the completion time of coverage
k=0 l=0 tasks is minimized. The linear objection function is T (U, R),
Therefore, the second aspect can be expressed by:
the linear constrains are Constraints from (C1) to (C5), and

m 
m
the exact formulation can be stated as:
∀i ∈ [1, n], ∀ j ∈ [1, m], x i,k, j = x i, j,l
mi n T (U, R)
k=0 l=0
s.t. (1), (2), (3), (4), (5)
Putting the two aspects together, Constraint (C3) can be
expressed by: Since the unknown variables in the above linear program-
ming are a mixture of integer variables (elements in the flight

n 
m 
n 
m
∀ j ∈ [1, m], x i, j,k = 1, x i,k, j =1 path matrix X and in the visiting order array S) and real
i=1 k=0 i=1 k=0
variables (time cost of UAVs T (Ui ) and task completion time

m 
m T (U, R)), this formulation is an MILP one.
∀i ∈ [1, n], ∀ j ∈ [1, m], x i,k, j = x i, j,l (3) Although the proposed MILP formulation could find an
k=0 l=0 exact solution for the coverage path planning problem, it has

Authorized licensed use limited to: NORTHWESTERN POLYTECHNICAL UNIVERSITY. Downloaded on May 05,2021 at 15:05:35 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

CHEN et al.: CLUSTERING-BASED COVERAGE PATH PLANNING METHOD FOR AUTONOMOUS HETEROGENEOUS UAVs 5

to search all of the solution space and spends a huge amount of Basically, ρ j is equal to the total time cost in flying to and
time in considering the possible flight paths. Its running time scanning all of regions that are closer than dc to region R j .
cost grows explosively along with the increase of numbers The cut-off distance dc has a significant influence over the
of UAVs and regions. This is because of the fact that the calculation of density of each region, and largely determines
coverage path planning of UAVs is NP-Hard and has huge the effectiveness of the center selection phase. From Eq. (6)
solution space. Inspired by density-based clustering methods, we can find that, the density of each region increases when a
we present a relatively efficient heuristic to settle the coverage larger value of dc is adopted. Meanwhile, when the value of dc
path planning problem of multiple UAVs in the following is fixed, for each given region R j , its density ρ j grows along
sections. with the increase of both the scanning areas and the relative
distances of its neighbours within the cut-off distance dc .
V. S PATIAL -T EMPORAL C LUSTERING -BASED A LGORITHM In the STCA algorithm, as did in [21], the value of dc should
In order to improve the efficiency and reliability of be reasonably set such that the average number of regions
problem-solving approaches, we propose a spatial-temporal in a cluster is around 1/n of the total number of regions.
clustering-based algorithm (STCA) to obtain approximately We sort the distances between regions in ascending order, and
optimal flight paths for autonomous heterogeneous UAVs. use {di }m(m−1)
i=1 to denote the ordered distance sequence. Then,
The proposed approach is inspired from the Clustering by the value of dc can be set to the
n1 m(m − 1)th element in
Fast Search and Find of Density Peaks (CFSFDP) algorithm the distance sequence, i.e.,
[21], which is a new density-based clustering method and
tries to classify elements into different clusters according to dc = d
1 m(m−1) (7)
n
their densities calculated via relative distances. The proposed δ j is defined as the minimum distance between region R j
STCA algorithm takes the influence of both flying features of and any other region which has a higher density than R j . For
heterogeneous UAVs and geographical locations of separated the region with the highest density, we set its distance to the
regions into consideration, adds time constraints on the basis of maximum distance to any other region. Therefore, δ j can be
the CFSFDP algorithm, and changes the calculation strategy of computed as follows:
key clustering parameters. With the proposed STCA algorithm, ⎧
regions are clustered and allocated to their best UAVs such that ⎨ min (D j,k ), if ρ j = max ρk ;
k:ρk >ρ j 1≤k≤m
δj = (8)
the task completion time is minimized. ⎩ max D j,k otherwise
The proposed STCA algorithm has its basis on the idea 1≤k≤m
that each cluster center is surrounded by neighbours with low In this paper, cluster centers are recognized as regions that
densities and has a relatively long distance from points with have relatively large values of both the density ρ j and the
higher densities. There are two major phases in the STCA distance δ j . In order to quickly find the cluster centers for
algorithm: a center selection phase for choosing centers for UAVs, two quantities γ and η are used to take into account
region clusters and a region clustering phase for classifying the synthetic influence of ρ and δ:
regions into clusters. The former phase computes the density of
∀ j ∈ [1, m], γ j = ρ j δ j , η j = ρ j + 2T F0, j (9)
each region, and selects the best cluster center for each UAV.
The latter phase classifies regions into clusters according to γ can be treated as an important criterion for the effective
their similarity, and achieves an approximate optimal visiting choice of cluster centers. Clearly regions with larger values
order of regions in the same cluster. of γ are more likely to be identified as centers of regions to
be covered. Therefore, we sort regions in descending order
A. Center Selection Phase according to their values of γ , pick up the regions with top n
Center selection is one of the most important steps in the values, and each selected region is the cluster center of regions
STCA algorithm. In this work, cluster centers are identified as scanned by one UAV.
regions that have higher densities than their neighbours and are η plays a significant role in allocating cluster centers to
relatively far away from regions with higher densities. In order UAVs. In this work, cluster centers are assigned to UAVs
to achieve the best cluster center for each UAV, we calculate sequentially, and regions with larger values of η are regarded
two parameters of each region R j , i.e., its density ρ j and its as the centers of clusters of UAVs that have faster flying
distance δ j , according to related relative distances. The density speeds. We sort the regions and UAVs in descending order
ρ j of region R j is defined as: according to values of η and according to the flying speeds,

m respectively, and the region with the i th (1 ≤ i ≤ n) largest
ρj = χ(D j,k , dc )(T F j,k + T Sk ) + T S j (6) value of η is allocated to the UAV with the i th fastest flying
k=1 speeds. We use an array {Oi }ni=1 to denote the indexes of the
sorted regions, then with this allocation strategy, R Oi is the
where dc is the cut-off distance of regions, T S j and T F j,k
center of the cluster of UAV Ui .
are the average scanning time of region R j and the average
flying time from regions R j to Rk , respectively. χ(D j,k , dc )
B. Region Clustering Phase
is computed via:
 After cluster centers have been chosen, each remaining
0, if D j,k > dc ; region is classified into the same cluster as its nearest neigh-
χ(D j,k , dc ) =
1, otherwise bour. We adopt C = {C1 , C2 , . . . , Cn } to denote the clusters

Authorized licensed use limited to: NORTHWESTERN POLYTECHNICAL UNIVERSITY. Downloaded on May 05,2021 at 15:05:35 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

6 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS

q
of UAVs. For each cluster Ci , a queue i = {πi1 , πi2 , .., πi } is is chosen. We use Rbest to denote the most suitable region
used to denote the indexes of regions that have been clustered chosen, then
into Ci and would be sequentially scanned by UAV Ui . Rπ 1
i Rbest = arg min min{D j,π 1 , D j,π q } (12)
and Rπ q are respectively the head (i.e., the first region) and R j ∈Ravail ,T D(C i )≤Ti i i
i
the tail (the last region) of this cluster.
We use T D(Ci ) to denote the time demand (i.e., time cost in If no available region is found according to Eq. (12),
covering all of clustered regions) of a given cluster Ci . Since we hold that this cluster C x is full and no more region can
UAVs should take off from and return back to the training be added into. In this case, C x should be removed from
base after its task is finished, T D(Ci ) can be computed via: the available cluster set, and a new cluster would be chosen
according to Step 2. On the contrary, if an available region

q−1 
q Rbest is found, Rbest should be removed from the unallocated
T D(Ci ) = T Fi,0,π 1 + T Fi,π l ,π l+1 + T Si,π l + T Fi,π q ,0 region set and added into cluster C x .
i i i i i
l=1 l=1 4) Cluster update. In this step, the selected region Rbest is
(10) moved into the region queue of cluster C x . If Rbest is closer to
the head than to the tail, Rbest is inserted before the head, and
From the above equation we can find that, the time demand
Rbest becomes the new head of the region queue; otherwise,
of clusters grows with the increase of number of regions in
Rbest is placed at the end and becomes the new tail of the
clusters. In order to satisfy the energy constraints of UAVs,
region queue of C x . Meanwhile, the time demand of C x is
a new region can be added into the cluster of a specified
updated according to Eq. (10).
UAV only if the updated time demand is not larger than the
After all regions have been allocated, the region cluster-
maximum flight time of this UAV.
ing phase terminates. Algorithm 1 gives the pseudo-code of
In this work, regions are classified into clusters sequentially,
the proposed spatial-temporal clustering-based algorithm with
and the clustering process is concluded as followings:
the nearest-to-end policy. Regions are sequentially scanned
1) Initialization. For each cluster Ci , its center R Oi is the
according to their orders in region queues of clusters, and the
first region to be added. Therefore, when the region clustering
task completion time of UAVs is the maximum time demand
phase begins, there is only one element in the region queue
of clusters, i.e.,
of each cluster, i.e.,
T (U, R) = max T D(Ci )
∀i ∈ [1, n], i = {Oi } C i ∈C

Meanwhile, all clusters are available and all regions except

the cluster centers are unallocated. We use Cavail and Ravail to C. Running Time Analysis
denote the sets constituted of available clusters and constituted In this subsection, running time cost of the proposed STCA
of unallocated regions, respectively. Then, algorithm with the nearest-to-end policy is analysed. As men-
 tioned, STCA has two major phases: a center selection phase
Cavail = C, Ravail = R \ R Oi
(lines from 1 to 6 in Algorithm 1) and a region clustering
i∈[1,n]
phase (lines from 7 to 23 in Algorithm 1). We analyse the
2) Cluster selection. In order to balance the time demands of running time complexity of the center selection phase first.
clusters, the cluster with the minimum time demand is chosen From Algorithm 1 we can find that, the most
from the available cluster set, and regions are preferably added time-consuming aspect of the center selection phase is
into this cluster. We use C x to denote the selected cluster, then calculating the cut-off distance dc , the density ρ and the
distance γ . Since all of these calculations run in O(m 2 ),
C x = arg min T D(Ci ) (11) the computation complexity of this selection phase is O(m 2 ).
C i ∈C avail
Now we start to analyse the time complexity of the region
3) Region selection. Region that is unallocated and most clustering phase. As can be seen in Algorithm 1, the region
suitable to join into cluster C x is selected. In this paper, clustering phase has a structure of double closed loops. In the
unallocated regions are classified into clusters according to inner loop (lines from 10 to 22), the cluster with the minimum
their relative distances, and basically allocated to the same time demand and the most suitable region for this selected
cluster as their nearest neighbours that have been allocated. cluster are found. According to Eq. (11) and (12), the inner
In order to ensure flight safety, new regions are added only if loop runs in O(n + m). In practical applications, the number
UAVs can finish the coverage tasks before their battery run out. of UAVs is usually far less than that of regions. Hence,
Therefore, the most suitable region is chosen from unallocated the computational of the inner loop is approximately equal to
ones that are nearer to allocated regions and satisfy the energy O(m). Given that the outer loop executes repeatedly m times
constraint of the cluster. at most, the computational complexity of this region clustering
Many strategies can be adopted to find out the suitable phase is O(m 2 ).
region, among which nearest-to-end policy is one of the most Combining the two phases, we can obtain that the compu-
frequently used due to the significant efficiency and reliability. tation complexity of the proposed STCA algorithm with the
In this policy, the unallocated region which is nearest to either nearest-to-end policy is O(m 2 ). Since this running time com-
the head or tail of the region queue of an examined cluster, plexity depends only on the number of regions m, our approach

Authorized licensed use limited to: NORTHWESTERN POLYTECHNICAL UNIVERSITY. Downloaded on May 05,2021 at 15:05:35 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

CHEN et al.: CLUSTERING-BASED COVERAGE PATH PLANNING METHOD FOR AUTONOMOUS HETEROGENEOUS UAVs 7

Algorithm 1 Pseudo-Code of the Spatial-Temporal satisfies the energy constraint and is nearest to any element in
Clustering-Based Algorithm With the Nearest-to-End Pol- the region queue of an examined cluster, is selected and added

into the current cluster. We use Rbest to denote the selected
icy
Input: the UAV set U and the region set R region with the nearest-to-any policy, then
Output: the flight path of each UAV and the task
Rbest = arg min min D j,π y (13)
completion time of UAVs y
R j ∈Ravail ,T (C i )≤Ti πi ∈i i

1 Compute the cut-off distance dc according to Eq. (7);

In this policy, Rbest is inserted into the region queue
2 Compute the density ρ and the distance δ for each region
according to Eq. (6) and Eq. (8); before or after its nearest region according to the updated
3 Compute the quantities γ and η according to Eq. (9);
time demands of the current cluster. If the new time demand
calculated when Rbest is inserted before its nearest region is
4 Sort regions in descending order, top n regions {Oi }i=1
n
smaller, Rbest is placed before its nearest region; otherwise,
are identified as centers;
5 Sort {Oi }i=1 and {Ui }i=1 in descending order according
n n Rbest is placed after its nearest region.
to values of η and according to the flying speeds; (2) Order optimization strategy. In the proposed STCA
6 ∀i ∈ [1, n], Oi is the center of the cluster of UAV Ui ;
algorithm, regions are visited and scanned by UAVs in the
7 ∀i ∈ [1, n], i = {Oi };
order that they appear in the region queues of clusters. The
visiting order of regions is not good enough in most of
8 Cavail = C, Ravail = R \ i∈[1,n] R Oi ;
9 while there are unallocated regions in the region list do
situations. An order optimization strategy is required to adjust
10 repeat the visiting order of regions classified into the same cluster,
11 i ← index of the cluster which is available and has such that the task consumption time of UAVs would be
the minimum time demand; reduced. In this paper, we adopt the genetic algorithm to search
12 (k, l) ← indexes of the head and the tail of the a better visiting order of regions for each UAV.
region queue of cluster Ci ; When the genetic algorithm is adopted to optimize the
13 j ← index of the unallocated region which is the visiting orders of regions allocated to an examined UAV,
nearest neighbour of Rk or Rl while satisfying the its solution domain contains all possible point-to-point flight
energy constraint; paths of regions, and its fitness function is defined as the total
14 if j == −1 then flight lengths of UAVs. The genetic algorithm proceeds to ini-
15 Remove Ci from the available cluster list; tialize a population of solutions randomly, improve solutions
16 end through repetitive application of mutation, crossover, inversion
17 else and selection operators, and achieve feasible solutions. The
18 Insert region R j into the region queue of cluster following experiment results demonstrate that, although the
Ci ; GA-based order optimization strategy requires more time,
19 Update the head, the tail and time demand of it can greatly improve the accuracy and robustness of path
the cluster Ci ; planning approaches.
20 Remove R j from the available region list;
21 end VI. E XPERIMENTS AND R ESULTS
22 until j = −1; In this section, the performance of the proposed STCA
23 end algorithm is evaluated by conducting simulation experiments
with randomly generated regions. Although the new approach
cannot provide exact solutions to settle the coverage path
planning issue, it does not require to fully search the possible
is a polynomial time algorithm, and normally performs better solution space and can find a reasonable flight path for each
to provide reasonable solutions for the coverage path planning
UAV in a relatively short time. In the following, regions are
problems of UAVs. generated with random parameters first, and then the perfor-
mance of approaches is evaluated in terms of task completion
D. Optimization Strategies time, execution time and deviation ratio.
Apart from the standard clustering strategy of classifying
regions into clusters according to their distances to the heads A. Generation of Regions
or tails of region queues, there exist some other strategies Regions are randomly generated according to the following
that may perform better but have higher time complexi- four parameters:
ties. In this study, we introduce two optimization strategies, • Number of separated regions (m).
named nearest-to-any policy and order optimization strategy, • Number of autonomous UAVs (n).
to improve the accuracy and reliability of solutions. • System utilization (u). u is the total utilizations of regions

(1) Nearest-to-any policy. Nearest-to-any policy is adopted in a system and can be expressed as u = 1≤ j ≤m u j ,
in the region clustering phase, and tries to classify unallocated where u i is defined as the ratio of the scanning area of
regions into clusters according to their shortest distances to any a region to the total area of active ranges of the system,
allocated ones. In this policy, the region which is unallocated, i.e., u j = A j /At ot al .

Authorized licensed use limited to: NORTHWESTERN POLYTECHNICAL UNIVERSITY. Downloaded on May 05,2021 at 15:05:35 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

8 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS

TABLE II
UAV S U SED IN THE S IMULATION E XPERIMENTS

• System drag factor (d). d determines the average

drag factors of regions when they are scanned by
UAVs, and is defined as the average ratio of scan-
ning speeds to the maximum flying speeds of UAVs,
1  Vi, j
i.e., d = mn 1≤i≤n,1≤ j ≤m V max . i
Fig. 2. Task completion times found by the spatial-temporal clustering-based
The generation steps are as follows. First, UUniFast algo- algorithm when the region number increases from 10 to 100.
rithm [22] is used to produce a random area ratio u j for each
region R j , and the scanning area of R j is constructed as A j =
u j At ot al . Then, for each region R j , two random numbers x¯j
and y¯j are selected randomly from [0, ai ], where ai represents
the side length of the flight range. (x¯j , y¯j ) is identified as
the coordinate of the center point of
R j . Distance of regions
R j and Rk is computed by D j,k = (x¯j − x¯k )2 + ( y¯j − y¯k )2 .
Finally, for each region R j , n drag factors di, j (1 ≤ i ≤ n) are
uniformly selected from [d −δ, d +δ], where δ = min(1−d, d)
and δ guarantees that the value of di, j ranges in [0, 1]. The
scanning speed of UAV u i on region R j is constructed as
Vi, j = di, j Vimax .
In this work, values of the above parameters are assigned
from the following sets:
• S E Tm = {10, 20, 30, 40, 50, 60, 70, 80, 90, 100} Fig. 3. Task completion times obtained by our approach when different
numbers of UAVs are used.
• S E Tn = {3, 4, 5, 6, 7, 8, 9}
• S E Tu = {0.005, 0.01, 0.015, 0.02, 0.025, . . .,
0.045, 0.05}
• S E Td = {0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9} optimization strategy, and both the two strategies are used.
In experiments of Fig. 2, the number of UAV is 3, and regions
Meanwhile, the flight range adopted in our experiments is
are generated when the system utilization is 0.01 and the
a square with sides that are 5000 meters in length and UAVs
system drag factor is 0.9.
used to carry out the mission are chosen from Table II.
From Fig. 2 we can find that, task completion times of the
four approaches have a similar change trend that they steadily
B. Task Completion Time Evaluation go up along with the increase of number of regions. The reason
We start by constructing experiments to evaluate the perfor- is that, when more regions are taken into consideration, flight
mance of the proposed approaches in terms of task completion lengths of UAVs increase and more time are required to fully
time, i.e., the time required by UAVs to complete the coverage cover all of regions. When the number of regions is fixed,
tasks. Since approaches adopt different strategies to find the STCA with the GA-based order optimization strategy has a
best fight paths for autonomous UAVs, task completion time better performance, which demonstrates that the GA-based
obtained by each approach is different even regions and UAVs order optimization strategy can play an active role in improv-
are the same. Task completion time is a recognized sign of ing the accuracy of solutions of path planning problems.
performances of path planning approaches. A more accurate An interesting or even surprising phenomenon revealed by
approach would provide a relatively smaller task completion experiments in Fig. 2 is, STCA_NE performs better than
time than others. STCA_NA when the number of regions is more than 40. This
By adjusting the number of regions, Fig. 2 shows the indicates that the basic nearest-to-end policy in the region
task completion time found by the proposed spatial-temporal clustering phase is more effective than the nearest-to-any
clustering-based algorithm (STCA) with different region allo- policy when a great number of regions are covered.
cation strategies. “STCA_NE” represents the basic STCA In experiments of Fig. 3, task completion time of our
algorithm with the nearest-to-end policy, and classifies regions approach is evaluated by adjusting the number of UAVs from
that are nearest to the head or tail into clusters one by 3 to 9. Regions are generated when the system utilization is
one. “STCA_NA”, “STCA_NE_GA” and “STCA_NA_GA” 0.005 and the system drag factor is 0.9. As expected, task com-
respectively represent the modified STCA algorithms, in which pletion times go down when the number of UAVs increases,
only the nearest-to-any strategy, only the GA-based order and approaches with the order optimization strategy perform

Authorized licensed use limited to: NORTHWESTERN POLYTECHNICAL UNIVERSITY. Downloaded on May 05,2021 at 15:05:35 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

CHEN et al.: CLUSTERING-BASED COVERAGE PATH PLANNING METHOD FOR AUTONOMOUS HETEROGENEOUS UAVs 9

Fig. 4. Time costs of the proposed approaches on regions when the system Fig. 5. Time consumption of our approaches when different numbers of
utilization is 0.03 and the system drag factor is 0.2. UAVs are adopted in experiments.

better than those without. In most cases, STCA_NE_GA has

the best performance than others. When three UAVs are
adopted in the experiments, task completion time obtained by
STCA_NE_GA is 4733 seconds, which is 8%, 6% and 3%
less than that of STCA_NA, STCA_NE and STCA_NA_GA,
respectively.

C. Execution Time Evaluation

In this section, performances of the proposed approaches
are evaluated in terms of execution time, i.e., the time cost of
each approach to obtain reasonable flight paths for autonomous
heterogeneous UAVs. Execution time is an important symbol
of measuring the efficiency of approaches. An approach which
Fig. 6. Deviation ratios of approaches on regions generated with different
can provide a feasible solution in a shorter time, is considered system utilizations.
to be more efficient.
In Fig. 4, average execution times of approaches are tested
by increasing the region number from 10 to 100. Regions are relative differences between task completion times obtained
produced when the system utilization is 0.03 and the system by a non-exact approach (STCA proposed in Sect. V here)
drag factor is 0.2. It can be found that the time consumption of and by an exact approach (MILP presented in Sect. IV here),
proposed approaches rises up with increase of the number of i.e., dr = (ST C A − M I L P)/M I L P. Deviation ratio is
regions, and the STCA_NE algorithm has a lower execution an efficient method to check the reliability of approaches.
time than others. This is because that both the GA-based An approach is considered to be more reliable and useful if it
order optimization strategy and the nearest-to-any strategy has a lower deviation ratio.
have a larger computational complexity than the nearest-to- In experiments of Fig. 6, deviation ratios of the proposed
end policy. The largest time consumption of these approaches approaches are evaluated by gradually varying the system
is 0.93 seconds, which demonstrates that our approaches are utilization from 0.005 to 0.05. Regions are generated when
efficient in settling the coverage path planning of multiple the system drag factor is 0.8 and the number of UAVs is 3.
autonomous UAVs. We can find that along with the increase of system utilizations,
Fig. 5 shows the average time costs of our approach with deviation ratios of the four approaches have the same change
different numbers of UAVs. In this experiment, regions are trend that they gradually drop down. The reason is that,
generated with m = 60, u = 0.005 and d = 0.9. As can a larger system utilization results the generation of regions
be seen, the execution times of approaches with the order with larger scanning areas, and more time is required by
optimization strategy grow gradually with the increase of UAVs to scan all of regions of interest. Since scanning time of
number of UAVs. The reason is that, when more UAVs are regions is constant and cannot be optimized, deviation ratios
taken into experiments, more time is spent in adjusting the of approaches become lower. It can be also found that, when
visiting orders of regions and in obtaining optimal flight paths the system utilization is more than 0.005, STCA_NA_GA has
for UAVs. Meanwhile, similar to experimental results given the smallest deviation ratio than others, which is consistent
in Fig. 4, STCA_NE requires the minimum time to achieve with the results shown in Fig. 2.
feasible solutions when the UAV number is fixed. Through varying the system drag factors, deviation ratios
of approaches are evaluated in Fig. 7. Regions are generated
D. Deviation Ratio Evaluation when the utilization is 0.03 and the region number is 25.
Now the performance of approaches is evaluated in terms of As can be seen, along with the increase of system drag factors,
deviation ratio. Deviation ratio dr , is defined as a ratio of the curves would be roughly divided into two main phases. In the

Authorized licensed use limited to: NORTHWESTERN POLYTECHNICAL UNIVERSITY. Downloaded on May 05,2021 at 15:05:35 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

10 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS

[6] Y. Li, H. Chen, M. J. Er, and X. Wang, “Coverage path planning

for UAVs based on enhanced exact cellular decomposition method,”
Mechatronics, vol. 21, no. 5, pp. 876–885, Aug. 2011.
[7] V. Roberge, M. Tarbouchi, and G. Labonte, “Comparison of parallel
genetic algorithm and particle swarm optimization for real-time UAV
path planning,” IEEE Trans. Ind. Informat., vol. 9, no. 1, pp. 132–141,
Feb. 2013.
[8] R. J. Szczerba, P. Galkowski, I. S. Glicktein, and N. Ternullo, “Robust
algorithm for real-time route planning,” IEEE Trans. Aerosp. Electron.
Syst., vol. 36, no. 3, pp. 869–878, Jul. 2000.
[9] C. Zheng, L. Li, F. Xu, F. Sun, and M. Ding, “Evolutionary route
planner for unmanned air vehicles,” IEEE Trans. Robot., vol. 21, no. 4,
pp. 609–620, Aug. 2005.
[10] H. Duan, P. Li, Y. Shi, X. Zhang, and C. Sun, “Interactive learning
environment for bio-inspired optimization algorithms for UAV path
planning,” IEEE Trans. Educ., vol. 58, no. 4, pp. 276–281, Nov. 2015.
Fig. 7. Deviation ratios of the four approaches when regions are generated [11] P. Yang, K. Tang, J. A. Lozano, and X. Cao, “Path planning for single
with different system drag factors. unmanned aerial vehicle by separately evolving waypoints,” IEEE Trans.
Robot., vol. 31, no. 5, pp. 1130–1146, Oct. 2015.
[12] E. Besada-Portas, L. D. L. Torre, J. M. D. L. Cruz, and
B. D. Andrés-Toro, “Evolutionary trajectory planner for multiple UAVs
first phase (i.e., when the system drag factor is less than in realistic scenarios,” IEEE Trans. Robot., vol. 26, no. 4, pp. 619–634,
0.5), approaches with the nearest-to-end policy have smaller Aug. 2010.
[13] I. K. Nikolos, K. P. Valavanis, N. C. Tsourveloudis, and A. N. Kostaras,
deviation ratios than those with the nearest-to-any policy. “Evolutionary algorithm based offline/online path planner for UAV
However, in the second phase (i.e., when the system drag navigation,” IEEE Trans. Syst., Man, Cybern. B. Cybern., vol. 33, no. 6,
factor is equal to or lager than 0.5), results are reversed. pp. 898–912, Dec. 2003.
[14] A. Janchiv, D. Batsaikhan, B. Kim, W. G. Lee, and S.-G. Lee, “Time-
These experiments demonstrate that different region allocation efficient and complete coverage path planning based on flow net-
policies directly influence the reliability of coverage path works for multi-robots,” Int. J. Control, Autom. Syst., vol. 11, no. 2,
planning solutions of autonomous UAVs. Meanwhile, when pp. 369–376, Apr. 2013.
[15] Z. Zhang, J. Li, and J. Wang, “Sequential convex programming for
the system drag factor is 0.3, all of the four curves peak, and nonlinear optimal control problems in UAV path planning,” Aerosp. Sci.
the maximum value of deviation ratios is 14.9%. This relative Technol., vol. 76, pp. 280–290, May 2018.
difference can be accepted when the time cost of approaches [16] B. Ji, Y. Li, D. Cao, C. Li, S. Mumtaz, and D. Wang, “Secrecy
performance analysis of UAV assisted relay transmission for cognitive
is taken into account. network with energy harvesting,” IEEE Trans. Veh. Technol., vol. 69,
no. 7, pp. 7404–7415, Jul. 2020.
VII. C ONCLUSION [17] T. Tang, T. Hong, H. Hong, S. Ji, S. Mumtaz, and M. Cheriet,
“An improved UAV-PHD filter-based trajectory tracking algorithm for
This paper studied the coverage path planning of multi-UAVs in future 5G IoT scenarios,” Electronics, vol. 8, no. 10,
autonomous UAVs, and tried to find a good enough fight p. 1188, Oct. 2019.
path to visit all of regions of interest. UAVs adopted in this [18] J. Chen, C. Du, P. Han, and X. Du, “Work-in-progress: Non-preemptive
scheduling of periodic tasks with data dependency upon heterogeneous
study are heterogeneous, which means their flying speeds, multiprocessor platforms,” in Proc. IEEE Real-Time Syst. Symp. (RTSS),
energy supplies and scanning widths of onboard sensors are Dec. 2019, pp. 540–543.
different. First, based on mixed integer linear programming, [19] C. D. Franco and G. Buttazzo, “Coverage path planning for UAVs
photogrammetry with energy and resolution constraints,” J. Intell. Robot.
an exact formulation was proposed to seek optimal flight paths Syst., vol. 83, nos. 3–4, pp. 1–18, 2016.
for UAVs. Then, a clustering-based algorithm was presented [20] S. A. A. Shahidian and H. Soltanizadeh, “Path planning for two
to classify regions into clusters and to obtain feasible flight unmanned aerial vehicles in passive localization of radio sources,”
Aerosp. Sci. Technol., vol. 58, pp. 189–196, Nov. 2016.
paths with an objective of minimizing the task completion time [21] A. Rodriguez and A. Laio, “Clustering by fast search and find of density
of UAVs. Experimental results demonstrate that the proposed peaks,” Science, vol. 344, no. 6191, pp. 1492–1496, Jun. 2014.
approaches can provide valid flight paths for UAVs such that [22] E. Bini, M. D. Natale, and G. Buttazzo, “Sensitivity analysis for fixed-
priority real-time systems,” Real-Time Syst., vol. 39, nos. 1–3, pp. 5–30,
all regions of interest are covered efficiently. Aug. 2008.

R EFERENCES
[1] J. Keller, D. Thakur, M. Likhachev, J. Gallier, and V. Kumar, “Coordi-
nated path planning for fixed-wing UAS conducting persistent surveil-
lance missions,” IEEE Trans. Autom. Sci. Eng., vol. 14, no. 1, pp. 17–24,
Jan. 2017.
[2] Z. Zhen, D. Xing, and C. Gao, “Cooperative search-attack mission
planning for multi-UAV based on intelligent self-organized algorithm,”
Aerosp. Sci. Technol., vol. 76, pp. 402–411, May 2018. Jinchao Chen (Member, IEEE) received the Ph.D.
[3] P. Yao, H. Wang, and Z. Su, “Cooperative path planning with applica- degree in computer science from Northwestern Poly-
tions to target tracking and obstacle avoidance for multi-UAVs,” Aerosp. technical University, Xi’an, China, in 2016. He is
Sci. Technol., vol. 54, pp. 10–22, Jul. 2016. currently an Assistant Professor with the School of
[4] P. Perazzo, F. B. Sorbelli, M. Conti, G. Dini, and C. M. Pinotti, “Drone Computer Science, Northwestern Polytechnical Uni-
path planning for secure positioning and secure position verification,” versity. He focuses on the multi-processor schedul-
IEEE Trans. Mobile Comput., vol. 16, no. 9, pp. 2478–2493, Sep. 2017. ing, embedded and real-time systems, simulation and
[5] H. Yu, K. Meier, M. Argyle, and R. W. Beard, “Cooperative path verification, decision-making, and intelligent con-
planning for target tracking in urban environments using unmanned air trol of unmanned aerial vehicles. He is a member
and ground vehicles,” IEEE/ASME Trans. Mechatronics, vol. 20, no. 2, of CCF.
pp. 541–552, Apr. 2015.

Authorized licensed use limited to: NORTHWESTERN POLYTECHNICAL UNIVERSITY. Downloaded on May 05,2021 at 15:05:35 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

CHEN et al.: CLUSTERING-BASED COVERAGE PATH PLANNING METHOD FOR AUTONOMOUS HETEROGENEOUS UAVs 11

Chenglie Du (Member, IEEE) received the Ph.D. Pengcheng Han received the master’s degree from
degree in computer science from Northwestern Poly- the School of Computer Science, Northwestern Poly-
technical University, China, in 1999. He is cur- technical University, China, in 2015, where he is
rently a Professor with the School of Computer currently pursuing the Ph.D. degree. He is also a
Science, Northwestern Polytechnical University. His Research Assistant with the School of Computer
research interests are real-time distributed computing Science, Northwestern Polytechnical University. His
systems, cyber-physical systems, intelligent trans- research interests are software engineering, real-time
portation, modeling, simulation, and verification of systems, simulation and verification of distributed
distributed systems. He is a member of CCF. systems, energy optimization and control, intelligent
control of unmanned aerial vehicles, and workflow
scheduling.

Ying Zhang received the Ph.D. degree in com- Wei Wei (Senior Member, IEEE) received the Ph.D.
puter science and technology from the College degree from Xi’an Jiaotong University, in 2010.
of Computer Science and Electronic Engineering, He is currently an Associate Professor with the
Hunan University, in 2020. He is currently an Asso- School of Computer Science and Engineering, Xi’an
ciate Professor with the School of Computer Sci- University of Technology, China. His research inter-
ence, Northwestern Polytechnical University, China. ests include wireless networks, sensor networks,
His research interests include parameter estimation, image processing, mobile computing, distributed
energy optimization and control, autonomous driving computing, pervasive computing, the Internet of
of ground vehicles, decision-making, and intelli- Things, sensor data, and cloud computing. He is a
gent control of unmanned aerial vehicles and other Senior Member of IEEE and CCF.
autonomous systems.

Authorized licensed use limited to: NORTHWESTERN POLYTECHNICAL UNIVERSITY. Downloaded on May 05,2021 at 15:05:35 UTC from IEEE Xplore. Restrictions apply.