Download as pdf or txt
Download as pdf or txt
You are on page 1of 11

Computers and Operations Research 105 (2019) 1–11

Contents lists available at ScienceDirect

Computers and Operations Research


journal homepage: www.elsevier.com/locate/cor

A vehicle routing problem arising in unmanned aerial monitoring


Lu Zhen a,∗, Miao Li a, Gilbert Laporte b, Wencheng Wang a
a
School of Management, Shanghai University, Shanghai, China
b
Department of Decision Sciences, HEC Montréal, Montréal, Québec, Canada

a r t i c l e i n f o a b s t r a c t

Article history: Unmanned aerial vehicles (UAVs) are widely used to perform monitoring tasks both in the military and
Received 16 March 2018 civilian areas, and the planning of their routes is critical. This study investigates a routing problem in
Revised 15 October 2018
which UAVs monitor a set of areas with different accuracy requirements. This problem is a variant of the
Accepted 2 January 2019
classical vehicle routing problem (VRP), where one must determine not only the order in which to visit
Available online 8 January 2019
a set of nodes located in the plane, but also the height at which to visit them, which impacts the accu-
Keywords: racy level and the service time. An integer programming model is formulated to optimize flight routes
Unmanned aerial vehicle and minimize the total time needed to complete the monitoring tasks. A tabu search metaheuristic is de-
Route planning veloped for the problem. Extensive numerical experiments are conducted to assess the efficiency of the
Vehicle routing problem heuristic.
Aerial monitoring
© 2019 Elsevier Ltd. All rights reserved.

1. Introduction model is suitable for scenarios in which the monitoring require-


ments are known in advance. In contrast to the well-known vehicle
Unmanned aerial vehicles (UAVs), commonly known as drones, routing problem (VRP) which only determines the order in which
constitute a new type of monitoring equipment. By loading dif- to visit a set of nodes, here we also have to determine the height at
ferent imaging sensors, UAVs can capture target images, and the which each node is visited by a UAV. As we will show, commercial
monitored images can be transmitted in real time to a control integer linear programming solvers such as CPLEX can obtain opti-
station via a wireless transmission system. Recent developments mal solutions within reasonable time only for small-scale instances
in aviation, microelectronics, computers, navigation, communica- of the problem. We have therefore developed a tailored tabu search
tions, sensors and other related technologies have yielded contin- metaheuristic to solve large-scale instances within relatively short
uous improvements in the performance of UAVs. These vehicles computing times.
can be used to monitor the water environment, traffic flows and The remainder of this paper is organized as follows.
oil pipelines (Alotaibi et al., 2018). The use of UAVs as a mean of Section 2 reviews the literature on UAV routing. Section 3 formally
monitoring offers numerous advantages, such as the capacity to defines the problem. A mathematical model is then formulated
observe wide areas (covering 30 0–50 0 m2 at a height of 150– in Section 4. Section 5 describes our tabu search metaheuristic.
300 m), low cost, excellent flexibility, high efficiency and real-time Section 6 presents the parameter settings and the results of
operations (Harwin and Lucieer, 2012). As a result, UAVs are exten- extensive numerical experiments. Conclusions are outlined in the
sively employed in meteorological exploration, disaster assessment, last section.
geological surveys, environmental monitoring, resources prospect-
ing and other fields. The monitoring accuracy grades requirements
vary from one area to another, and the time needed to perform a 2. Related literature on UAV routing
monitoring task depends on the height of the UAV. Given a fleet
of UAVs with limited maximum flight time and several monitor- There are two categories of relevant research streams: the
ing areas with different required accuracy grades, it is critical to VRP and UAV routing. The VRP literature has already been exten-
optimize the routing and heights of UAVs. sively surveyed in recent years. For a comprehensive coverage, see
In this paper, we formulate an integer linear programming Toth and Vigo (2014). In this section, we mainly review the litera-
model for the UAV routing and height selection problem. This ture on UAV routing.
Because UAVs are widely used in both military and civil ap-
plications, and UAV-related studies have become a popular re-

Corresponding author. search topic in recent years. Several academic papers have been
E-mail address: lzhen@shu.edu.cn (L. Zhen). written about UAV mission path planning problems in military

https://doi.org/10.1016/j.cor.2019.01.001
0305-0548/© 2019 Elsevier Ltd. All rights reserved.
2 L. Zhen, M. Li and G. Laporte et al. / Computers and Operations Research 105 (2019) 1–11

environments. Specifically, these problems aim at planning UAV industry, where UAVs are used to collect information about pos-
flight paths to complete their tasks and avoid enemy threats sible emergencies. The main goal of their model was to create
such as radar detection, obstacles or collisions with other UAVs an optimal operational schedule for UAVs in order to satisfy the
(Alotaibi et al., 2018; Beard et al., 2002; Casbeer and Holsap- monitoring needs in each of several time periods by considering
ple, 2011; Xia et al., 2017; Zhang et al., 2016). In civil applica- fixed and operating costs. Chow (2016) proposed a real-time data-
tions, UAV routing problems arise in the planning of monitoring driven method to systematically deploy UAVs during multiple pe-
and delivery tasks, mostly in the context of physical infrastructure, riods. They developed a deterministic arc-inventory routing model
agriculture, transport, and entertainment and media applications and derived a stochastic dynamic policy for their problem which
(Otto et al., 2018). was formulated as a selective VRP. This methodology was applica-
The planning of area coverage and UAV routing is a major ble to scenarios with strong uncertainty and high costs, such as
optimization problem in the field of UAV monitoring. Classical city monitoring during major events. To solve the problem, they
UAV coverage path planning problems refer to finding UAV paths proposed an approximate dynamic programming algorithm based
so that all nodes of an area can be covered (Otto et al., 2018). on least squares Monte Carlo simulation. Coelho et al. (2017) in-
The objective is to minimize the distance-related cost (Oh et al., troduced a fleet of heterogeneous UAVs routing problem to deal
2014; Torres et al., 2016; Xu et al., 2014), energy consumption with the vehicles’ limited autonomy by considering multiple charg-
(Di Puglia Pugliese et al., 2016; Li et al., 2016; Zorbas et al., 2016), ing stations and respecting operational requirements. They used a
or completion time (Avellar et al., 2015; Barrientos et al., 2011; mixed integer linear programming model to minimize seven dif-
Nedjati et al., 2016). Some recent work exists in this field which ferent objective functions inspired by a multi-criteria view of a
formulates this problem in a three-dimensional (3D) space net- real system. Their model was solved by a metaheuristic algorithm
work. Thus Mozaffari et al. (2016) studied the optimal deployment that filters the non-dominated solutions from the pool of solutions
of multiple UAVs equipped with directional antennas used as wire- found in a branch-and-bound optimization tree, using a black-box
less base stations. They proposed a solution approach based on dynamic search algorithm.
circle packing theory to determine the 3D locations of UAVs with The contribution of our research is threefold. First, we solve a
the aim of maximizing the coverage area and the coverage lifetime VRP arising in unmanned aerial monitoring in a three-dimensional
of UAVs. They also explored the relationship between the altitude space rather than in a two-dimensional plane, where each flight
and locations of UAVs, and between the number of UAVs and the height corresponds to a monitoring accuracy level: the lower the
beamwidth of the directional antennas. Zorbas et al. (2016) pro- flight, the higher the monitoring accuracy. To the best of our
posed the static and dynamic UAV location problems considering knowledge, this paper is the first work that explicitly defines a
3D deployment of the UAVs with the aim of minimizing the num- VRP by considering the visiting heights of vehicles in the context
ber of UAVs and the total energy consumption, which is related to of UAVs monitoring. Second, the UAV routing and height selection
the UAV positions. They formulated this problem by defining an problem is mathematically formulated. We provide a model based
integer linear and a mixed integer non-linear optimization models, on that of the VRP to minimize the total time spent by the fleet,
and they designed centralized and localized heuristics to solve the while guaranteeing the monitoring accuracy of all areas and satis-
problems. Caillouet and Razafindralambo (2017) presented a 3D fying the maximum flight time allowed to perform tasks for each
model for the problem of covering ground targets, in which UAVs UAV. Third, we design a tabu search metaheuristic that includes a
monitor the targets and collect the information from the ground. tailored heuristic for the generation of initial solutions. Our meta-
Their goal is to cover all the targets while minimizing deployment heuristic yields efficient solutions to the problem.
cost and the UAVs altitude. They computed a Pareto front to ob-
tain non-dominated solutions to their bi-objective model whithin
reasonable computing times. 3. Formal problem description
If the area is discretized, the UAV coverage path planning prob-
lem becomes a UAV routing problem, which can be modeled as The routing problem on which the UAV monitoring problem is
a basic routing problem, such as the traveling salesman problem based is defined on a directed network G = (N, D ), where N is a
(TSP) (Murray and Chu, 2015; Phung et al., 2017), the multiple TSP set of nodes in the plane and 0 ∈ N is a depot. We define a set
(Manyam et al., 2017; Sundar and Rathinam, 2017; Vilar and Shin, of heights H, where 0 ∈ H corresponds to the ground height at the
2013), or the VRP (Cho et al., 2015; Chow, 2016; Coelho et al., depot. The arc set is D = {((i, r ), ( j, s ))|i, j ∈ N; r, s ∈ H }. Moving a
2017; Di Puglia Pugliese et al., 2017; Guerriero et al., 2014; Kim UAV from (i, r) to (j, s) incurs a time consumption uir + tir js , where
et al., 2013). Among these papers, VPR-based UAV routing prob- uir is the observation time at (i, r) and tirjs is the travel time be-
lems are particularly relevant to our study. Kim et al. (2013) de- tween the two locations. We are given a set K of UAVs, each with
veloped a mixed integer linear program (MILP) to model the prob- a limited maximum flight time Tmax , and a set A of monitoring
lem of scheduling UAVs with multiple shared bases in several lo- areas with different accuracy grades. A binary coefficient bair in-
cations, considering that the duration of UAV missions is limited dicates whether monitoring area a ∈ A can be covered by a UAV
by battery or fuel capacity. They proposed a genetic algorithm stopping at node i ∈ N at height r ∈ H. The problem is to select an
to obtain feasible solutions when solvers such as CPLEX cannot. optimal flight route and heights for the UAVs in order to minimize
Guerriero et al. (2014) proposed a distributed dynamic schedul- the total monitoring time. Here we define two groups of variables:
ing system to film a football match by autonomous UAVs and for- the first group determines the order of the space nodes to be tra-
mulated the problem as a multi-criteria optimization model based versed by the UAVs, and the second group makes decision on the
on a VRP with soft time windows (VRPTW), in which the objec- height at which each UAV should stay for a given time period in
tive are the total distances traveled by the UAVs, customer satis- order to perform its monitoring task.
faction and the number of UAVs used in the solution. Similarly, Fig. 1 depicts an example of this problem. In this figure, two
(Di Puglia Pugliese et al., 2017) considered a problem of film- UAVs are sent to monitor a number of areas with different accu-
ing sport events with UAVs. Unlike their 2014 paper, the aim of racy grades. The monitoring areas covered by some smaller mon-
their bi-objective model is to ensure a maximum viewer satisfac- itoring scopes (the darker circles) have a higher accuracy require-
tion and to minimize the distance traveled by the UAVs under a ment, which forces the UAV to remain at a lower altitude for mon-
static scenario. Cho et al. (2015) designed a VRP-based mathemat- itoring. In contrast, if a UAV uses a higher altitude, it will cover a
ical model for UAV-aided security operations in the oil and gas wider area (the lighter circle).
L. Zhen, M. Li and G. Laporte et al. / Computers and Operations Research 105 (2019) 1–11 3

Fig. 1. Illustration of two UAVs monitoring routes.

• All UAVs are isomorphic and every UAV can execute any moni-
toring task.
• The UAV velocity is constant and UAVs encounter the same air
resistance at different altitudes.
• Each UAV tour must start and end at the same depot.

4. Model formulation

In this section, we formulate an integer linear programming


model for the UAV monitoring problem. The model determines the
UAV flight routes as well as the monitoring heights. Each moni-
toring area has a specific monitoring requirement level, which is
known in advance and should be satisfied by the UAVs’ monitoring
Fig. 2. A simplified planar case of UAVs monitoring.
activities. Since the UAVs are battery-powered, limited battery ca-
pacity affects their flight time and monitoring time. Therefore the
We discretize large monitoring areas into several smaller objective of the model is to minimize the total time required for
squares, as shown in Fig. 2. Each square represents a unit monitor- completing all monitoring tasks.
ing area. In the figure, a UAV stays at a higher altitude to monitor
areas 1, 4 and 5 simultaneously, and with the same accuracy re-
quirements. However, monitoring area 2 requires a higher accuracy 4.1. Notations
than areas 1, 4 and 5. Although area 2 is covered by a monitoring
scope, which also covers the areas 1, 4 and 5, its accuracy require- We introduce the following notations for our mathematical
ment is not satisfied. Monitoring area 2 should be covered by an- model.
other monitoring scope with a higher accuracy requirement, i.e., a
smaller circle with dark color covering area 2. It should be noted Indices and sets
that a monitoring area can be covered by more than one moni- i, j index of a node in the plane; 0 denotes the depot.
toring scope it there exists at least one monitoring scope with an N set of all nodes in the plane.
accuracy at least equal to that required by the monitoring area. r, s index of a flight height; 0 denotes the ground height at
Before modeling the UAV monitoring problem, we list a number the depot.
of assumptions: H set of all possible flight heights.
a index of a monitoring area.
• The planar projections of the nodes in the space either are lo- A set of all monitoring areas.
cated at the center of discrete squares or at the vertices of the k index of a UAV.
K set of all UAVs.
squares, as shown in Fig. 2.
4 L. Zhen, M. Li and G. Laporte et al. / Computers and Operations Research 105 (2019) 1–11

Parameters
tir js time needed by a UAV to travel from node (i, r ) to ( j, s );
i, j ∈ N, r, s ∈ H. xir jsk ∈ {0, 1} i, j ∈ N, r, s ∈ H, k ∈ K
uir time needed by a UAV staying at node (i, r ) to monitor (10)
reachable areas; i ∈ N\{0}, r ∈ H \{0}.
bair binary parameters; equal to one if and only if monitoring
area a can be covered by the monitoring scope of a UAV
that stops at node (i, r ) and the monitoring area’s yirk ∈ {0, 1} i ∈ N\{0}, r ∈ H \{0}, k ∈ K.
monitoring accuracy requirement is also satisfied; a ∈ A, (11)
i ∈ N\{0}, r ∈ H \{0}.
Tmax maximum flight time of each UAV. The objective function (1) minimizes the total time required to
perform all tasks. It contains two terms: the total traversal time
Decision variables and the total monitoring time. Constraint (2) guarantee that each
binary variables; equal to one if and only if UAV k travels monitoring area is covered by at least one monitoring scope of a
xir jsk along the link from (i, r ) to ( j, s ); i, j ∈ N, r, s ∈ H, k ∈ K. UAV while satisfying the monitoring accuracy requirements. Con-
yirk binary variables; equal to one if and only if UAV k stops straint (3) ensure that there will not be more than one UAV stop
at node (i, r ) to monitor; i ∈ N\{0}, r ∈ H \{0}, k ∈ K. at the same node. Constraint (4) link the decision variables xirjsk
and yirk . Constraints (5) and (6) indicate that all UAV tours start
4.2. Mathematical model and end at the depot. Constraint (7) ensure flow conservation. Con-
straint (8) are subtour elimination constraints. Constraint (9) spec-
ify that the traversal time consumption and the monitoring time
   consumption of each UAV cannot exceed its maximum flight time.
Minimize Z = tir js xir jsk + uir yirk Constraints (10) and (11) define the domains of the variables.
i, j∈N r,s∈H k∈K i∈N r∈H k∈K Constraint (8) are commonly used to eliminate subtours in VRP
(1) models, but they are not suitable for direct use in CPLEX since it is
impractical to enumerate all subsets S of N × H with |S| ≥ 2. Here
subject to
we design an alternative way to eliminate subtours by defining
  
bair yirk ≥ 1 a∈A some additional variables and constraints.
i∈N\{0} r∈H \{0} k∈K Added variables
(2) μirk number of space nodes that have been visited by UAV k;
i ∈ N, r ∈ H, k ∈ K. μ00k = 0, and μirk is not defined if
space node (i, r ) is not visited by UAV k.

yirk ≤ 1 i ∈ N\{0}, r ∈ H \{0} Added constraints
k∈K
(3) μ00k = 0 k∈K
(12)

xir jsk = yirk i ∈ N, r ∈ H, k ∈ K; (i, r ) = ( j, s )
j∈N s∈H μ jsk ≥ μirk + 1 − |N| × |H |(1 − xir jsk )
(4) i ∈ N, r ∈ H, j ∈ N\{0}, s ∈ H \{0}, k ∈ K; (i, r ) = ( j, s )
(13)

x00 jsk = 1 k∈K
j∈N s∈H

(5) 0 ≤ μirk ≤ |N | × |H | − 1 i ∈ N, r ∈ H, k ∈ K.
(14)
 As in Miller et al. (1960), the above constraints eliminate
xir00k = 1 k∈K
subtours. Although more variables and constraints are defined,
i∈N r∈H
the number of constraints and the model complexity are signifi-
(6)
cantly reduced compared with the number of Constraint (8). Con-
straint (8) will therefore be replaced with Constraints (12)–(14) in
  our implementation.
xir jsk = x jsirk i ∈ N, r ∈ H, k ∈ K; (i, r ) = ( j, s )
j∈N s∈H j∈N s∈H
5. Algorithmic strategy
(7)
The model presented in Section 4 can be solved by means
 of a commercial optimization software such as CPLEX for small-
xir jsk ≤ |S| − 1 k ∈ K; S ⊆ N × H, |S| ≥ 2
scale instances. However, for medium- and large-scale instances,
i∈N r∈H j∈N s∈H
the model becomes intractable. We have therefore developed a
(8) tailored tabu search metaheuristic in order to solve the proposed
model which is based on that of the classical VRP. Tabu search is
    a heuristic solution approach developed by Glover (1986), which
tir js xir jsk + uir yirk ≤ Tmax k∈K extends some traditional local search algorithms and has been suc-
i, j∈N r,s∈H i∈N\{0} r∈H \{0} cessfully applied to a wide variety of routing problems (Chen et al.,
(9) 2006; Lai et al., 2016; Wen et al., 2009; Xu et al., 2017).
L. Zhen, M. Li and G. Laporte et al. / Computers and Operations Research 105 (2019) 1–11 5

Algorithm 1 Algorithm framework of tabu search metaheuristic. • Initial solution


1: Generate an initial solution as the current solution yn
according We present in Algorithm 2 a heuristic to generate an initial
to Algorithm 2 and obtain its objective value by using CPLEX to solution corresponding to the core variables yirk . In Algorithm 2,
solve the mathematical model. Au ⊆A is the set of all areas that have not been monitored.
2: Set the iteration number n = 1, the number of iterations u = 0
for which the best known solution has remained unchanged, Algorithm 2 Procedure for generating an initial solution of yirk .
the tabu list T L = ∅, and the tabu length l.
1: Generate the index of UAV k randomly chosen in K for each
3: while the iteration number n has not reached a preset maxi-
areas a ∈ A.
mum value iter_max or the number of iterations u for which
2: For each area a ∈ A and the UAV k ∈ K, record the list of space
the best known solution has been unchanged has not reached
nodes (i, r ), i ∈ N\{0}, r ∈ H \{0} where the UAV k can stay for
a preset maximum value unc_max do
monitoring area a with the area’s monitoring accuracy require-
4: Generate neighborhoods of the current solution yn and ob-
ment satisfied.
tain their objective values.
3: if the areas a ∈ A only can be monitored by the UAV k ∈ K stop-
5: Sort neighborhood solutions according to their objective
ping at one space node (i, r ), i ∈ N\{0}, r ∈ H \{0} then
value and identify the one with the minimum objective value.
4: Update the status indicating that these areas a have be
6: if elements of a neighborhood solution are not in the tabu
monitored and set yirk = 1.
list then
5: end if
7: Replace the current solution yn with this neighborhood
6: while there are areas that have not been monitored do
solution.
7: Calculate the number of areas a ∈ Au that can be monitored
8: else if elements of a neighborhood solution are in the tabu
by the UAVs k ∈ K stopping over at each space node (i, r ), i ∈
list but this neighborhood solution satisfies the aspiration cri-
N\{0}, r ∈ H \{0}.
terion then
8: if there is a space node (i, r ), i ∈ N\{0}, r ∈ H \{0} where
9: Replace the current solution yn with this neighborhood
UAV k ∈ K can stay and monitor the most areas a ∈ Au then
solution.
9: Find all the areas a corresponding to this space node
10: else Impose the release criterion to the neighborhood solu-
(i, r ) and the UAVs k that monitored these areas a, update the
tion.
status indicating that these areas have be monitored and set
11: end if
yirk = 1.
12: Update the tabu list T L.
10: end if
13: if the new current solution yn is better than the best known
11: end while
solution y∗ then 
12: if k∈K yirk > 1, i ∈ N\{0}, r ∈ H \{0} then
14: y∗ ← yn , and set u = 0.
13: Consider all cases for which yirk = 1 and set one such vari-
15: else set u = u + 1. 
able to one in order to make k∈K yirk ≤ 1.
16: end if
14: end if
17: if u > unc_max then
15: Update the last |A| elements of the sequence list SL.
18: Regenerate the current solution yn according to diversifi-
cation strategy and set u = 0.
19: else set u = u + 1.
20: end if • Neighborhood search strategy and tabu list
21: set n = n + 1.
If a monitoring area is not mapped with one space node, for the
22: end while
first |A| elements in the sequence list SL, we randomly choose one
23: Return y∗ .
of the space nodes that have not been chosen by the current solu-
tion. This choice is made from the set of space nodes that can sat-
isfy the accuracy requirement of the monitoring area to perform an
We first describe in Algorithm 1 the tabu search algorithm exchange with the current space node associated with it. The first
framework. |A| elements of the sequence list SL are then updated. One of the
We explain the algorithmic elements next. last |A| elements recording the index of UAVs, is randomly selected
to plus or minus one in order to change the UAV that monitors
• Solution representation and evaluation
an area. This number is plus one if the index is the smallest, and
The core variables yirk determining whether UAV k stops at it is minus one if the index is the largest. In addition, we need to
space node (i, r) to monitor are coded as follows. For the UAVs check the mapping among UAVs, monitoring areas and space nodes
routing problem with |A| monitoring areas, we use a sequence list according to Constraint (3) of the model and we update the last |A|
SL with 2|A| elements; the first |A| elements record the index of elements of the sequence list SL.
space nodes, and the last |A| elements record the index of UAVs. Tabu elements are set according to the neighborhood search
Then the variables yirk are known parameters and we solve the procedure. Elements of a solution induced by a move are assigned
mathematical model using CPLEX to obtain the value of variables a tabu status. A monitoring area a just chosen cannot be chosen
xirjsk (the traversal order of UAV) and its objective value. again while it is tabu. The two swapped space nodes (i, r) and (j,
The solution constrained by the yirk variables may be infeasible. s) that can satisfy the accuracy requirement of the same monitor-
As we mentioned previously, these variables determine at which ing area cannot be swapped again while they are tabu.
space nodes the UAV stops. It may happen that a UAV may stop at • Diversification strategy
too many space nodes and the solution will be infeasible because it
will violate the maximum flight time. This solution will not satisfy The current solution is regenerated to help avoid local optima
Constraint (9) and the algorithm will return a large positive num- if the best known objective value has not improved after a certain
ber as a penalty. Note that this positive number must be greater number of iterations. Several feasible solutions are generated ran-
than the value of the objective function associated with any feasi- domly, and the solution with the best objective value is selected
ble solution. as the current solution. The method applied to generate feasible
6 L. Zhen, M. Li and G. Laporte et al. / Computers and Operations Research 105 (2019) 1–11

Fig. 3. A case of generating an initial solution and neighborhood exchange.

solutions is the same as that used for the generation of the initial accuracies. The low flight height (50 m), the medium flight height
solution, except for the steps in the do while loop, which are as (100 m) and the high flight height (300 m) mean that the monitor-
now follows: for an area a ∈ Au , randomly choose a space node (i, ing scope of UAVs cover exactly one monitoring area (50 m2 ), four
r), i ∈ N\{0}, r ∈ H\{0} that can be monitored by UAV k. Then, find monitoring areas (200 m2 ) and nine monitoring areas (450 m2 ),
all the areas a corresponding to this space node and the UAVs that respectively, as shown in Fig. 4. On small-scale instances, the flight
can monitor them; update the status indicating that these areas height of UAVs is set to low and medium levels; and on medium-
have be monitored and set yirk = 1. and large-scale instances, it is set to low, medium and high lev-
els, that is the set H of flight heights is {0, 1, 2, 3}. For the small-
• Aspiration criterion
scale instances involving six areas and eight areas, if the accuracy
If a neighborhood solution is better than the best known solu- demand of an area is the low accuracy level, the UAV will cover
tion, it is accepted as the initial solution for the next iteration and all areas at the same time, which is not meaningful. Therefore, we
replaces the best known solution. consider only low and medium levels for the small-scale instances.
We define several notations in Table 1 to explain how the two
• Release criterion parameters bair and tirjs are generated. For the setting of param-
Elements of several solutions in the tabu list can be unblocked eters bair , the accuracy requirements for each monitoring area a
with a certain probability. are first generated randomly. The center node i ∈ Nc of each mon-
Fig. 3 provides a simple example based on the scale of nine ar- itoring area a corresponding to vertices j ∈ Nv is then determined.
eas, 25 nodes and two UAVs to explain the coding of variables yirk Finally we select the maximum distance dij between nodes i ∈ N
and the neighborhood search strategy. The detailed locations of the and vertices j ∈ Nv of each monitoring area a, and compare it to
areas and of the nodes can be seen in Fig. 7(a). In Fig. 3, the space radii r1 , r2 , r3 of monitoring scopes with different monitoring lev-
node corresponding to area 1 is (4,1). This means that a UAV can els. If the accuracy requirement of the monitoring area a is high,
stay at space node (4,1) to monitor area 1, which can satisfy area i.e., max {dij } < r1 , then this area to which vertices j ∈ Nv belong can
1’s accuracy requirement. The relationship between the other areas be covered by the monitoring scope of the UAV that stops at space
and their corresponding space nodes is the same as above. nodes (i, 1), and the parameter bai1 is set to 1. Similarly, we gener-
ate the values of the parameter bair under the other two monitor-
6. Numerical experiments ing levels.
To set the parameters tirjs , we use the space distance between
In order to assess the effectiveness of the solution method, we two space nodes dirjs to measure the time consumption of UAVs
have performed extensive numerical experiments on small- (from traveling from space node (i, r) and node (j, s). It is possible
six areas to nine areas), medium- (from 12 areas to 20 areas), and that the displacements between space nodes (i1 , r1 ) and (j1 , s1 )
large-scale (from 36 areas to 81 areas) instances. We have used the and space nodes (i2 , r2 ) and (j2 , s2 ) are equal, but the space an-
IBM ILOG CPLEX Optimization Studio 12.6.1 (Visual Studio 2015, gles formed by these two pairs of space nodes may be differ-
C#) on a DELL Precision 7600 workstation with two Xeon E5-2643 ent, which means that the UAVs traveling from (i1 , r1 ) to (j1 , s1 )
V3 CPUs (24 cores) of 3.4 GHz processing speed and 128 GB of and from (i2 , r2 ) to (j2 , s2 ) could spend unequal times. Therefore,
memory. the different flight paths of UAVs between any two space nodes
should be considered as well. There are five flying modes: upward
6.1. Parameter settings flight, obliquely upward flight, downward flight, obliquely down-
ward flight, and level flight. The lift forces of UAVs during a flight
For the flight heights, we distinguish between low, medium and with the above five modes are illustrated in Fig. 5. Based on the
high levels, which correspond to high, medium and low monitoring
L. Zhen, M. Li and G. Laporte et al. / Computers and Operations Research 105 (2019) 1–11 7

Fig. 4. Illustration of the relationship between monitoring scopes and monitoring areas.

Table 1
Definition of the auxiliary parameters.

Notation Definition

ma accuracy requirements of the monitoring areas


Nc set of all center nodes, Nc ∈ N
Nv set of all vertices, Nv ∈ N
la side length of a square monitoring area √ √ √
r1 , r2 , r3 radii of monitoring scopes that correspond to high, medium and low monitoring levels; r1 = 2/2la , r2 = 2la , r3 = 3 2/2la
dij planar distance from node i to node j
hrs height difference between space nodes (i, r) and (j, s)
dirjs distance from space nodes (i, r) to (j, s), dir js = (di j )2 + (hrs )2
G gravity met by a UAV during flight
f air resistance met by a UAV during flight; f < G
F lift force met by a UAV during flight

Fig. 5. Force analysis of UAVs during flight.

above analysis, the lift force of UAVs during flight can be calcu- • Case 4: obliquely downward flight
lated as follows:
• Case 1: upward flight

hrs 2 h
F = [G − f sin(arctan )] + [ f cos(arctan rs )]2
di j di j
i, j ∈ N, r, s ∈ H (18)
F =G+ f
(15) • Case 5: level flight

• Case 2: obliquely upward flight



F= G2 + f 2 .

hrs 2 h (19)
F = [G + f sin(arctan )] + [ f cos(arctan rs )]2
di j di j The greater is the required lift force of the UAV traversing from
i, j ∈ N, r, s ∈ H (16) space nodes (i, r) and (j, s), the longer is the time it spends (Fig. 6).
Thus, the required lift force of UAVs from space nodes (i, r) and (j,
• Case 3: downward flight s) is considered as the weight of the flight routing selection and
is used to measure the time consumption between any two space
nodes with the flight distance dirjs .
F =G− f In addition, in the tabu search procedure, the maximum it-
eration number iter_max is between 30 and 50, the maximum
(17)
number for which the best known solution has been unchanged
unc_max is 5, and the tabu length l is between 10 and 30.
8 L. Zhen, M. Li and G. Laporte et al. / Computers and Operations Research 105 (2019) 1–11

is computed as the convex combination p × T_ub + (1 − p) × T _lb,


with p ∈ {0.3, 0.5, 0.7}; p is the parameter of the convex combina-
tion. Intuitively, different maximum flight times will influence the
objective value and the computation time. Specifically, too tight
maximum flight time can mean that a feasible solution will not be
obtained within a reasonable time, for example, in instance ‘36-
85-4-2-1’, we cannot obtain a feasible solution within two hours
by using CPLEX or tabu search. In contrast, if the maximum flight
time is too loose, it will not be sufficiently constrainting, for exam-
ple, in instance ‘9-25-3-2-2’, ‘9-25-3-2-3’ and ‘16-41-4-2-3’, where
two UAVs are not fully used. Overall, the total route duration, the
flight route and the computation time are sensitive to the maxi-
mum flight time of UAVs. Therefore, we set the same maximum
flight time for each group of experiments and the same upper
bound of maximum flight time for all experiments of Section 6.3.

Fig. 6. Illustration of hrs , dij and dirjs . 6.3. Performance of the proposed algorithmic strategy

The following numerical experiments on small- medium- large-


6.2. Sensitivity analysis of maximum flight time scale instances were performed by applying both CPLEX and tabu
search. The solution obtained by CPLEX is used for the comparison.
In reality, the monitoring tasks performed by the UAVs have dif- The results are shown inTables 3–5, where the objective values ‘Z’,
ferent accuracy requirements. Accordingly, the UAVs need different the computation time ‘T’, the gap of objective values ‘Obj gap’ and
times to complete these tasks. Different maximum flight times for the computation time ratio ‘Time ratio’ are provided. The instance
the UAVs may lead to different flight routes and hence to different ID has the same meaning as in Section 6.2.
durations. The way in which the route duration is influenced by From Tables 3–5, we can observe that CPLEX can only solve
the maximum flight time of UAVs is worth discussing. We have small-scale instances to optimality. When the number of UAVs in
conducted a sensitivity analysis on the maximum flight time to the instances changes from one to two, the computation time in-
test its influence on the objective value and on the computation creases sharply with an average of over one hour, and the compu-
needed to solve the model of Section 4, based on the instance in tation time of the instance ‘9-25-3-2-2’ is more than three hours.
Fig. 7. For small-scale instances, the tabu search heuristic can also ob-
In Table 2, the instance ID numbers (e.g., ‘9-24-3-2-1’) denote tain the optimal solutions, but it only takes an average of 19.82 s
the number of monitoring areas ‘9’, the number of nodes ‘24’, the of computation time. For 15 medium-scale instances and the first
number of flight heights ‘3’, the number of UAVs ‘2’, and the in- two groups of large-scale instances, the feasible solutions obtained
dex of the case ‘1’, respectively. We report the maximum flight by the proposed solution method are at least as good as the best
time T_max , the lower bound T_lb of T_max , the upper bound ones generated by CPLEX in over half of the cases. For some of
T_ub of T_max , the objective values ‘Z’, the computation time ‘T’ these instances, even if the feasible solutions by CPLEX are better
and the number of UAVs enabled ‘No. UAVs enabled’. Here, T_max than those obtained by the proposed method, the average objective

Fig. 7. The instance of Section 6.2.


L. Zhen, M. Li and G. Laporte et al. / Computers and Operations Research 105 (2019) 1–11 9

Table 2
Sensitivity analysis of maximum flight time.

Instance ID p T_lb T_ub T _ max CPLEX Tabu search No. UAVs enabled

Zc Tc (s) Zt Tt

9-25-3-2-1 0.3 2370 3307.14 6919.72 3307.14 61.84 2


9-25-3-2-2 0.5 300 7200 3750 2953.52 13.86 2953.52 10.45 1
9-25-3-2-3 0.7 5130 2953.52 11.14 2953.52 10.26 1
16-41-4-2-1 0.3 30 0 0 4942.61 720 0.0 0 5018.29 1875.04 2
16-41-4-2-2 0.5 1200 7200 4200 4605.32 2392.12 4637.59 967.65 2
16-41-4-2-3 0.7 5400 4438.46 51.40 4438.46 2.36 1
36-85-4-2-1 0.3 3840 N.A. – N.A. – –
36-85-4-2-2 0.5 2400 7200 4800 6999.18 720 0.0 0 7143.52 4224.52 2
36-85-4-2-3 0.7 5760 7236.90 720 0.0 0 7143.52 3805.76 2

Notes: (i) T_max = p × T _ub + (1 − p) × T _lb; (ii) ‘Zc ’ is the solution obtained by CPLEX; ‘Zt ’ is the global best solution
obtained by the tabu search algorithm; (iii) ‘Tc ’ is the CPLEX running time in seconds, ‘Tt ’ is the tabu search running time
in seconds; (iv) ‘N.A.’ means the solution cannot be obtained within 7200 s; (vi) ‘Zc ’ of instance ‘16-41-4-2-1’, ‘36-85-4-
2-2’ and ‘36-85-4-2-3’ is the solution obtained in 7200 s.

Table 3
Performance of the tabu search metaheuristic for small-scale instances.

Instances CPLEX Tabu search Time ratio

Scale ID Zc Tc (s) Zi Zt GAPit (%) Tt (s) Tt /Tc

Six areas 6-18-3-1-1 1378.33 4.32 1471.59 1378.33 6.338% 0.41 0.094
and 54 6-18-3-1-2 1627.60 4.64 1631.15 1627.60 0.218% 1.59 0.342
space nodes 6-18-3-1-3 1631.15 0.84 1725.11 1631.15 5.447% 0.34 0.405
6-18-3-1-4 1891.54 6.65 1895.09 1891.54 0.187% 1.73 0.260
6-18-3-1-5 1934.73 3.01 1989.05 1934.73 2.731% 1.27 0.422
Eight areas 8-23-3-1-1 2556.18 5.97 2610.50 2556.18 2.081% 2.03 0.340
and 69 8-23-3-1-2 1937.00 10.44 2280.76 1937.00 15.072% 4.20 0.402
space nodes 8-23-3-1-3 2450.86 8.26 2544.13 2450.86 3.666% 4.96 0.600
8-23-3-1-4 2200.36 7.03 2280.76 2200.36 3.525% 3.61 0.450
8-23-3-1-5 2454.26 8.03 2548.05 2454.26 3.680% 3.29 0.410
Nine areas 9-25-3-2-1 2624.14 4473.31 2941.26 2624.14 10.782% 78.59 0.018
and 75 9-25-3-2-2 3049.35 17382.44 3339.55 3049.35 8.690% 85.71 0.005
space nodes 9-25-3-2-3 2717.90 2082.75 2879.18 2717.90 5.601% 50.30 0.024
9-25-3-2-4 3192.73 4240.91 3507.81 3192.73 8.982% 13.66 0.003
9-25-3-2-5 3247.96 4572.59 3536.25 3247.96 8.152% 46.11 0.010
Average 2184.53 5.677% 19.82 0.009

Notes: (i) ‘Zc ’ is the optimal solution obtained by CPLEX; ‘Zi ’ is the initial solution obtained by the tabu search
algorithm; ‘Zt ’ is the global best solution obtained by the tabu search algorithm; (ii) ‘Tc ’ is the CPLEX running time
in seconds, ‘Tt ’ is the tabu search running time in seconds; (iii) GAPit = (Zi − Zt )/Zt ; (iv) T ime ratio = Tt /Tc .

Table 4
Performance of the tabu search metaheuristic for medium-scale instances.

Instances CPLEX Tabu search Obj gap

Scale ID Zf Zi Zt GAPit (%) Tt (s) GAPtf (%) GAPft (%)

12 areas 12-32-4-2-1 3678.18 3927.79 3674.77 6.442% 328.69 ࢨ 0.093%


and 128 12-32-4-2-2 3521.91 3843.58 3490.75 9.180% 462.61 ࢨ 0.893%
space nodes 12-32-4-2-3 3293.79 3545.10 3293.79 7.089% 332.57 0.0 0 0% 0.0 0 0%
12-32-4-2-4 3535.02 3843.58 3535.02 8.028% 589.23 0.0 0 0% 0.0 0 0%
12-32-4-2-5 3042.11 3369.91 3042.11 9.727% 482.23 0.0 0 0% 0.0 0 0%
16 areas 16-41-4-2-1 3745.72 4212.21 3745.72 11.075% 1148.59 0.0 0 0% 0.0 0 0%
and 164 16-41-4-2-2 4016.44 4638.81 4021.17 13.315% 869.51 0.118% ࢨ
space nodes 16-41-4-2-3 3942.47 4195.49 3942.68 6.026% 832.82 0.005% ࢨ
16-41-4-2-4 4605.29 5380.52 4608.72 14.344% 1111.61 0.074% ࢨ
16-41-4-2-5 4498.83 5059.16 4495.80 11.135% 766.93 ࢨ 0.067%
20 areas 20-50-4-2-1 5935.39 6463.56 5817.84 9.990% 2019.73 ࢨ 2.021%
and 200 20-50-4-2-2 4449.53 5671.62 4462.19 21.324% 2134.33 0.285% ࢨ
space nodes 20-50-4-2-3 5070.44 5767.87 5096.35 11.642% 2640.24 0.511% ࢨ
20-50-4-2-4 5388.98 5771.07 5384.82 6.693% 1489.47 ࢨ 0.077%
20-50-4-2-5 4287.27 5130.17 4289.21 16.392% 2450.68 0.045% ࢨ
Average 10.827% 1177.28 0.104% 0.350%

Notes: (i) ‘Zf ’ is the feasible solution obtained by CPLEX in 7200 s; (ii) GAPt f = (Zt − Z f )/Z f , GAPf t = (Z f − Zt )/Zt ; (iii)
‘ࢨ’ means the value is negative.

gap between them is within one percent. On large-scale instances is 8.585%, which means that the tailored heuristic for generating
where CPLEX cannot obtain feasible solutions, our tabu search initial solutions and the tabu search strategy are both efficient.
metaheuristic can provide feasible solutions within three hours. In To sum up, our tests suggest that the proposed solution method
addition, the average gap between the initial objective values and is appropriate for solving the formulated model and has a definite
the global best objective values of tabu search among all instances
10 L. Zhen, M. Li and G. Laporte et al. / Computers and Operations Research 105 (2019) 1–11

Table 5
Performance of the tabu search metaheuristic for large-scale instances.

Instances CPLEX Tabu search Obj gap

Scale ID Zf Zi Zt GAPit (%) Tt (s) GAPtf (%) GAPft (%)

36 areas 36-85-4-2-1 8927.95 9284.87 8771.45 5.853% 3717.17 ࢨ 1.784%


and 340 36-85-4-2-2 6376.44 6758.33 6336.73 6.659% 4237.01 ࢨ 0.632%
space nodes 36-85-4-2-3 7597.73 8588.96 7597.73 13.046% 3863.69 0.0 0 0% 0.0 0 0%
36-85-4-2-4 7734.31 8590.02 7723.16 11.224% 3430.07 ࢨ 0.144%
36-85-4-2-5 7706.90 8388.84 7706.90 8.849% 4387.99 0.0 0 0% 0.0 0 0%
64 areas 64-145-4-2-1 13294.27 14978.76 13557.89 10.480% 6475.82 1.983% ࢨ
and 580 64-145-4-2-2 13864.48 14363.47 12970.08 10.743% 6098.18 ࢨ 6.896%
space nodes 64-145-4-2-3 13960.55 14753.34 14006.22 5.334% 6128.70 0.327% ࢨ
64-145-4-2-4 15455.08 16989.77 15455.08 9.930% 5888.28 0.0 0 0% 0.0 0 0%
64-145-4-2-5 14687.20 16592.13 14687.20 12.970% 5751.95 0.0 0 0% 0.0 0 0%
81 areas 81-181-4-3-1 N.A. 20958.09 19601.66 6.920% 9503.18 - -
and 724 81-181-4-3-2 N.A. 18638.92 17731.09 5.120% 8743.50 - -
space nodes 81-181-4-3-3 N.A. 22371.88 19960.64 12.080% 9039.80 - -
81-181-4-3-4 N.A. 19163.78 17416.87 10.030% 9904.12 - -
81-181-4-3-5 N.A. 20082.60 18336.93 9.520% 9209.13 - -
Average 9.251% 6425.24 0.385% 1.182%

Notes: (i) ‘N.A.’ means the solution cannot be obtained within 7200 s.

advantage over CPLEX in terms of computation time. Moreover, in Acknowledgments


contrast to CPLEX, it always yields a feasible solution.
This work was supported by the National Natural Science Foun-
dation of China [grant numbers 71831008, 71671107, 71422007]
7. Conclusions and the Canadian Natural Sciences and Engineering Research Coun-
cil [grant number 2015-06189]. Thanks are due to the reviewers for
Motivated by the increasing use of UAVs to perform monitoring their valuable comments.
tasks both in the military and civilian areas, we have introduced a
new UAV routing problem that considers the height at which the
nodes are visited. The problem was formulated as an integer lin-
References
ear programming model in order to minimize the total duration
needed by all UAVs for monitoring a set of areas with given ac- Alotaibi, K.A., Rosenberger, J.M., Mattingly, S.P., Punugu, R.K., Visoldilokpun, S., 2018.
curacy requirements, and without exceeding the maximum flight Unmanned aerial vehicle routing in the presence of threats. Comput. Ind. Eng.
time of each UAV. While CPLEX can solve the model for small-scale 115, 190–205.
Avellar, G.S.C., Pereira, G.A.S., Pimenta, L.C.A., Iscold, P., 2015. Multi-UAV routing
instances, we had to develop a tabu search metaheuristic for the for area coverage and remote sensing with minimum time. Sensors 15 (11),
solution of medium- and large-scale instances. Extensive numerical 27783–27803.
experiments were conducted on several families of instances. Our Barrientos, A., Colorado, J., Cerro, J.D., Martinez, A., Rossi, C., Sanz, D., Valente, J.,
2011. Aerial remote sensing in agriculture: a practical approach to area coverage
results demonstrate that the proposed tabu search metaheuristic and path planning for fleets of mini aerial robots. J. Field Rob. 28 (5), 667–689.
has a good solution efficiency and always yields a feasible solution. Beard, R.W., McLain, T.W., Goodrich, M.A., Anderson, E.P., 2002. Coordinated target
When instance size grows, its relative advantage compared with assignment and intercept for unmanned air vehicles. IEEE Trans. Robot. Autom.
18 (6), 911–922.
CPLEX becomes more significant. The model and the algorithm de- Caillouet, C., Razafindralambo, T., 2017. Efficient deployment of connected un-
veloped in this paper can pave the way for the development of manned aerial vehicles for optimal target coverage. In: Proceedings of the
effective decision support systems for the routing and scheduling Global Information Infrastructure and NETWORKING Symposium (GIIS), pp. 1–8.
Casbeer, D.W., Holsapple, R.W., 2011. Column generation for a UAV assignment
of UAVs that must perform monitoring tasks.
problem with precedence constraints. Int. J. Robust Nonlinear Control 21 (12),
This work constitutes a starting point to the study of a new 1421–1433.
UAV routing problem. There is a certain gap between the problem Chen, P., Guo, Y., Lim, A., Rodrigues, B., 2006. Multiple crossdocks with inventory
we proposed and a realistic UAV monitoring application. The model and time windows. Comput. Oper. Res. 33 (1), 43–63.
Cho, J., Lim, G., Biobaku, T., Kim, S., Parsaei, H., 2015. Safety and security man-
formulated for the problem in this paper considered a static sce- agement with unmanned aerial vehicle (UAV) in oil and gas industry. Procedia
nario, the velocity of UAVs was assumed to be constant, whereas in Manuf. 3, 1343–1349.
reality the velocity of UAVs varies according to different altitudes. Chow, J.Y.J., 2016. Dynamic UAV-based traffic monitoring under uncertainty as
a stochastic arc-inventory routing policy. Int. J. Transp. Sci. Technol. 5 (3),
In addition, the objective function used in this paper was the min- 167–185.
imization of the total time required to perform all tasks of UAVs. Coelho, B.N., Coelho, V.N., Coelho, I.M., Ochi, L.S., Roozbeh, H.K., Zuidema, D.,
However, minimizing the number of UAVs and other practical fac- Lima, M.S.F., Costa, A.R.D., 2017. A multi-objective green UAV routing problem.
Comput. Oper. Res. 88, 306–315.
tors should be taken into account in UAV monitoring applications. Di Puglia Pugliese, L., Guerriero, F., Natalizio, E., Zema, N.R., 2017. A biobjective
Future work will be devoted to extending the model from a formulation for filming sport events problem using drones. In: Proceedings of
static scenario to a dynamic scenario in order to be closer to prac- the IEEE International Conference on Intelligent Data Acquisition and Advanced
Computing Systems: Technology and Applications, pp. 21–23.
tical applications. In a dynamic scenario, the UAV velocity is not
Di Puglia Pugliese, L., Guerriero, F., Zorbas, D., Razafindralambo, T., 2016. Modelling
constant. As a result, the enhanced model will consider the energy the mobile target covering problem using flying drones. Optim. Lett. 10 (5),
consumption of each UAV. The problem we have addressed in this 1021–1052.
Glover, F., 1986. Future paths for integer programming and links to artificial intelli-
work was formulated as a single-objective optimization model. We
gence. Comput. Oper. Res. 13 (5), 533–549.
will consider adding other objective functions which are the mini- Guerriero, F., Surace, R., Loscr, V., Natalizio, E., 2014. A multi-objective approach for
mization of the number of UAVs and the maximization of average unmanned aerial vehicle routing problem with soft time windows constraints.
monitoring satisfaction. In this context, each area is assigned a soft Appl. Math. Model. 38 (3), 839–852.
Harwin, S., Lucieer, A., 2012. Assessing the accuracy of georeferenced point clouds
time window and the time at which a UAV starts monitoring an produced via multi-view stereopsis from unmanned aerial vehicle (UAV) im-
area can be used to measure satisfaction. agery. Remote Sens. (Basel) 4 (6), 1573–1599.
L. Zhen, M. Li and G. Laporte et al. / Computers and Operations Research 105 (2019) 1–11 11

Kim, J., Song, B.D., Morrison, J.R., 2013. On the scheduling of systems of UAVs and Phung, M.D., Cong, H.Q., Dinh, T.H., Ha, Q., 2017. Enhanced discrete particle swarm
fuel service stations for long-term mission fulfillment. J. Intell. Robot. Syst. 70 optimization path planning for UAV vision-based surface inspection. Autom.
(1–4), 347–359. Constr. 81, 25–33.
Lai, D.S.W., Demirag, O.C., Leung, J.M.Y., 2016. A tabu search heuristic for the het- Sundar, K., Rathinam, S., 2017. Algorithms for heterogeneous, multiple depot, mul-
erogeneous vehicle routing problem on a multigraph. Transp. Res. Part E: Logist. tiple unmanned vehicle path planning problems. J. Intell. Robot. Syst. 88 (2–4),
Transp. Rev. 86, 32–52. 513–526.
Li, D., Wang, X., Sun, T., 2016. Energy-optimal coverage path planning on topo- Torres, M., Pelta, D.A., Torres, J.C., 2016. Coverage path planning with unmanned
graphic map for environment survey with unmanned aerial vehicles. Electron. aerial vehicles for 3D terrain reconstruction. Expert Syst. Appl. 55 (C), 441–451.
Lett. 52 (9), 699–701. Toth, P., Vigo, D., 2014. Vehicle Routing: Problems, Zzmethods, and Applications.
Manyam, S.G., Rathinam, S., Darbha, S., 2017. Computation of lower bounds for a MOS-SIAM Series on Optimization. SIAM, Philadelphia.
multiple depot, multiple vehicle routing problem with motion constraints. J. Vilar, R.G., Shin, H.S., 2013. Communication-aware task assignment for UAV cooper-
Dyn. Syst. Meas. Control 137 (9), 1–5. ation in urban environments. IFAC Proceedings 46 (30), 352–359.
Miller, C.E., Tucker, A.W., Zemlin, R.A., 1960. Integer programming formulation of Wen, M., Larsen, J., Clausen, J., Cordeau, J.-F., Laporte, G., 2009. Vehicle routing with
traveling salesman problems. J. ACM 7 (4), 326–329. cross-docking. J. Oper. Res. Soc. 60 (12), 1708–1718.
Mozaffari, M., Saad, W., Bennis, M., Debbah, M., 2016. Efficient deployment of multi- Xia, Y., Batta, R., Nagi, R., 2017. Controlling a fleet of unmanned aerial vehicles
ple unmanned aerial vehicles for optimal wireless coverage. IEEE Commun. Lett. to collect uncertain information in a threat environment. Oper. Res. 65 (3),
20 (8), 1647–1650. 674–692.
Murray, C.C., Chu, A.G., 2015. The flying sidekick traveling salesman problem: opti- Xu, A., Viriyasuthee, C., Rekleitis, I., 2014. Efficient complete coverage of a known
mization of drone-assisted parcel delivery. Transp. Res. Part C: Emerg. Technol. arbitrary environment with applications to aerial operations. Auton. Robots 36
54, 86–109. (4), 365–381.
Nedjati, A., Izbirak, G., Vizvari, B., Arkat, J., 2016. Complete coverage path planning Xu, D., Li, K., Zou, X., Liu, L., 2017. An unpaired pickup and delivery vehicle routing
for a multi-UAV response system in post-earthquake assessment. Robotics 5 (4), problem with multi-visit. Transp. Res. Part E: Logist. Transp. Rev. 103, 218–247.
26. Zhang, S., Zhou, Y., Li, Z., Pan, W., 2016. Grey wolf optimizer for unmanned combat
Oh, H., Kim, S., Tsourdos, A., White, B.A., 2014. Coordinated road-network search aerial vehicle path planning. Adv. Eng. Softw. 99, 121–136.
route planning by a team of UAVs. Int. J. Syst. Sci. 45 (5), 825–840. Zorbas, D., Di Puglia Pugliese, L., Razafindralambo, T., Guerriero, F., 2016. Optimal
Otto, A., Agatz, N., Campbell, J., Golden, B., Pesch, E., 2018. Optimization approaches drone placement and cost-efficient target coverage. J. Netw. Comput. Appl. 75,
for civil applications of unmanned aerial vehicles (UAVs) or aerial drones: a sur- 16–31.
vey. Networks 00, 1–48.

You might also like