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

fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TMC.2021.3051273, IEEE
Transactions on Mobile Computing

SARDO: An Automated Search-and-Rescue

Drone-based Solution for Victims Localization
Antonio Albanese, Student Member, IEEE, Vincenzo Sciancalepore, Senior Member, IEEE,
Xavier Costa-Pérez, Senior Member, IEEE

Abstract—Natural disasters affect millions of people every year. Finding missing persons in the shortest possible time is of crucial
importance to reduce the death toll. This task is especially challenging when victims are sparsely distributed in large and/or
difficult-to-reach areas and cellular networks are down.
In this paper we present SARDO, a drone-based search and rescue solution that leverages the high penetration rate of mobile phones
in the society to localize missing people. SARDO is an autonomous, all-in-one drone-based mobile network solution that does not
require infrastructure support or mobile phones modifications. It builds on novel concepts such as pseudo-trilateration combined with
machine-learning techniques to efficiently locate mobile phones in a given area. Our results, with a prototype implementation in a
field-trial [1], show that SARDO rapidly determines the location of mobile phones (∼ 3 min/UE) in a given area with an accuracy of few
tens of meters and at a low battery consumption cost (∼ 5%).
State-of-the-art localization solutions for disaster scenarios rely either on mobile infrastructure support or exploit onboard cameras for
human/computer vision, IR, thermal-based localization. To the best of our knowledge, SARDO is the first drone-based cellular
search-and-rescue solution able to accurately localize missing victims through mobile phones.

Index Terms—UAV-based cellular coverage, Search and Rescue Operations, Single-UAV localization, Pseudo-Trilateration,
Convolutional Neural Network (CNN), Long Short-Term Memory (LSTM).

1 I NTRODUCTION e.g., smart phones and wearables, that can be detected by UAV
mobile networks.
When a natural disaster unexpectedly strikes, the most compelling Currently, the users Global Navigation Satellite System
challenge is to quickly establish a first contact with affected (GNSS)-based positions are not directly accessible by network
victims that might be trapped or hidden from rescue teams without providers as 3rd Generation Partnership Project (3GPP) standards
any possibility to communicate. It is estimated that in 2017, do not include any Application Programming Interface (API) for
335 natural disasters affected over 95.6 million people, killing this purpose. Specifically, the only way for a network provider
an additional 9, 697 and costing a total of 335 billion USD. to obtain such information would be a client application running
The deadliest event in 2017 was the landslide in Sierra Leone on each User Equipment (UE) thereby actively making its GNSS
in August, with 1102 reported dead or missing, followed by position available. Conversely, while building an emergency local-
Cyclone Okchi in December with 884 reported dead or missing ization system, we may not rely on any active users intervention or
in India. Notably, these two events are characterized by a high client application running on the UEs, but rather seek a standalone
number of missing people, representing over half of the total death design to cope with the unexpected events of a disaster. Moreover,
toll [2]. In addition to natural disasters, human-generated threats GNSS signals may be heavily attenuated by debris or rubble,
(e.g. fires, electrical outages, terrorism) might as well require ad- which would likely not completely block the cellular coverage
hoc solutions to improve first responders’ capabilities and further provided by a UAV hovering close enough to the victims UEs.
support their modus operandi within such critical situations. As additional attenuation may be introduced by tree foliage, metal
Unmanned Aerial Vehicles (UAVs) or Drones have recently or wooden objects, few centimeters of water, shadowing in urban
emerged as a cost-efficient alternative to address emergency canyons and so on, GNSS is unreliable for such a mission critical
scenarios for multiple reasons [3]. First, UAVs can be rapidly operation [4].
deployed in disaster areas providing on-demand mobile networks. In this paper, we present our Search-And-Rescue DrOne-
Second, UAVs may rapidly approach difficult-to-reach locations, based solution, SARDO, an all-in-one localization system that
such as mountains, deserts, or devastated areas and cover large supports first responders to quickly identify and localize victims
search areas with sparse victims distribution. Finally, given the in disaster areas even if the terrestrial communication network is
high penetration rate of mobile devices in our society, it can be down in the aftermath of a natural disaster. SARDO i) scans and
reasonably assumed that victims are equipped with smart devices, spots target users by means of an integrated International-Mobile-
Subscriber-Identity(IMSI)-catcher, ii) applies machine-learning
A. Albanese is with NEC Laboratories Europe and University Carlos III of principles on the distance measurements to perform our novel
Madrid, 28911 Leganés, Spain. pseudo-trilateration localization technique, iii) relies on a neural
V. Sciancalepore is with NEC Laboratories Europe, Heidelberg, Germany. network to predict future target positions, iv ) closes the feedback
X. Costa-Pérez is with NEC Laboratories Europe, i2CAT Foundation and
ICREA, Barcelona, Spain. loop with a control component that automatically adjusts the UAV
Emails: {antonio.albanese, vincenzo.sciancalepore, xavier.costa}@neclab.eu. trajectory for improving the localization accuracy. SARDO has

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TMC.2021.3051273, IEEE
Transactions on Mobile Computing
2

been implemented and tested in a field-trial scenario with com- ToF measurements User position
mercial off-the-shelf (COTS) components [1]. Our results prove process estimate
Disaster area Time-stamp
the feasibility of the solution and provide quantitative system coordinates User position

performance figures.

Section 6 Sections 4.1 - 4.2

1.1 Notation
UAV Position Future user positions
For the analysis section, we use conventional notation. We use Rn , Update prediction

and Rn×m to represent the sets of n-dimensional real vectors, and

m × n real matrices, respectively. Vectors are denoted by default
as column vectors and written in bold font. Subscripts represent Section 5
Pointwise operation
Layer
Section 4.3
an element in a vector and superscripts elements in a sequence.
(t) (t)
For instance, x(t) = [x1 , . . . , xn ]T is a vector from Rn and Fig. 1. Overview of SARDO’s Building Blocks
(t)
xi is its ith component. Operation [·]T represents the transpose
localization tasks executed within a predefined short time frame,
operator while || · ||2 denotes the L2-norm of a vector.
e.g., tens of seconds, by suitably adjusting the number of ToF
measurements. Upon completion of each task, which returns an
2 SARDO F RAMEWORK OVERVIEW estimate of the user trajectory as well as a prediction of its future
State-of-the-art localization solutions rely on trilateration and positions, the UAV relocates to a different location and resumes
triangulation methods, which combine measurements collected by the measurement process. The variable duration of each task im-
different anchors. In contrast, hereafter we present SARDO that pacts on the number of collected measurements that, in turn, drive
aims to find victims’ locations by keeping track of their mobile the accuracy of the pseudo-trilateration and the forecasting phase.
phone signals in disaster areas with the information collected by We design SARDO to accurately provide users positions within
a single UAV that sweeps a given predefined area and acts as a few minutes and explain its parameters in detail in Section 7.
portable cellular base station. The building blocks of SARDO are
depicted in Fig. 1. 3 G EO -L OCALIZATION M ODEL
Time-of-flight measurement process. A UAV equipped with The simplest localization technique proposed in the literature for
a light-weight base station scans on a predefined disaster area to static objects requires distinct reference points, dubbed as anchors,
discover victims1 . No protocol stack modifications are introduced to measure their distances from the target. It is known as trilater-
such that compatibility with commercial mobile phones is guar- ation, if 2-dimensional coordinates are sought, or multi-lateration
anteed, as described in Section 6. This allows measuring the time technique, in case of 3-dimensional coordinates2 . However, these
of flight (ToF) of a user uplink signal that is fed into the novel techniques suffer from major limitations when the measurements
pseudo-trilateration algorithm, as detailed in Section 3.2. get affected by noise and/or mobility. In the following, we describe
User position estimate. A single view-point, the UAV, ex- the multi-lateration technique pointing out its limitations and then
ploits time diversity to retrieve different user ToF values that are we introduce our novel technique, namely pseudo-trilateration,
combined to estimate the current user position as described in that is robust to noisy measurements and mobile objects.
Section 4. In addition, when the user moves an estimate of its
motion trajectory is derived by means of a Convolutional Neural
Network (CNN) that extracts features from different ToF values. 3.1 Legacy multi-lateration
Such features are processed by the output deep Feed-Forward Satellite localization systems, encompassed under the Global Nav-
Neural Network (FFNN) that learns and implements the concept igation Satellite System (GNSS) umbrella, effectively achieve up
of pseudo-trilateration. to 5-meter localization accuracy [5]. Unfortunately, 3rd Gener-
Future user positions prediction. The returned set of pre- ation Partnership Project (3GPP) standards do not provide any
vious user positions is used to predict future locations. A Long technical mean (e.g., Application Programming Interfaces (APIs))
Short-Term Memory (LSTM) neural network keeps memory of for cellular service providers to directly access the GNSS in-
previous system states and forecasts a set of future user positions. formation of User Equipments (UEs). Therefore, retrieving such
This is explained in detail in Section 4.3. information would require a client application running at the UE
UAV Relocation. A large prediction time window results in a side in charge of periodically transmitting it to a remote collection
lower accuracy of the obtained prediction. Thus, we implement server. This calls for active users participation, which is infea-
a component that leverages control theory to adjust the next sible in emergency situations. Moreover, GNSS signals heavily
UAV motion trajectory. On the one hand, a high accuracy in the suffer from shadowing (induced by, e.g., foliage, wooden objects,
forecasting process fosters the UAV to focus on the expected user shallow water), thus becoming not very useful whenever victims
location and reduce the scanning area. On the other hand, detected are covered in rubble. As reliability is key in mission-critical
errors in the prediction process forces the UAV to enlarge the operations, telco operators would prefer to deploy and control
scanning area in order to recover from previous wrong decisions, their own robust localization solutions rather than depending on
as shown in Section 5. GNSS technologies, which are developed and managed by third
The execution of the entire localization process requires some (untrusted) parties.
time. Therefore, we design the system as a sequence of atomic On the other hand, current cellular networks—where user
localization techniques are implemented assuming base stations
1. Our analysis focuses on single-target detection. However, SARDO can
be used to locate multiple targets in a sequential manner or can be readily 2. Note that the terms “trilateration” and “multilateration” are used inter-
extended to locate users’ clusters by means of a customized implementation. changeably throughout the paper.

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TMC.2021.3051273, IEEE
Transactions on Mobile Computing
3

acting as anchors—may provide a localization accuracy up to where

hundreds of meters [6]. This is mostly due to the fact that Long- q
Term Evolution (LTE) leverages the Received Signal Strength fi = γi − γˆi = (px − xi )2 +(py − yi )2 +(pz − zi )2 − γˆi (3)
Indicator (RSSI) that roughly provides an estimation of the dis-
is a shortening for fi (x, y, z). We can calculate the gradient vector
tance based on pre-defined fingerprints. However, a close enough
g = [∂F/∂x, ∂F/∂y, ∂F/∂z] of F as
cellular base station would likely radiate enough power (at a
lower frequency than GNSS) to overcome the possible additional N N N
∂F X ∂fi ∂F X ∂fi ∂F X ∂fi
attenuation introduced by blocking debris and enable the reception =2 fi ; =2 fi ; =2 fi , (4)
∂x i=1
∂x ∂y i=1
∂y ∂z i=1
∂z
of useful cellular signals by the victims UEs. Additionally, for
each mobile user only few base stations—the ones covering and express it in matrix form as g = 2J T f , where f = {fi , ∀i ∈
the area wherein the user moves—may retrieve useful distance I} and J is the Jacobian matrix. Now applying the Newton
information thereby requiring full cooperation of a high number of iteration3 technique [8], we can find the approximate solution at
anchors in the area in order to jointly calculate the instantaneous step r k+1 = {x, y, z} as follows
user position. Ideally, a full deployment of base stations within
the considered area would minimize the localization error but, at r k+1 = r k − (JkT Jk )−1 JkT f k . (5)
the same time, increase the overhead due to the combination of
Note that Jk and f k are evaluated at r k whereas r1 = {x̂, ŷ, ẑ}
different measurements.
denotes the initial condition. The time complexity of this approach
Such localization solutions apply the well-known concept
is driven by the convergence speed of the Newton iteration pro-
of multi-lateration that involves a number of viewpoints (more
cess. Indeed, a fine-tuning can be applied to retrieve the position
than four in case of 3-dimensional geo-localization) simultane-
of the target with high accuracy within affordable time. However,
ously measuring the distance from the target. Analytically, we
dealing with mobile targets may exacerbate the problem complex-
assume I = |I| anchors where each anchor position xi =
ity thereby reducing the accuracy of the localization system.
{xi , yi , zi }, ∀i ∈ I is known a-priori. We can jointly compute
the position of the target object p = {px , py , pz } solving the
following set of equations 3.2 A novel technique: Pseudo-Trilateration
2 2 2
(px − xi ) + (py − yi ) + (pz − zi ) = γi2 , ∀i ∈ I, (1) The above-mentioned technique suffers from the following major
limitations: i) uncertainty: wireless channels are strongly affected
where the distance γi between the anchor i and the target object by fading and shadowing that may alter the distance measurements
t identifies the radius of a sphere centered in xi and passing and, in turn, the calculation process and ii) time-variability: the
through p. If the retrieved distances are not biased, the above classical solution requires more than four anchor nodes that might
set of equations admits a unique solution p represented by the be covered by different devices (space diversity) or by the same
intersection of all involved spheres. However, this result might device on different locations (time diversity); in both cases the user
not hold in real wireless environments due to the presence of may move and the trilateration process may report inconsistent
channel fading where the measured distance—denoted as γˆi —is results.
calculated through Time of Flight (ToF) or Time of Arrival (ToA). To overcome such limitations, we propose a novel localization
Indeed, this might lead to multiple intersection points between approach that does not require a highly-dense coverage of anchors
any possible pair of spheres and require an approximation on the and easily handles high-mobility users: the Pseudo-trilateration
location of the target. We can rewrite Eq. (1) as a linear system of concept. The core idea is to use a single anchor that retrieves mul-
(I − 1) equations in matrix form Sp = ψ , where tiple distance measurements over a time window moving through
xI − x1 yI − y1 zI − z1
 
  different points, namely along some anchor motion trajectory.
 xI − x2 y I − y2 zI − z2  px Such measurements are properly combined to identify the set of
S=  , p =  py 
 
.. .. .. positions covered by the target (if moving) within the considered
 . . . 
pz time window.
xI − xI−1 yI − yI−1 zI − zI−1 Let us denote the position of our target as a vector of co-
and (n) (n) (n)
ordinates p(n) = {px , py , pz } at time n, where n ∈ N
(γ̂12 − γ̂I2 ) − (x21 − x2I ) − (y12 − yI2 ) − (z12 − zI2 )
 
and N = |N | is the length of the time window. The motion
 (γ̂22 − γ̂I2 ) − (x22 − x2I ) − (y22 − yI2 ) − (z22 − zI2 )
 trajectory of the anchor point is defined as the set of positions
ψ= . (n) (n) (n) (n)
 
.. xd = {xd , yd , zd } at each time n ∈ N . We can rewrite
 . 
2 2 2 2 2 2 2 2 Eqs. (2) and (3) with the ensuing equation4 :
(γ̂I−1− γ̂I )−(xI−1−xI )−(yI−1−yI )−(zI−1−zI ) r
Let us denote as p∗ the approximated target position due to (n) 2 (n) 2 (n) 2
     
(n) (n) (n)
f = px −xd + py −yd + pz −zd −γ̂ (n) . (6)
(n)
the measurements noise, namely p∗ satisfying Sx∗ ≈ ψ . Finding
the derivative of the sum of the squares of the residuals yields As per the classical trilateration technique, the
 objective is to
that S T Sp∗ = S T ψ . While a non-singular S T S exhibits p∗ =

(n) (n) (n)
minimize the function f (n) xd , yd , zd that returns the
(S T S)−1 S T ψ as unique solution, we apply the non-linear least
squares method when it is close to singular that minimizes the sum error between the calculated distance γ̂ (n) and the real distance
of the squares of the errors on the distances. 3. We adopt the Newton iteration due to its high convergence speed towards
Let us denote the function to minimize as the following the steepest decent. However, other techniques can be readily applied, e.g., the
I I gradient descent as reported in [7].
X X
F (x, y, z) = (γi − γˆi )2 = fi2 , (2) 4. To avoid notation
 clutter, hereafter
 we refer to the error function
(n) (n) (n)
f (n) x(n) = f (n) xd , yd , zd as f (n) .


i=1 i=1

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TMC.2021.3051273, IEEE
Transactions on Mobile Computing
4

Mobile anchor

Wrong Correct Mobile anchor Wrong target Correct target

Mobile anchor
target target Mobile
Target target Mobile anchor

(a) 2D localization problem with static (b) 2D localization problem with static (c) 2D localization problem with mo- (d) 2D localization problem with mo-
target and single solution target and double solution bile target and single solution bile target and double solution

The solution of Problem 1 provides a set of feasible positions

covered by the target within the time window N . Note that the
Mobile Mobile anchor (1) (2) (N )
anchor Mobile
target
anchor trajectory S = {xd , xd , . . . , xd } may affect the set
Target of solutions of the problem and, in some cases, it might result in a
double optimal solution.
Solution
For this discussion we assume full knowledge of the distance
Solution
space
space values between the anchor and the target. We discuss the 2-
(e) 3D localization problem with (f) 3D localization problem with mo- dimensional case, as it makes our problem tractable, however
static target and infinite solutions bile target and infinite solutions it may be extended to the 3-dimensional case by applying the
same ideas. Let us consider a static target user, i.e., p(n) =
Fig. 2. Geometrical considerations on the localization p(m) , ∀n 6= m as well as a linear motion trajectory of the anchor,
(n) (n) (n) (n+1)
among the anchor point and the target every time n. However, i.e., yd = rxd + q with xd 6= xd . An example is
in this case, only one equation is expressed per time n with provided in Fig 2(b), where two optimal solutions are depicted.
the objective to find the solution that provides an error equal Thus, we can formulate the following lemma:
to 0. Unfortunately, this problem admits multiple solutions for Lemma 1. Considering a static target and a linear motion
each time n as they depict the coordinates of the points that trajectory of the anchor, Problem 1 always admits two distinct
geometrically lie on a sphere with radius equal to γ̂ (n) centered solutions. However, if the static target lies on the motion trajectory
(n)
on xd . of the anchor, Problem 1 admits one single solution with double
To overcome this problem, we can minimize the sum of multiplicity.
the Euclidean distances—namely L2-norm—between every two
(n)
consecutive solutions, each of which is obtained at time n ∈ N . Proof. Let us consider two different positions of the anchor xd
Thus, the overall path length is minimized ([9]). We can formulate at any time n, m where n 6= m. Using Eq. (1), we can derive the
it with a convex optimization model: following set of equations
(n) (n) (n) (n)
Problem 1 (Pseudo-trilateration). (x−xd )2+(y−(rxd +q))2−(px −xd )2−(py−(rxd +q))2 = 0
(m) (m) (m) (m)
N
X   2
(x−xd )2+(y−(rxd +q))2−(px−xd )2−(py−(rxd +q))2=0
min p(n) − p(n−1) (8)
p(n) ∈R3
n=2
2qr+r 2 px −2rpy −px
s.t. f(n)
= 0, ∀n ∈ N ; (7) that provides a double root x = [px − r 2 +1 ] and
2q+r 2 py +2rpx −py
p(n) ∈ R3 , ∀n ∈ N . y= [py r 2 +1 ].
Clearly, when the target position lies
on the motion trajectory, i.e., py = rpx +q , the double root results
Problem 1 reveals its sublinearity and can be solved with in x = [px py ] and y =[py px ]. This proves the lemma.
commercial tool for convex optimization. We can then formulate Graphically, we show the double solution (intersection point)
the following theorem. of Eqs. (8) in Figs. 2(a)-2(b) wherein each circumference repre-
Theorem 1. Problem 1 is NP-Hard. sents one equation.
Now let us consider a nomadic target. From Lemma 1, we can
Sketch of Proof: The proof goes by reduction. Let us consider as claim the following theorem.
input γ̂ (n) = 1, ∀n ∈ N . Now we can find an infinite (with N
multiplicity) number of subsets p(n) that satisfy the first constraint Theorem 2. Problem 1 admits a double solution iff the motion
(1) (2) (N )
of Problem 1. The new problem is to find a subset of such values trajectory S = {xd , xd , · · · , xd } of the anchor is linear.
with the minimum sum. This can be easily (within polynomial
Proof. Similarly to the proof of Lemma 1, we write and solve
time) mapped onto a subset-sum problem that is known to be NP-
Eqs. (8) for each position of the target. The solutions (single or
Complete. The NP-completeness property implies that Problem 1
double) of each system of equations exhibit the symmetry property
belongs to the NP class as well as to NP-hardness class.
with respect to the anchor trajectory. If the anchor is moving along
With the above theorem, we prove that the solution of Prob- the same direction of the target, the solution of each system is
lem 1 cannot be found within affordable time: while a low- single with double multiplicity, as previously proved.
complex heuristics can be applied to solve it, its optimality cannot
be verified within polynomial time. The heuristics only solves the A graphical illustration is provided in Figs. 2(c)-2(d). As
problem locally for each n ∈ N by finding the coordinates of the expected, when the anchor (blue dot) is moving along the direction
next point p(n) with the shortest distance from the previous one of the target (red dot), the solution of Problem 1 reveals the exact
p(n−1) . This results in a suboptimal global solution. position of the target (with double multiplicity). Conversely, when

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TMC.2021.3051273, IEEE
Transactions on Mobile Computing
5

the motion trajectory is different from the direction of the target

0
but still linear, the optimization problem results in two distinct 210
optimal solutions. 180

25
row index

Distance [m]
Analogously, in 3-dimensional space the motion trajectory 150

50
affects the solution of Problem 1. In this case, when the anchor
direction is linear and the target position is static, there are infinite 120

m
75
optimal solutions that geometrically lie on the circumference 90
centered on the motion trajectory line, orthogonal to it and passing
0 25 50 75 100

100
through the target position, as depicted in Fig. 2(e). When the tar- m column index
get moves, all solutions lie on analogously defined circumferences
as depicted in Fig. 2(f). Fig. 3. Heatmap of matrix Γm of distance measurements in a single-
As a result, to avoid ambiguity in the solution of Problem 1, obstacle scenario.
based on Theorem 2 the anchor trajectory must change its direc- the UAV trajectory, we concatenate each Γm with a matrix Dm
tion within a finite time. Nevertheless, this may cause the mobile that contains the 3D coordinates of all UAV measurements spots,
anchor to get far away from the target reducing the receive signal- as depicted in the following. We name the above mentioned matrix
to-noise ratio and so the accuracy. Therefore, we rely on a close Φm and express it as follows
trajectory of the mobile anchor. For the sake of simplicity, we  (1) (1) 
model the anchor trajectory as circular. γ̂1 . . . γ̂L x1 y1 z1
 (2) (2)
While considering shortest paths in our calculation may appear  γ̂1 . . . γ̂L x2 y2 z2 

reasonable, in this work we go one step beyond by proposing Φm =   .. .. .. .. .. .
..  (9)
 . . . . . . 
in the next section an AI-based localization system that applies (N ) (N )
the pseudo-trilateration method, automatically learns the common γ̂1 ... γ̂L xN yN zN
| {z }
motion behaviors and calculates the positions with high accuracy.
| {z }
Γm Dm

To understand the rationale behind our neural network design, let

4 AI-BASED E NHANCED LOCALIZATION us assume a scenario wherein the target UE is partially covered by
The pseudo-trilateration method is based on the usage of a single a fixed obstacle, e.g. rubble. As shown in Fig. 3, the heatmap of
anchor node, e.g. a UAV flying along a predefined motion trajec- the corresponding Γm presents a stripe pattern due to the excess
tory locating a target user under its cellular coverage, as detailed error at the measurements spots where the wireless channel suffers
in Section 6. For the sake of simplicity, we assume that the UAV from the shadowing effect induced by the obstacle. Different
keeps flying on a certain fixed altitude5 . propagation environments lead to different patterns in Dm thereby
We develop a machine learning framework to solve the posi- suggesting to process each corresponding Φm as a single-channel
tioning problem, that is to infer the position of the target, dubbed picture. In such regard, we employ a 2D CNN whose usage is
as User Equipment (UE) according to the LTE nomenclature, well established in indoor localization and have been shown to be
given the raw ranging measurement series. In the following, we capable of achieving sub-meter accuracy with custom Ultra Wide
provide some insights on the measurement data preprocessing, Band (UWB) signals in indoor scenarios [10].
the Convolutional Neural Network (CNN) design as well as on its
training dataset. Note that, given the low-complexity of pre-trained 4.2 2D CNN design and training
neural networks, SARDO can be scaled to several target UEs. 2D CNNs are a class of feed-forward neural networks that make
use of convolutional layers to extract information from 2D input
4.1 Data preprocessing data. Generally, they consist of a bunch of modules that include
Let us consider any close trajectory along which the UAV may fly one convolutional and one pooling (or subsampling) layers, and
during the ToF measurements collection. For the sake of tractabil- they may be repeated to build a deeper model. In addition, some
ity, we assume that the UAV takes measurements in a subset of fully connected layers are stacked onto the last module to provide
equally-spaced points on its trajectory, namely measurement spots. the final output of the network. In particular, for each Φm , we
Moreover, we assume that all measurements are taken at the same train the network to regress Um , which denotes the matrix of the
time instant, meaning that there is no time delay between any 2D user coordinates p(n) , ∀n ∈ N corresponding to each UAV
measurement in the same spot6 . Let us denote as N the number measurement spot. We denote as Ûm the matrix of the regressed
of measurement spots and L the number of measurements taken user coordinates p̂(n) , ∀n ∈ N .
at each spot. The set of measurements taken by the UAV during From an initial shallow CNN design, we incrementally in-
(n)
a single revolution is γ̂l , ∀n ∈ N , l ∈ L, with N = |N | and crease the deepness of the network stopping at the level that pro-
L = |L|. For each revolution, we arrange the measurement series vides the highest performances while not overfitting our dataset.
into a 2D data structure where the time and spatial relationships Indeed, as our problem is NP-Hard by Theorem 1, the network
among subsequent measurements may be better represented. In input-output relationship is complex, thus likely to be better
particular, we build a matrix Γm , ∀m ∈ M, of size N × L, approximated by a deeper CNN. Our final design involves two
where M is the set of all close trajectories. To take into account modules and three fully connected layers. During the training
phase, we aim at offering the network a broad range of labeled
5. While this assumption complies with national regulations, it may be samples. To do so, we simulate several UAV and user trajectories,
simply relaxed by adding the UAV altitude to the retrieved information.
and generate a synthetic dataset as described in Section 4.1. In
6. This assumption makes our analysis tractable. However, in practice
the ToF measurements are performed within few milliseconds, as shown in Section 3.2, we have shown that a close line satisfies Theorem 2.
Section 7. Indeed, to take into account any possible perturbation of the

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TMC.2021.3051273, IEEE
Transactions on Mobile Computing
6

Suggested Algorithm 1 UAV Relocation procedure

UAV trajectory

1) Initialise the offset values δx , δy , δρ to 0.

Perturbed
UAV trajectory
2) Update the UAV trajectory parameters with the following
equations xc += δx , yc += δy , ρ+= δρ .
3) Calculate the spatial average p̄ = {p̄x , p̄y } of the predicted
(f ) (f )
user positions p̂(f ) = {p̂x , p̂y } with the following
F F
1 X (f ) 1 X (f )
Real user p̄x = p̂x , p̄y = p̂y .
trajectory F F
f =1 f =1
Expected
user trajectory 4) Calculate the average predicted speed of the user as follows
F
1 X (f )
v̄t = ||p̂ − p̂(f −1) ||2 .
F −1
f =2

5) Set δx = xc − x̄t δy = yc − ȳt δρ = v̄v̄dt max ||p(f ) − p̄||2 .

f ∈F
6) Go to Step 2.
Fig. 4. Real example of localization error on simulated data applying
the pseudo-trilateration mechanism considering UAV circular perturbed
trajectory.

imposed UAV trajectory, we simulate UAV trajectories considering

sinusoidal characteristics [11] according to the following equation: R + Controller δx,δy ,δρ AI-based System γ
(n) - C(z) H(z)
xd = {(ρ + asin(2πn/N )) cos(2π/N ) + xc ,
(10)
(ρ + asin(2πn/N )) sin(2π/N ) + yc , h}, Feedback loop
(z-1)
∀n ∈ N , where {xc , yc , h} and ρ denote the coordinates of its
center and radius, respectively, and a ∈ R is a parameter such
that 0 < a < ρ. Some examples of such trajectories are shown in Fig. 5. Building blocks of the controller designed to automatically adjust
Fig. 4. Besides, to simulate realistic user trajectories, we employ the UAV position.
the so called SLAW mobility model [12].
adjustment process tractable, we assume that the UAV is instructed
to change the radius ρ of the circular motion trajectory by an
4.3 Encoder-Decoder LSTM design and training offset, namely δρ , and its center xc by a space offset, namely
We set the properties of the next UAV trajectory on the base δx , δy (and δz in case of 3−dimensional scenarios). The pseudo-
of the predicted future user behavior. In order to perform this code is listed in Algorithm 1.
prediction, we employ a multi-layered Long Short-Term Memory Predicted user positions p̂(f ) within the future time window F
(LSTM) that takes as input the current 2D-CNN output, namely are retrieved from the Encoder-Decoder LSTM block, as explained
the estimated current user trajectory, and outputs its expected evo- in Section 4.3. Such coordinates are spatially averaged to get the
lution. This problem is usually denoted as sequence-to-sequence next center value of the UAV trajectory xc , yc . The maximum
prediction problem. distance between two subsequent centers will determine the radius
Our method is based on the so-called Encoder-Decoder LSTM of the UAV trajectory. Additionally, we compute a confidence
Recurrent Neural Network [13]. The core idea is to map the input value that depends on the user average speed v̄t and the UAV
sequence to a fixed-length vector using an LSTM, which is the speed v̄d : if the user is moving faster our algorithm adds a safety
Encoder, and then map the latter vector onto the target sequence, margin to the radius offset δρ to keep the user close to the UAV
which is the Decoder LSTM. The Encoder and the Decoder trajectory coverage.
LSTMs are then followed by a fully connected feed-forward layer, Recalling from control theory, the designed controller aims at
which represents the output layer. In this way, the network creates reducing the difference between a reference signal R(t) and an
an internal fixed-dimensional vector representation of the input output signal O(t). Specifically, in our case we select the output
sequence and learns how to generate an output sequence of the signal from the overall system as the distance between the UAV
same or different length, namely the prediction. and the user, i.e., γ (n) whereas the reference is set to 0, as shown
Specifically, we train the network with a dataset made of in Fig. 5. Assuming our system running in a discrete domain, we
input-output pairs (Um , Fm ), ∀m ∈ M, where Fm is the matrix can write the z -transform [14] of the controlled system as the
containing the future user coordinates, that is p(f ) , ∀f ∈ F , with following7
F = |F|. We name F̂m the matrix containing the predicted user H(z) = 2δx (xc +(ρ+δρ)cos ω− p̄x )+2δy (yc +(ρ+δρ )sin ω− p̄y )
coordinates p̂(f ) , ∀f ∈ F .
+ 2δρ ((ρ+δρ +(xc +δx − p̄x ) cos ω+(yc +δy − p̄y ) sin ω,
(11)
5 DYNAMIC UAV R ELOCATION Ki
whereas the PI controller8 is defined as C(z) = Kp + z−1 . Fol-
Predicted user positions are used to improve the localization
process. After a complete revolution, SARDO automatically ma- 7. The transfer function is linearized at (0,0).
noeuvres the UAV by changing its position to be closer to the user 8. We design a Proportional-Integral (PI) controller due to its simplicity
and retrieve more accurate distance measurements. To make the while guaranteeing zero error in the steady-state.

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TMC.2021.3051273, IEEE
Transactions on Mobile Computing
7

UAV Smartphone
DJI Matrice Intel NUC SARDO

UAV Locations
UAV Trajectory User Trajectory
Relay
GPS module 2D CNN DJI SDK
②
① ③

New UAV New UAV

Trajectory Future User Trajectory
⑥ ⑤
Flight Trajectory UAV
Controller E-D LSTM Relocation
④

Fig. 7. SARDO prototype architecture with message flow sequence

numbers.
Instruments (NI) USRP B210 [20], equipped with omnidirectional
antennas. In order to devise an all-in-one solution, we deploy the
Fig. 6. SARDO in action. Video available at network backhaul and core domain as part of the srsLTE suite.
https://www.youtube.com/watch?v=9v3NNghq3so. This enables a quick and direct interaction with any single 3rd
Generation Partnership Project (3GPP) architectural component,
lowing the Ziegler-Nichols rules [15], we can calculate Kp = 0.1
such as Serving Gateway (SGW), Packet Data Network Gateway
and Ki = 0.11 to keep our system stable.
(PGW), Home Subscriber Server (HSS) or Mobility Management
Considering multiple UAV positions n ∈ N , the con-
Entity (MME) (hereafter described).
troller adaptively tunes the parameters to reduce the distances
Testbed. We securely set up this module of about 1kg weight
γ (n) , ∀n ∈ N that in turn translates in having the UAV trajectory
on board of an advanced DJI Matrice 600 Pro UAV [21] that is
accurately covering the expected positions of the user p̂x , p̂y .
able to carry up to 6kg payload, as shown in Fig. 6. In addition, to
Specifically, when the user moves around few points showing a
control the UAV trajectory, we developed a control iOS application
limited motion area, our automated solution reduces the circular
by means of the DJI Mobile Software Development Kit (SDK)9 ,
trajectory area, i.e., reduces δρ to focus onto a narrow area and
which has a twofold function: i) it fetches the UAV coordinates
retrieve higher accurate measurements (in the extreme case when
and relays them to the NUC board, where first the 2D CNN and
the user is static, the UAV is covering the minimum trajectory
then the Encoder-Decoder LSTM are executed to calculate the
area). Conversely, when the user moves within a larger area the
new UAV trajectory parameters according to Algorithm 1, ii) it
controller increases the coverage area trying to keep the user
retrieves such new settings and calculates the next UAV positions
within the close trajectory. In Section 7, we show the performance
that are set back on the UAV. This is achieved by means of a 2.4
results in terms of system stability over time and we finally show
GHz WiFi control channel that delivers information to the Intel
the overall system performance (running our algorithm) when the
NUC while a proprietary DJI wireless communication interface is
feedback loop is active against different user speeds.
used by the DJI framework, i.e. by the ad-hoc iOS application, to
deliver new UAV motion patterns. An anonymous online video is
6 SARDO I MPLEMENTATION available at [1].
We designed and prototyped SARDO going through an intensive To make our solution mobile infrastructure independent, as
engineering process resulting in the architecture summarized in highlighted in Section 1, we design the ToF measurements pro-
Fig. 7. SARDO runs on off-the-shelf equipment (listed in Table cessing module such that it does not require successful associa-
1) and does not require any protocol stack change to mobile users tions with the user equipments (UEs). For the sake of clarity, in
devices. the following we do not distinguish between the above mentioned
module and the UAV itself.
TABLE 1
Equipment for SARDO prototype

6.1 Mobile Infrastructure Independence

Equipment Model
FGPA Board NI USRP B210 SARDO does not rely on the successful UE attachment thereby
Embedded Computer Intel NUC7i7DNBE reducing the overall complexity of the system and improving the
UAV DJI Matrice 600 PRO effectiveness of rescue operations in emergency scenarios.
High-Gain Amplifier Mini-Circuits ZX60-V63+ The 3rd Generation Partnership Project (3GPP) prescribes that
Low-Noise Amplifier ZX60-33LNR-S+ [16]
the UE performs the Random Access Procedure (RAP) whenever
Directional Antennas 2 × 10 dBi
UE Samsung Galaxy Tab S2 it attempts to establish a connection with a base station (namely
eNodeB or eNB), e.g., initial access to the network or han-
dovers [22]. However, the completion of the RAP does not imply
Prototype architecture. The first building block is in charge
that the UE is attached to the eNB. Indeed, the UE establishes
of collecting ToF measurements from the target UE and processing
a Radio Resource Control (RRC) connection with the selected
them as described in Section 4.1. This block relies on the Software
eNB to access the required network resources. In particular, the
Defined Radio (SDR) technology building on top of srsLTE [17],
an open-source LTE-compliant software suite [18]—deployed on
9. The iOS application is only needed to control the UAV motion patterns
an Intel NUC Board [19] with 32GBs of RAM and 1.9 GHz 7th while processing runtime information. However, first responders may run a
generation CPU—that interfaces with an FPGA board, National standalone application that automatically triggers new UAV directions.

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TMC.2021.3051273, IEEE
Transactions on Mobile Computing
8

UE UAV eNB UAV MME cause #10: Implicitly detached. In this way, the UAV
Listening to SIB 4,5,6,7
MME advertises that the UE is set to deregistered state forcing
the UE to perform a new Attach Procedure. The UE sends
an ATTACH_REQUEST message to the UAV MME containing
Set highest priority only temporary identity information, e.g., the Globally Unique
frequency Temporary Identifier (GUTI).
Once the Globally Unique Temporary Identifier (GUTI) is re-
Set different TAC
ceived, the UAV MME sends an IDENTITY_REQUEST message,
TAU_REQUEST notifying its inability to derive the UE IMSI from its temporary
TAU_REJECT (code 10) identity information. The UE responds with its real IMSI in the
next IDENTITY_RESPONSE message.
Deregistered State After having received the IMSI, being the UAV not able to
ATTACH_REQUEST (GUTI) authenticate the UE because of missing authentication information
IDENTITY_REQUEST
in the Home Subscriber Server (HSS), the Attachment Procedure
is dropped with an ATTACH_REJECT. Albeit the UE fails to
IDENTITY_RESPONSE (IMSI)
attach with the UAV, it exchanges many messages making use
Unable to of several physical layer channels. In particular, we focus our
authenticate UE analysis on the Physical Uplink Shared CHannel (PUSCH) on
ATTACH_REJECT
which the so-called DeModulation Reference Signal (DMRS) is
periodically transmitted. This Reference Signal allows the UAV
Fig. 8. IMSI-Catcher message sequence chart to successfully demodulate the PUSCH that, in the following, is
exploited to perform UE distance measurements. As a high-speed
RRC layer is responsible for radio resource configuration and data connection is never established between the UAV and the
mobility management of connected UEs. In addition, RRC serves UE, this approach may be readily extended to several UEs without
as transport protocol for Non-Access Stratum (NAS) signaling saturating SARDO computing capabilities.
messages between a UE and its Mobile Management Entity
(MME).
We introduce specific changes to the srsLTE software but keep 6.2 Tunable precision
our system in full compliance with 3GPP standard guidelines. The UAV stores its current GNSS position (e.g., GPS) infor-
We update the NAS signaling10 for the Tracking Area Update mation and the uplink Demodulation Reference Signal (DMRS)
(TAU) Procedure so that the UE reveals its International Mobile received from the UE11 . The Time of Flight (ToF) is calculated
Subscriber Identity (IMSI) while exchanging messages with the by exploiting the ideal autocorrelation property of the sequences
base station, i.e., our UAV cell. In this way, it is straightforward used for DMRS. Indeed, the DMRS uses Constant-Amplitude-
to identify the UE and start the localization process within a short Zero-Autocorrelation (CAZAC) sequences known as Zadoff-Chu
disruption time window. Such an identity-awareness feature opens (ZC) [24]. Theh ZC sequence of odd-length
i N can be written as
n(n+1)/2+ln
up new use cases, e.g. searching for specific missing people or xq (n) = exp −j2πq N , where n = 0, 1, . . . , N −1,
locating specific targets for public safety purposes. Thus, our ToF l ∈ N and q ∈ {1, . . . , N − 1} is called the ZC sequence root
measurements processing module is built as an IMSI-catcher [23]. index. For the sake of simplicity, in LTE, l is set to 0. Moreover, for
Hereafter, we detail the minimal steps of our approach. ZC sequences of any length N , the zero autocorrelation property
IMSI-catcher. The message sequence chart is depicted in holds, namely it yields the following
Fig. 8. The UAV listens to System Information Block (SIB)
N −1
messages transmitted by existing ground base stations in the X
Rxx (m) = xq (n) x∗q (n + m) = δ(m), (12)
disaster area. SIB Type 4, 5, 6 and 7 messages carry the parameters
n=0
of the Cell Reselection Procedure. Such procedure aims at moving
the UE to the best cell of the selected operator. In LTE, this is where Rxx (m) and δ(·) denote the discrete periodic autocorrela-
accomplished by letting the UE assess all the frequencies and tion function of xq (n) at lag m and the Dirac delta, respectively,
Radio Access Technologies (RATs) based on the priority list and (·)∗ denotes the complex conjugate operation.
specified in the above mentioned SIB messages and then rank them Inspired by [25], we consider the known and the received
according to the respective radio link quality. The TAU procedure DMRSs in the discrete-frequency domain, namely X(k) and
is triggered as soon as the UE selects a new Tracking Area Code Y (k), respectively. Using the cross-correlation property of the
(TAC) different from the one it is currently camping on. Therefore, Discrete Fourier Transform (DFT), we calculate the circular cross-
we setup our UAV on the highest-priority frequency. correlation of the two sequences as IDFT{X(k) Y ∗ (k)}, where
This deteriorates the received power of the serving eNB, thus IDFT{·} denotes the Inverse Discrete Fourier Transform. Thus, we
enabling the inter-frequency and inter-RAT Cell Search Procedure, look for the magnitude peak of the sequence as its position returns
which eventually leads to a Cell Reselection towards our UAV. the delay of the received DMRS. In other words, this returns the
Moreover, we set a different TAC with respect to the camping ToF of the uplink signal transmitted by the UE. Regardless of
one in order to trigger a TAU_REQUEST from the UE. To this the ideal autocorrelation property of the DMRS, this procedure
request, our UAV MME responds with a TAU_REJECT with is constrained by the sampling the frequency of the time-domain
signal. Indeed, the above mentioned cross-correlation is sampled
10. SARDO achieves the disclosure of the UE identity relying on a design
choice of the current LTE standard. However, next cellular network generations 11. We have amended the public source code of srsLTE [18] to implement
may natively provide such information. the IMSI-catcher procedure and the ToF calculations.

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TMC.2021.3051273, IEEE
Transactions on Mobile Computing
9

at the same sampling frequency ∆f of the original signals and TABLE 2

the position of the peak is approximated to the closest time Empirical channel parameters
offset at the current sampling frequency. Therefore, its resolu-
Parameter Value Parameter Value
tion depends on the time interval ∆t between two subsequent
ηLoS 2.3 dB ηN LoS 34 dB
samples, that is, for an LTE signal bandwidth of 20 MHz and a 27.23 b 0.08
sampling frequency of 30.72 MHz, ∆d = c ∆t = c/∆f ≈ fc 1.8 GHz σsh 4 dB
3 × 108 m/s/30.72 MHz ≈ 9.8 m, where c is the speed of light.
Non Light of Sight (NLoS) communication. P (h, r) denotes the
To workaround this limitation in terms of resolution, the two
probability of LoS and is defined as follows
signals may be upsampled by a factor K before computing their
cross-correlation. Indeed, by tuning this parameter it becomes 1
P (h, r) = h



, (14)
possible to tune the desired precision of the ToF measurements. 1 + a exp −b arctan r −a
Unfortunately, in practical implementations parameter K cannot
where a and b are tuneable parameters depending on the environ-
be increased indefinitely. In particular, there is a tradeoff between
ment. As SARDO is designed to operate in a disaster scenario,
the upsampling factor and the accuracy of the ToF measurements
we take into account the case of a UE covered in rubble. To
given that the higher K , the lower the Signal-to-Noise Ratio
the best of our knowledge, there are no available measurement
(SNR) of the autocorrelation magnitude peak, this lowers the
campaigns deriving an excess path loss model for such scenario.
ability to recognize the peak that is involved in the receiver noise.
Nevertheless, we build upon similar works carried out for ground-
In our trials, choosing K = 4 provides the best performance.
to-ground propagation environments (e.g., [30]) and model the
additional loss as a constant value, uniformly drawn between
0 and 60 dB. Note that we consider the same rubble loss for
7 P ERFORMANCE EVALUATION every measurement taken in the same spot given the slow-varying
In this section we evaluate the performance of SARDO through an nature of the phenomenon. We assume that losses obtained from
exhaustive simulation campaign with synthetic traces followed by different UAV measurement spots are independent and identically
experimental results with a proof-of-concept implementation in a distributed.
rural environment. Performance metrics. To analyze and compare SARDO
Simulations data. We assume that the UAV has a constant against the ground-truth, we select two metrics, the former being
linear speed. In particular, we consider a discrete range of average the mean localization error for each single UAV revolution, the
user speeds while we set the UAV speed v̄d to 5 m/s12 . Moreover, latter being a similarity index inspired by the Jain’s Fairness
we set the UAV altitude h to 100 m, as suggested in [26]. This Index [31] and defined as follows:
is compliant with national regulations [27]. For testing purposes,  N 2
P (n) (n) 2
we generate circular UAV trajectories with center coordinates ||p − p̂ ||
{xc , yc , h} and radius ρ as per Eq. (10), updated over time via n=1
SI = N  2 . (15)
Algorithm 1. We limit ρ to lay within the range 50 − 250m as the N
P
||p(n) − p̂(n) ||2
radius length is driven by the following trade-off: on the one hand, n=1
the revolution time along the corresponding trajectory should be SI assumes values ranging from the worst case 1/N to the best
minimized to allow for a quick localization; on the other hand, we case 1 that shows the same localization error for every point of the
should account for a safety margin to ensure that the target user is user trajectory, i.e., the punctual localization error. The higher the
reachable, i.e., within a certain distance from the UAV trajectory SI, the higher the accuracy of the user trajectory reconstruction
coverage, even in cases where the Encoder-Decoder LSTM fails to by means of the 2D CNN, after the deduction of any bias in
predict future user positions. Note that we choose N and L to be the localization error. As per simulation environment, we use
equal to 100 and that we generate 5000 input data samples to infer Python 3.6.8 with Keras 2.2.4 [32] as a front-end for TensorFlow
the user trajectory. For any neural network training, we randomly 1.12 [33]. In addition, we use MATLAB R2018b to generate user
select 66.6%, 22.2% and 11.1% of the available data to build the trajectories according to the SLAW model with different settings.
training, testing and validation datatasets, respectively.
1e2
2 1e2
Channel models. The communication channel between the
UAV and the UE is an air-to-ground channel. Specifically, we
2
Error [m]

OPT (rand) OPT (rand, ideal) CNN

capitalize on the path loss model proposed in [28] by considering OPT (true) OPT (true, ideal)
y[m]

the slow fading effect, thereby modeling it as Ground truth 1

CNN estimate

4πfc
 0 OPT solution
0100
p 
P L(h, r) = 20 log + 20 log h2 + r2 + 1 0 1 2 3 700 1300 1900
c x[m] 1e2 Number of SLAW waypoints
P (h, r) ηLoS + (1 − P (h, r)) ηN LoS + ũ, (13)
Fig. 9. Example of the user trajec- Fig. 10. Localization error for different
where fc , h and r denote the carrier frequency, the UAV altitude tory for different solutions. SLAW waypoints.
and the 2D distance between the UAV and the UE, respectively
(and reported in Table 2 [29]). ũ is a log-normal random variable
7.1 Learning accuracy
with standard deviation σsh , whereas ηLoS and ηN LoS represent
the average additional losses in case of Line of Sight (LoS) and We evaluate the 2D CNN output based on the pseudo-trilateration
process. We develop and solve Problem 1 using a simple heuristics
12. When the UAV flight speed is set up to 5m/s the battery drain is limited, (as explained in Section 3.2). In Fig. 9, we compare the real
as reported in [26]. motion pattern of the user (target), the output of the AI-based

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TMC.2021.3051273, IEEE
Transactions on Mobile Computing
10

1e2 2.0 1e2 3 1e2 static

static static 2 m/s 4 m/s
1 m/s 3 m/s 4 m/s
1 m/s
2 2 m/s 1.5
Radius [m]
2

Error [m]
3 m/s

y [m]
4 m/s 1.0
1 1
0.5
1 3 5 7 1 3 5 7 0 100 200 300
Time [UAV revolutions] Time [UAV revolutions] x[m]

Fig. 11. Average UAV radius over time for different Fig. 12. Average localization error over time for Fig. 13. UAV centers trajectory for two differ-
user speeds. different user speeds. ent user speeds.

enhanced localization process (as described in Section 4.2) and very few outliers around 50 m. Conversely, when rubble is in
the solution of Problem 1 when the user is moving following place we obviously observe an increased error due to scattering
the SLAW mobility model considering 100 waypoints, namely and attenuation phenomena whereas the curve behavior is still
points of interest. In particular, the optimization problem solution invariant to the number of considered waypoints. Note that the
exhibits worse performance in case of noisy channel conditions. standard deviation augments due to the worsening of the receive
The figure shows that the optimization problem seeks the shortest SNR. Counter-intuitively, the SI for both case-scenarios is close to
path between two subsequent distance measurements. However, 1: this proves the robustness of the 2D CNN estimate against the
the performance of the optimization problem strongly depends error, even in case of large error, such as the one observed with
on the initial condition, i.e., on the first position of the solution rubble.
vector. In Fig. 10, we benchmark our 2D CNN against the
optimal solution using a random initial condition as well as a true
7.3 SARDO stability performance
localization value (only for the first point of the solution vector).
In addition, we evaluate the optimization problem solution when Being SARDO a closed-loop system, we evaluate its performance
noisy or ideal channel conditions are considered. As shown, the over time considering multiple UAV revolutions. For this purpose,
2D CNN outperforms the optimization problem with noisy and we simulate several scenarios with different average user speeds.
ideal channel conditions by a factor of 2 and 6, respectively. Note that the time is indexed by performed UAV revolutions
and that, being the UAV speed constant, the revolution time is
7.2 Pseudo-trilateration validation proportional to the trajectory radius.
The localization error increases with the user speed, being
We show the robustness of the introduced 2D CNN against diverse
the UAV speed set to the maximum user speed that allows for a
channel fading settings and user trajectories, the latter being
reasonably accurate localization. Albeit a user speed bound equal
generated by differently-tuned SLAW model instances.
to 5 m/s is compatible with a disaster scenario, it is always possible
To generate multiple user trajectories in such a way that they
to increase the UAV speed—up to 19 m/s for our particular UAV
are not correlated with the training set used during the neural
model—and cope with higher user speeds. It is worth noting that
network training phase, we set different numbers of waypoints
the localization error and the trajectory radius do not have a strictly
considered in the SLAW model. Indeed, the number of waypoints
monotonic trend due to the feedback loop that might fail and
influences how fast the user is moving and the type of motion pat-
needs to recover from previous wrong decisions, as depicted in
tern, i.e., few waypoints lead to a quasi-linear trajectory whereas
Figs. 11,12. Finally, we show in Fig. 13 the variation of the center
a huge number of waypoints drives the user to change the motion
of the UAV trajectory for both static users and 4m/s user speed
directions quite often. We let this parameter range from 100 to
scenarios. While the static user scenario allows the UAV to move
1900, being the network trained on a dataset of trajectories with
and converge exactly on the user position, a nomadic user may
100 waypoints.
drive the UAV towards different locations but, still, reducing the
1e2 trajectory radius and increasing the localization accuracy.
1
2 No rubble Rubble
Error [m]

7.4 Proof of Concept Experimental Results

SI

1
Hereafter, we test SARDO in a field-trial taking real measurements
No rubble Rubble
0100 700 1300 1900 0100 700 1300 1900
with the prototype described in Section 6 in a rural area.
Number of SLAW waypoints Number of SLAW waypoints 1 1e2
static 1 m/s 2 m/s
Error [m]

115
Fig. 14. 2D CNN Localization error Fig. 15. 2D CNN Similarity index 1 88 98 98
PDF

85 84
with different channel conditions. with different channel conditions. 60 70 72

Figs. 14, 15 show the distribution of the localization error and

0 200 400 600 100 120 140
the SI considering rubble and no rubble propagation scenarios, Measured distance [m] UAV Altitude [m]
respectively. Specifically, the 2D CNN is only trained on a
generated dataset without the rubble effect. For the no-rubble Fig. 16. Distribution of two sets of Fig. 17. 2D CNN performances with
scenario, we show that the median accuracy is quite stable around real distance measurements. different UAV altitudes and user
31 m (regardless of the increasing number of waypoints) with speeds.

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TMC.2021.3051273, IEEE
Transactions on Mobile Computing
11

1e2 2 1e2

Total time [min]

Battery Impact
static 1 m/s 2 m/s Error Time Baseline SARDO
4.30 4.19 1
Error [m]
1 4

Error [m]
Error = 82.6 m 3.46
3.01
Radius = 100 m
1
Time = 1 min
Error = 51.1 m 2
Radius = 50 m
0.16 0.09 0.08 0.05
0 30 50 100 150 0
Time = 2 min
1 3 5 7 0 0 30 50 100
Number of UAV revolutions UAV Altitude [m] UAV Altitude [m]
Fig. 18. Time evolution of SARDO Fig. 19. Localization of a single user Fig. 20. SARDO’s localization error Fig. 21. SARDO’s impact on UAV
localiz. error for different user with SARDO in a real scenario. and time for different UAV altitudes. battery life in a real scenario.
speeds.
the authors boil down the complexity of a 3D localization system
Dataset and Neural Network. First, we show in Fig. 16
by means of a vector rotation and mobile beacon node, respec-
the distribution of a dataset containing SARDO’s distance mea-
tively, whereas visible light communication is exploited in [36],
surements for a victim in two different locations. As it can be
[37]. Recently, sensor networks have been identified as the main
observed, the obtained values (depicted as bar plots) are distributed
application for more accurate localization systems. In [38], [39],
as Gaussian variables (solid lines) with different variance. In
the authors propose a novel mechanism based on the trilateration
particular, the higher the actual distance, the higher the variance of
solution to identify and localize moving objects exploiting coop-
the distribution. In Fig. 17, we showcase the 2D CNN performance
eration between different anchor nodes. The concept of pseudo-
on real measurements collected at different UAV altitudes and
triangulation is proposed in [40] as a walk-location algorithm.
user speed values. As expected, the localization error tends to
Lastly, [41] provides a survey on localization methods applied to
increase for larger speed values and higher altitudes. As in the
different mobile network generation deployments.
rubble scenario, we have trained the neural network on a dataset
AI-based localization. The usage of machine learning tech-
generated with a fixed UAV altitude equal to 100m but we get
niques to solve localization problems has been introduced in early
similar error values for different UAV altitudes.
works like [42], where the authors compare different families of
Localization error. Hereafter, we assess the performance
neural networks against the well-known Kalman filter or in [43],
of the full SARDO prototype. Fig. 18 shows the average error
to process fingerprinting data. More recently, CNNs have inspired
evolution over time for different user speeds. Notably, the error
several works addressing the problem of indoor localization. [44]
curve exhibits a decreasing slope that proves the convergence of
demonstrates a low-cost intrusion detection mechanism built on
our solution. When the user speed gets closer to the UAV speed
COTS hardware able to detect intruders up to 99% accuracy at
(set to 5 m/s) the convergence rate reduces as the UAV trajectory
2% additional energy cost. Along these lines, [45]–[47] tackle the
radius is kept constant (or increases), as explained in Section 5.
localization accuracy in harsh environments, showing that CNNs
In Fig. 19, we report an example of a full localization process
are able to mitigate NLoS effects and obtain sub-meter accuracy
for a moving victim. In particular, the missing person moves
with Ultra Wide Band (UWB) signals. Even though Channel State
along a polygonal chain with a speed of 1m/s while the initial
Information (CSI) and Angle-of-Arrival (AoA) measurements can
UAV trajectory radius is 100m. After the first UAV revolution, the
be leveraged to perform direct localization [10], CNNs have been
radius of the UAV trajectory is set to the minimum value (50m),
used to forecast channel information [48], which could be fed to
thereby significantly reducing the localization error (from 82.6m
any system based on the trilateration technique. Differently from
to 51.1m). In this experiment, SARDO is able to localize the
the above-mentioned literature that exploits several anchor nodes,
victim within 3 minutes.
SARDO makes use a single anchor and achieves time and spatial
Rescue operations. We focus on a static victim showing
diversity by moving the UAV along controlled trajectories.
SARDO’s applicability during emergency situations. In particular,
in Fig. 20 we outline the total time needed to localize the UAV Swarms. Lately, UAVs have been gaining interest
target (upon convergence) for different UAV altitudes and the among researchers thanks to their usage flexibility even though
corresponding average localization error. It is worth noting that efficient UAV deployment procedures involve several technical
at 100m, both metrics are minimized. Such an optimal altitude is challenges [49]. Indeed, UAV trajectories should take into ac-
obtained as a trade-off between the limited aperture illumination count the flight time, energy constraints, ground users’ Quality-
(due to the antenna radiation pattern effect at low altitudes) and the of-Service (QoS) and collision avoidance requiring an iterative
low received power (due to the stronger path loss at high altitudes). complex optimization problem resolution that provides, time by
This trend is confirmed by the cell radius model derived in [28]. time, the next UAV position [50]–[52]. While building an airborne
Battery cost. Fig. 21 shows the relative cost of our solution on network, [53] compares centralized and distributed approaches,
the UAV battery life. As expected, the battery impact of SARDO while [54] sets up a two-tier mesh network to provide connectivity
decreases as the drone altitude increases. In the optimal operation in emergency conditions. UAV swarms deployment is further
point reported in Fig. 20 (100m) the value is ∼ 5%, which is a exacerbated when additional degrees of freedom, such as altitude,
reasonable cost for the added localization functionality. speed and trajectory, are introduced in the optimization process, as
discussed in [29], [50], [55]. Conversely to the above-mentioned
works, which focus on building an emergency UAV-based net-
8 R ELATED W ORK work, [56] introduces a UAV-based system for accurate (2.5 m)
Localization systems. Geo-localization has been exhaustively localization and tracking in indoor scenarios (mostly buildings)
investigated providing reasonable results in the field of positioning based on unconstrained optimization and multi-hop ad-hoc UWB
systems. A mathematical formulation of the user position based connectivity.
on multiple distance measurements is provided in [34], where a Drone-based Commercial Localization Solutions. The cur-
closed-form stochastic position algorithm is described. In [35], rently available commercial solutions, summarized in Table 3,

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TMC.2021.3051273, IEEE
Transactions on Mobile Computing
12

TABLE 3 [6] I. Leontiadis et al., “From cells to streets: Estimating mobile paths with
Relevant cutting-edge positioning solutions cellular-side data,” in Proceedings of the 10th ACM International on
Conference on Emerging Networking Experiments and Technologies, ser.
#BSs or Infrastr. CoNEXT, 2014, pp. 121–132.
Method Accuracy
#Drones agnostic [7] Naraghi-Pour et al., “A novel algorithm for distributed localization in
LTE [57] E-CID 1 5 > 150 m wireless sensor networks,” ACM Trans. Sen. Netw., vol. 11, no. 1, pp.
OTDoA ≥3 5 50 − 200 m 1:1–1:25, Sep. 2014.
A-GNSS 1 5 < 10 m
AERYON Camera (visible, human visual [8] F. Izquierdo et al., “Performance evaluation of a TOA-based trilateration
1 X method to locate terminals in wlan,” in International Symposium on
Skyranger [58] IR, Thermal) perception
DELAIR Aerial Camera (visible, human visual Wireless Pervasive Computing, 2006.
1 X [9] M. Gutierrez-Roig et al., “Active and reactive behaviour in human
Intell. Platform [59] IR, Thermal) perception
SARDO Pseudo-Trilateration 1 4 ≤ 50 m mobility: the influence of attraction points on pedestrians,” Royal Society
Open Science, vol. 3, no. 7, Jul 2016.
either require an up-and-running mobile infrastructure or are [10] M. Comiter and H. T. Kung, “Localization Convolutional Neural Net-
equipped with IR or thermal cameras to perform image-based works Using Angle of Arrival Images,” in 2018 IEEE Global Communi-
cations Conference (GLOBECOM), 2018, pp. 1–7.
localization. Therefore, such solutions do not provide any useful [11] R. Austin, “Control and stability,” in Unmanned Aircraft Systems. John
synthetic localization error figure and rather rely on the operator Wiley & Sons, Ltd, 2010, ch. 10, pp. 155–167.
experience and skill set. Conversely, SARDO is a single-drone [12] K. Lee et al., “SLAW: A New Mobility Model for Human Walks,” in
cellular-based localization system able to localize users without IEEE INFOCOM 2009, Apr 2009, pp. 855–863.
[13] I. Sutskever, O. Vinyals, and Q. V. Le, “Sequence to Sequence Learning
relying on mobile infrastructure or users GNSS information. with Neural Networks,” in Proc. NIPS, Montreal, CA, 2014. [Online].
Available: http://arxiv.org/abs/1409.3215
[14] D. G. Manolakis and V. K. Ingle, Applied Digital Signal Processing:
9 C ONCLUSIONS Theory and Practice. Cambridge University Press, 2011.
[15] G. F. Franklin, J. D. Powell, and M. L. Workman, Digital Control of
Due to the flexible deployment possibilities and capabilities of Dynamic Systems, 2nd ed. Addison-Wesley, 1990.
modern drones, unmanned aerial vehicles (UAVs) are ideal can- [16] Mini-Circuits amplifiers. https://www.minicircuits.com/WebStore/
didates for novel localization systems when victims are sparsely Amplifiers.html.
[17] Gomez-Miguelez et al., “srsLTE: An Open-source Platform for LTE
distributed in large and/or difficult-to-reach areas. Evolution and Experimentation,” in Proceedings of the Tenth ACM
In this context, we presented here SARDO which, to the best of International Workshop on Wireless Network Testbeds, Experimental
our knowledge, is the first cellular-based drone search and rescue Evaluation, and Characterization, ser. WiNTECH ’16. ACM, 2016,
localization system. SARDO localizes missing people (assumed to pp. 25–32.
[18] srsLTE. https://www.softwareradiosystems.com/.
be close to their phones) with an accuracy of few tens of meters. [19] Intel NUC NUC7i7DNBE. https://ark.intel.com/content/www/us/en/ark/
It requires a few minutes per phone to locate them and achieves products/130394/intel-nuc-board-nuc7i7dnbe.html.
this at a low battery cost. [20] NI USRP B210. https://www.ettus.com/all-products/ub210-kit/.
The properties of the SARDO solution can be summarized as [21] DJI MATRICE 600 Pro. https://www.dji.com/de/matrice600.
[22] 3GPP, “Evolved Universal Terrestrial Radio Access (E-UTRA); User
follows: i) drone-based cellular localization solution for disaster Equipment (UE) procedures in idle mode,” 3rd Generation Partnership
scenarios where the mobile infrastructure is out of service and UE Project (3GPP), Technical Specification (TS) 36.304, 01 2019, version
GNSS information is not available, ii) support for localization of 15.2.0. [Online]. Available: https://portal.3gpp.org/desktopmodules/
multiple victims by running SARDO sequentially within a given Specifications/SpecificationDetails.aspx?specificationId=2432
[23] S. F. Mjølsnes and R. F. Olimid, “Easy 4G/LTE IMSI Catchers for Non-
area, iii) AI-driven improvement of the localization accuracy Programmers,” in Computer Network Security, J. Rak, J. Bay, I. Kotenko,
through a feedback control loop and iv ) automated localization L. Popyack, V. Skormin, and K. Szczypiorski, Eds. Cham: Springer
operations given a GNSS-defined search area. International Publishing, 2017, pp. 235–246.
SARDO has been implemented with COTS components and [24] S. Sesia et al., LTE, The UMTS Long Term Evolution: From Theory to
Practice. Wiley Publishing, 2009.
tested in a field-trial on a rural area [1]. Our results proved the [25] A. Chakraborty et al., “SkyRAN: A Self-organizing LTE RAN
feasibility of the solution and provided quantitative results on the in the Sky,” in Proceedings of the 14th International Conference
expected performance in practice. on Emerging Networking EXperiments and Technologies, ser.
CoNEXT ’18. ACM, 2018, pp. 280–292. [Online]. Available:
http://doi.acm.org/10.1145/3281411.3281437
[26] B. Galkin et al., “UAVs as Mobile Infrastructure: Addressing Battery
ACKNOWLEDGMENTS Lifetime,” IEEE Communications Magazine, pp. 2–7, 2019.
[27] “Navigating Regulatory Compliance for UAV Electronics Development,”
The research leading to these results has been supported by the Tech. Rep., 2014, NASA Tech Briefs.
H2020 5G-CARMEN Project (Grant Agreement No. 825012). [28] A. Al-Hourani et al., “Optimal LAP Altitude for Maximum Coverage,”
IEEE Wireless Communications Letters, vol. 3, no. 6, pp. 569–572, Dec
2014.
R EFERENCES [29] R. I. Bor-Yaliniz, A. El-Keyi, and H. Yanikomeroglu, “Efficient 3-D
placement of an aerial base station in next generation cellular networks,”
[1] Anonymous online video of SARDO in action. https://www.youtube. in 2016 IEEE International Conference on Communications (ICC), May
com/watch?v=9v3NNghq3so. 2016, pp. 1–5.
[2] ReliefWeb, “Natural Disasters 2017,” Report from Centre for Research [30] A. DiCarlofelice et al., “Experimental Characterization of Electromag-
on the Epidemiology of Disasters, Sep 2018. netic Propagation Under Rubble of a Historic Town After Disaster,” IEEE
[3] K. Namuduri, “Flying cell towers to the rescue,” IEEE Spectrum, vol. 54, Transactions on Vehicular Technology, vol. 64, no. 6, pp. 2288–2296, Jun
no. 9, pp. 38–43, Sep 2017. 2015.
[4] K.-D. Park and J. Won, “The foliage effect on the height time series [31] R. Jain, et al., “A Quantitative Measure of Fairness and Discrimination
from permanent GPS stations,” Earth, Planets and Space, vol. 62, pp. for Resource Allocation in Shared Computer Systems,” Tech. Rep., 1984,
849–856, 2008. dEC Research Report TR-301.
[5] C. Specht et al., “Comparative analysis of positioning accuracy of [32] F. Chollet et al., “Keras,” https://keras.io, 2015.
GNSS receivers of Samsung Galaxy smartphones in marine dynamic [33] M. Abadi et al., “TensorFlow: Large-Scale Machine Learning on Hetero-
measurement,” Advances in Space Research, 2018. geneous Systems,” https://www.tensorflow.org/, 2015.

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TMC.2021.3051273, IEEE
Transactions on Mobile Computing
13

[34] S. Nardi and M. Pachter, “GPS estimation algorithm using stochastic Antonio Albanese (S’18) received the M.Sc. in
modeling,” in Proceedings of the 37th IEEE Conference on Decision and Telecommunications Engineering from Politec-
Control (Cat. No.98CH36171), vol. 4, Dec 1998, pp. 4498–4502 vol.4. nico di Milano in 2018. Currently, he is pursuing
[35] Junbo Wang et al., “Design of a 3D localization method for searching his Ph.D. in Telematic Engineering at Universi-
survivors after an earthquake based on WSN,” in 2011 3rd International dad Carlos III de Madrid while being appointed
Conference on Awareness Science and Technology (iCAST), Sep 2011, as Research Associate at NEC Laboratories Eu-
pp. 221–226. rope GmbH. His research field covers optimiza-
[36] S. Zhu and X. Zhang, “Enabling high-precision visible light localization tion, machine learning techniques, blockchain
in today’s buildings,” in Proceedings of the 15th Annual International and MEC topics with a particular interest in posi-
Conference on Mobile Systems, Applications, and Services, ser. MobiSys tioning and prototyping.
’17. ACM, 2017, pp. 96–108.
[37] L. Li et al., “Rainbowlight: Towards low cost ambient light positioning
with mobile phones,” in Proceedings of the 24th Annual International
Conference on Mobile Computing and Networking, ser. MobiCom ’18.
ACM, 2018, pp. 445–457.
[38] M. Balaji and S. A. Chaudhry, “A cooperative trilateration technique for
object localization,” in 2018 20th International Conference on Advanced
Communication Technology (ICACT), Feb 2018, pp. 758–763.
[39] Tang, S. and others, “Study of portable infrastructure-free cell phone
detector for disaster relief,” Natural Hazards 86, pp. 453–464, 2017.
[40] I. Kolingerová et al., “The Stochastic Walk Algorithms for Point Location
in Pseudo-triangulations,” Adv. Eng. Softw., vol. 42, no. 8, pp. 577–585,
Aug 2011.
[41] J. A. del Peral-Rosado et al., “Survey of Cellular Mobile Radio Localiza-
tion Methods: From 1G to 5G,” IEEE Communications Surveys Tutorials,
vol. 20, no. 2, pp. 1124–1148, 2018.
[42] A. Shareef et al., “Localization using neural networks in wireless sensor
networks,” in Proceedings of the 1st International Conference on MO- Vincenzo Sciancalepore (S’11–M’15–SM’19)
BILe Wireless MiddleWARE, Operating Systems, and Applications, ser. received his M.Sc. degree in Telecommunica-
MOBILWARE ’08, 2007, pp. 4:1–4:7. tions Engineering and Telematics Engineering in
[43] C. Laoudias et al., “Indoor Localization Using Neural Networks with 2011 and 2012, respectively, whereas in 2015,
Location Fingerprints,” in Artificial Neural Networks – ICANN 2009. he received a double Ph.D. degree. Currently,
Springer Berlin Heidelberg, Sep 2009, pp. 954–963. he is a senior researcher at NEC Laboratories
[44] F. Devoti et al., “PASID: Exploiting Indoor mmWave Deployments Europe in Heidelberg, focusing his activity on
for Passive Intrusion Detection,” in IEEE INFOCOM 2020 - IEEE network virtualization and network slicing chal-
Conference on Computer Communications, 2020, pp. 1479–1488. lenges. He is the Chair of the ComSoc Emerging
[45] T. Van Nguyen et al., “Machine Learning for Wideband Localization,” Technologies Initiative (ETI) on Reconfigurable
IEEE Journal on Selected Areas in Communications, vol. 33, no. 7, pp. Intelligent Surfaces (RIS) and an editor of IEEE
1357–1380, Jul 2015. Transactions on Wireless Communications.
[46] B. Tariq et al., “Neural Networks for Indoor Human Activity Reconstruc-
tions,” IEEE Sensors Journal, vol. PP, pp. 1–1, June 2020.
[47] K. Bregar and M. Mohorčič, “Improving Indoor Localization Using
Convolutional Neural Networks on Computationally Restricted Devices,”
IEEE Access, vol. 6, pp. 17 429–17 441, 2018.
[48] C. Luo et al., “Channel state information prediction for 5G wireless
communications: A deep learning approach,” IEEE Transactions on
Network Science and Engineering, 2018.
[49] K. G. Panda et al., “Design and Deployment of UAV-Aided Post-Disaster
Emergency Network,” IEEE Access, vol. 7, pp. 102 985–102 999, 2019.
[50] Y. Zeng et al., “Wireless communications with unmanned aerial vehicles:
opportunities and challenges,” IEEE Communications Magazine, vol. 54,
no. 5, pp. 36–42, May 2016.
[51] B. Li et al., “UAV Communications for 5G and Beyond: Recent Ad-
vances and Future Trends,” IEEE Internet of Things Journal, vol. 6,
no. 2, pp. 2241–2263, 2019.
[52] H. Shakhatreh et al., “Unmanned Aerial Vehicles (UAVs): A Survey on Xavier Costa-Pérez (M’06–SM’18) is ICREA
Civil Applications and Key Research Challenges,” IEEE Access, vol. 7, Research Professor, Scientific Director at the
pp. 48 572–48 634, 2019. i2Cat Research Center and Head of 5G Net-
[53] H. Zhao et al., “Deployment algorithms for UAV airborne networks works R&D at NEC Laboratories Europe. He has
toward on-demand coverage,” IEEE Journal on Selected Areas in Com- served on the Organizing Committees of sev-
munications, vol. 36, no. 9, pp. 2015–2031, Sep 2018. eral conferences, published papers of high im-
[54] L. Ferranti et al., “HIRO-NET: Self-Organized Robotic Mesh Network- pact and holds tenths of granted patents. Xavier
ing for Internet Sharing in Disaster Scenarios,” in 2019 IEEE 20th In- received his Ph.D. degree in Telecommunica-
ternational Symposium on ”A World of Wireless, Mobile and Multimedia tions from the Polytechnic University of Catalo-
Networks” (WoWMoM), 2019, pp. 1–9. nia (UPC) in Barcelona and was the recipient of
[55] X. Wang et al., “Real-Time Multi-Target Localization from Unmanned a national award for his Ph.D. thesis.
Aerial Vehicles,” MDPI Sensors, vol. 17, p. 33, 12 2016.
[56] A. Dhekne et al., “TrackIO: Tracking first responders inside-out,” in 16th
USENIX Symposium on Networked Systems Design and Implementation
(NSDI 19). USENIX Association, 2019, pp. 751–764.
[57] Y. Liu et al., “Prospective Positioning Architecture and Technologies in
5G Networks,” IEEE Network, vol. 31, no. 6, pp. 115–121, Nov 2017.
[58] AERYON Skyranger. https://www.aeryon.com/skyranger/.
[59] DELAIR Aerial Intelligence Platform. https://delair.aero/
aerial-data-solutions/.

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/