RF-Based Device-Free Counting of People Waiting in Line: A Modular Approach

Abstract

Device-free RF sensing and monitoring of human/crowd behavior has gained great attraction and consensus. Many different systems have been proposed for different applications, e.g. crowd counting, people localization and tracking or activity recognition. This paper focuses on a relatively new application for device-free RF sensing, i.e. counting of people waiting in line. The proposed solution follows a modular approach which splits an indefinitely long and complex waiting line into chunks (modules) of small size. In each module we apply a naïve Bayes classification algorithm to statistical features of RF power measurements over links crossing the queue. We prove through experiments and combinatorial calculus that the proposed approach achieves very promising accuracy.

Full text

IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 71, NO. 10, OCTOBER 2022 10471
RF-Based Device-Free Counting of People
Waiting in Line: A Modular Approach
Mauro De Sanctis , Simone Di Domenico, Davide Fioravanti , Eduardo Ballesteros Abellán ,
Tommaso Rossi , and Ernestina Cianca , Member, IEEE
Abstract—Device-free RF sensing and monitoring of hu-
man/crowd behavior has gained great attraction and consensus.
Many different systems have been proposed for different applica-
tions, e.g. crowd counting, people localization and tracking or activ-
ity recognition. This paper focuses on a relatively new application
for device-free RF sensing, i.e. counting of people waiting in line.
The proposed solution follows a modular approach which splits an
indefinitely long and complex waiting line into chunks (modules)
of small size. In each module we apply a naïve Bayes classifica-
tion algorithm to statistical features of RF power measurements
over links crossing the queue. We prove through experiments and
combinatorial calculus that the proposed approach achieves very
promising accuracy.
Index Terms—Crowd counting, people waiting in line, queue
counting, RF sensing, RSSI, WiFi.
I. INTRODUCTION
THE ability to monitor people behavior in public spaces
is fundamental for both safety and business purposes. In
particular, in the field of transportation, it would help in making
train/bus stations or in airports more secure and comfortable
places, by reducing waiting time and the occurrence of long
queues [1]. As a matter of fact, real-time information on the
length of the queue could be transmitted to an active dashboard
and configured to offer real time alert to management levels,
which could deploy existing staff when there is a shortage
of workers manning the desk. Moreover, historical data could
help with staff deployment planning based on peak hours when
passengers may grow agitated due to excessive queue lengths.
Two main categories of solutions can be applied: device-based
and device-free approaches. In case of device-based systems,
monitored people need to carry an electronic device. Today, most
of the device-based approaches assume that people carry a smart-
phone and use either actively or passively the data transmitted
Manuscript received 12 July 2021; revised 23 November 2021, 26 Febru-
ary 2022, 7 April 2022, and 25 May 2022; accepted 30 May 2022. Date of
publication 13 June 2022; date of current version 17 October 2022. This work
was supported by Regione Lazio through the Project WIFI-SUITE, under Grants
A0375-2020-36625 and CUP E85F21000920002, under the Programme Gruppi
di Ricerca 2020. The review of this article was coordinated by Prof. Daniele
Tar ch i. (Corresponding author: Mauro De Sanctis.)
The authors are with the Department of Electronics Engineering, University
of Rome, 00133 Rome, Italy (e-mail: mauro.de.sanctis@uniroma2.it;
simone.didomenico@uniroma2.it; davidefioravanti90@gmail.com; balles-
teros_eba@hotmail.com; tommaso.rossi@uniroma2.it; cianca@ing.uniroma2.
it).
Digital Object Identifier 10.1109/TVT.2022.3182548
by the smartphone. In case of active device-based systems, the
smartphone has an active APP that triggers the transmission of
data through WiFi or Bluetooth radio signals, which are then
received, and eventually processed, by other devices deployed
in the monitored area [2], [3]. Passive device-based systems
use signaling data transmitted by the radio interfaces of the
smartphone (i.e. WiFi, Bluetooth, LTE) [4], [5]. Device-based
solutions do not require the deployment of a complex infras-
tructure in the monitored area. On the other hand, in some
application scenarios it is not feasible for a person to carry a
device or, in other scenarios, most of the people could be kids
(such as in theme parks); this kind of device-based approach is in
general used for crowd monitoring inside an area but is not very
effective for counting people waiting in a specific line (unless
highly directional antennas are used to capture only the signals
transmitted by devices inside the queue area). Furthermore, the
radio interface should be switched on and this is not always the
case, especially for WiFi and Bluetooth interfaces.
In case of device-free systems, people are not required to
carry any device and a monitoring infrastructure must be de-
ployed in the area of interest. The most common device-free
approach is based on the use of videocameras and image process-
ing algorithms [6], [7]. The videocamera-based infrastructure
could be expensive and it also raises privacy concerns. Other
device-free approaches use sensors at lower cost such as infrared
or ultrasound sensors. Another important class of device-free
approaches belongs to the so-called RF sensing approach which
is based on the use of radio signals transmitted by and received
from transceivers deployed in the monitored area [8]. The idea
is that the presence of people in an environment changes the
propagation channel of any RF signal that propagates through
it and there is a correlation between the activities/behaviour
of people and some metrics calculated on the received sig-
nal [9]–[15]. The changes of the propagation channel are usually
measured through signal power indicators or frequency-domain
transfer function, i.e. Received Signal Strength Indicator (RSSI)
or Channel State Information (CSI) respectively, [16], [17].
Then, the estimation step usually consists in a classification
process requiring a training phase [18]. Although device-free
RF sensing solutions have attracted a lot of interest in the wider
framework of activity recognition and crowd monitoring (e.g.
gesture recognition, people counting, localization, pedestrian
flows estimation), both in wide areas [19] and in limited en-
vironments [20], only few works have applied device-free RF
0018-9545 © 2022 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.
Authorized licensed use limited to: Universita degli Studi di Roma Tor Vergata. Downloaded on December 10,2022 at 09:34:43 UTC from IEEE Xplore. Restrictions apply.

10472 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 71, NO. 10, OCTOBER 2022
sensing to the estimation of the number of people waiting in
line [21], [22].
Summary of the Original Contribution
This paper represents one of the few attempts to extend the
device-free RF sensing approach to the specific application
of estimating the number of people waiting in line, which is
characterized by the following specific challenges:
1) The waiting line may take non-linear shapes, may be
indefinitely long and people density may not be uniform
along the line.
2) The line where people queue may be located inside more
or less large environments with various types of furniture
and the line might be surrounded by people not belonging
to the line. Due to the multipath propagation, received RF
signals are affected from what happens inside the line, but
also outside the line.
Our proposed system approach aims to deal with the men-
tioned challenges while providing flexibility and ease of use. In
summary, this study investigates the following key aspects of
the proposed system approach:
rFast, small and low-cost installation to collect RSSI mea-
surements - the proposed system configuration is based on
the use of retractable belt stanchions with small and low
cost RF transceivers positioned on top of each stanchion.
Such RF transceivers allow to collect RSSI measurements
that can be processed in real time to estimate the number
of people waiting in line. It is worth noting that, for our
experimental setup we selected WiFi routers to perform
the function of RF transceivers, but other RF technologies
such as Bluetooth and ZigBee may be used to extract RSSI
measurements. It has to be underlined that these WiFi
routers are not the ones that may be already present in
the monitored environment, they are deployed for the RF
sensing application. This type of configuration allows fast
deployment and reconfiguration also supporting a flexible
selection of the waiting line shape. Furthermore, since low
cost RF transceivers may be used, system installation for a
long waiting line requiring a large number of devices does
not lead to a high system cost.
rModular approach - the property of modularity is the key
aspect of this work. The counting system splits the waiting
line into square modules delimited by 4 retractable belt
stanchions with a RF transceiver on top of each stanchion
and estimates the number of people inside each module.
Then, the overall number of people waiting in line is
simply the sum of the estimated number of people within
each single module. Specifically, as it will be explained
later, by properly using the RSSI measurements and the
training data for one single module, it is possible to assume
the independence in the estimation performed inside one
module with respect to the adjacent modules. Therefore, it
is possible to perform an accurate estimation of the people
waiting in a line where more modules are consecutively
inserted. It is worth noting that, this modular approach is
very useful to not limit the length of the queue, which
can be extended simply by using additional modules in the
installation phase and by modifying the overall number
of modules defined in the processing software. In fact,
if the classification/estimation process is performed over
the whole waiting line, the data collection phase required
to build the training dataset must be carried out with
a variable number of volunteers up to the number that
completely fill the queue. This may not be feasible for
long queues. The modular approach greatly reduces the
computational complexity of the system as it reduces the
number of classes to be considered in the classification step
as it increases the sensitivity of the system only to people
waiting inside the module. A classification algorithm is
applied to features which provide a statistical character-
ization of the RSSI measurements over a window of W
packets. We demonstrate through experiments that, in this
particular application, the proposed system configuration
and the selected features based on the statistics of RSSI
measurements provide very good estimation accuracy.
rRelaxed assumptions - no strong assumptions are made on
the density and distribution of people along the line. In
other words, our monitoring system works both in case
of low density and high density of people and also in
case of not uniform density of people waiting in line. In
fact, we consider that in some section of the line there
might be more people and in some other parts less people,
depending e.g. on the relationships between people. It
is worth noting that stronger assumptions are made by
previous works. In particular, some previous works assume
people waiting perfectly “in line” one behind another and
with fixed minimum and/or maximum distance from each
other in line, [21], [23], [24] and [22]. Inside each module,
people can stand in any position, they can change their
relative position inside the module and there could be from
0 up to 4 persons inside one module.
rExpression of the average accuracy for any modular count-
ing system - the paper presents an Equation providing the
average counting accuracy over a waiting line including
any number of modules. This Equation is achieved using
combinatorial calculus and the average counting accuracy
of a single module of a M-module counting system.
Furthermore, the paper proposes and compares two different
feature extraction approaches based on statistical analysis of
RSSI sequences. In both cases, specific patterns of the RSSI
values measured on multiple WiFi links are identified and
associated to the number of people waiting in line through a
well-established classification algorithm. The proposed modular
approach assumes that the estimated number of people in a line
containing Mmodules is the sum of the estimated people in
each module. This assumption holds as long as the training
experiment for one single module is performed including the
interference effect caused by people surrounding the considered
module.
The paper is organized as follows. Section II presents a
survey of related works. Section III describes the key idea of the
proposed approach. The experimental setup is discussed in Sec-
tion IV. The feature extraction and selection step is introduced in
Section V. Section VI analyses the experimental results, while
conclusions are discussed in Section VII.
Authorized licensed use limited to: Universita degli Studi di Roma Tor Vergata. Downloaded on December 10,2022 at 09:34:43 UTC from IEEE Xplore. Restrictions apply.

DE SANCTIS et al.: RF-BASED DEVICE-FREE COUNTING OF PEOPLE WAITING IN LINE: A MODULAR APPROACH 10473
II. RELATED WORKS
A generic crowd counting system cannot be used for counting
people waiting in line for two reasons: 1) people waiting in line
may be very close to each other, hence, crowd counting systems
designed to count people spaced from each other may easily fail
to estimate the correct number of people; 2) waiting lines may be
located in large areas where other people not waiting in line are
present and such people, which may be considered as a source
of interference, should not be included in the overall counting.
In this Section, with reference to the wider framework of
crowd counting or crowd behavior recognition, we review
the main works related to crowd queuing behavior such as
counting the number of people in a queue, detecting people in
a queue, estimating the waiting time, etc. Two main approaches
can be identified: a) device-based, in which the user carries a
device (smartphone or another specific device); b) device-free,
when people are not required to carry any device. In both
cases, such approaches could be further classified into active
and passive. In case of active approaches, signals or data are
actively transmitted by the devices (i.e. smartphones in case
of device-based approaches or APs in case of device-free
approaches), hence, they are somehow controlled by the
monitoring system through software/App installed on the
devices for the specific monitoring purpose. In case of passive
approaches, the monitoring system uses opportunistically the
signals transmitted by the devices, e.g. Beacon frames sent from
available APs or WiFi Probe Request packets autonomously
sent by smartphones with active WiFi cards.
A. Device-Based Approaches
These approaches are grouped on the basis of the used tech-
nology.
WiFi/Bluetooth Signals: Many of the device-based ap-
proaches use WiFi or Bluetooth Low Energy (BLE) signals
either actively transmitted, and eventually activated by an APP,
or passively transmitted by a smartphone. In both cases, signals
are received (sniffed in case of passive approach) and analyzed
by the receiver nodes deployed in the monitored area. In sev-
eral works, RSSI is measured by the received signals and the
estimation of the queue parameters (queuing time, entry and
exit time) are performed by analyzing the RSSI fluctuations [3].
In [25] and [26], only one single WiFi monitor is located at
the head of the queue and an active approach is considered. A
similar system, but using a passive approach, is proposed in [27],
where multiple signal sniffers are deployed, one located at the
starting point of a potential queue such as a service counter, and
two located at the left-hand side and right-hand side along the
queue. In [2], a collaborative approach for queuing recognition
has been proposed, based on mobile phones. One of the issues
that arise is related to the power consumption involved both by
the individual sensing step and by the collaborative sensing step
(transmitting data to the neighbor users). It is worth noting that
none of them count the number of people in line but they only
detect when people start to queue up and estimate the queuing
time.
Radio Frequency Identification (RFID): RFID readers, de-
ployed in the monitored area, detect the existence of nearby tags
carried by the user. The queuing parameters can be tracked on
the basis of the number of detected tags and signal strength [28].
Indoor localization technologies: It uses Bluetooth or UWB
signals with positioning methods such as proximity, trilateration,
and fingerprinting. Once the position has been acquired, the
queue entry and exit times can be determined based on the
relationship between a queue process and a spatial location to
estimate the queuing time [29], [30].
B. Device-Free Approaches
Traditional device-free solutions for queuing behaviour mon-
itoring are based on the use of videocameras [6], [7] or other
sensors (infrared and ultrasound) deployed in the area. In the
wide framework of device-free activity recognition and crowd
behavior monitoring, a lot of research has focused on the use of
RF signals transmitted by opportunistic transceivers deployed
in the environments [12]. Different solutions use different radio
signals (WiFi [31], LTE [14], etc.) and different signal mea-
surements such as RSSI [11] or CSI [22], [32]. It is worth
noting that few works based on RF-sensing can be found on
the specific application of queuing behaviour. Moreover, most
of them do not really count the number of people in line, but
they either detect when people are queuing up or they estimate
the expected waiting time. For instance, in [33], a device-free
passive solution is proposed for detecting people passing through
a line where they might queue up. The proposed system consists
of pairs of BLE transceivers deployed repeatedly along the line
to be monitored, transmitting and receiving packets. Then, the
variance and the mean of the recorded RSSI values, while people
in the queue cross the RF links, are used to trigger detection
events that are then combined by an algorithm.
About the specific task of counting people in a queue, it
must be outlined that the existing studies on crowd density
estimation or people counting cannot be directly applied as it
is an application with peculiar characteristics. First of all, when
people queue up, the density of people might be high and this
is a challenging situation for many of the proposed systems,
as outlined in [34] where an Impulse Radio UWB (IR-UWB)
radar approach is proposed. IR-UWB radar systems adequately
distinguish multipaths and count signals reflected by the peo-
ple. However, they count each signal separately suffering from
instability, superposition and obstruction of signals, which limit
the counting performance in congested environments. Moreover,
the queuing line could be located within one big environment
where other people are present and moving, but the counting
system should only count people waiting in line. In other words,
the changes of the radio signals should be affected only by
what happens in one specific and usually small area of the
entire indoor space in which the queue is located. Information
about what happens outside the line where people queue should
be filtered out and not taken into account by the monitoring
system. Therefore, a different type of system deployment with
respect to crowd density estimation or people counting in single
environments is required.
Authorized licensed use limited to: Universita degli Studi di Roma Tor Vergata. Downloaded on December 10,2022 at 09:34:43 UTC from IEEE Xplore. Restrictions apply.

10474 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 71, NO. 10, OCTOBER 2022
A recent interesting solution for counting people in dense
scenarios is presented in [21]. As in [34], the system is based on
IR-UWB radars but they use the hybrid curvelet transform based
features - distance bin based features (CTF-DBF) to face the
challenges of rapid variations between signals and superposed
multipaths of each signal in congested scenarios. They have
assessed the performance of the proposed system in several
experimental scenarios and in particular, in a scenario where
up to 15 people stand in a queue with an average distance
of 10 cm. With a Random Forest classifier, they achieve an
accuracy of 98.7%. For what concerns the set-up, with one single
UWB radar they detect a line of 5 meters. For longer queues,
more radars must be placed. Moreover, the authors have not
considered the case in which the environment around the queue
is not static but there might be interference from other people
moving around. A more complex and dynamic environment
is considered in the system proposed in [22], which is the
only paper, at the best of authors’ knowledge, which uses CSI
measurements from WiFi to count queuing people with a passive
approach. In particular, in [22], in the signal preprocessing part,
they leverage the moving average filter to remove the outliers and
denoising the signal amplitude and restore the phase information
by phase sanitization. To identify the in-line people number and
people behaviour, they use a dual recognition model, i.e. a static
recognition based on deep learning and a dynamic recognition
based on the Fresnel zone model. In the experimental results,
they assume the use of two WiFi transceivers positioned at
a distance of 5 meters. Moreover, even if the people in the
queue are free to perform any activity, their distance, set to
0.6 m, is rather high for some typical queue scenario. More-
over, they only count 4 people in this scenario, achieving an
accuracy of 95%.
In [35], an RSSI based system using Bluetooth is proposed
to detect people crossing portions of a waiting line. As a conse-
quence, they plan to use this information to infer the number of
people in the queue and/or the average waiting time; however,
this final result was not shown. Specifically, they use Bluetooth
transceivers at the sides of the waiting line at a distance of 1 m
measuring the RSSI of the received packets.
This paper proposes a novel device-free RF sensing approach
that uses WiFi signals, statistical features extracted from RSSI
values and a classification step using a conventional algorithm.
One key element of the proposed approach is the modularity
which can greatly increase the counting accuracy when long
queues must be monitored and reduces the complexity and
duration of the training phase when a complex surrounding
environment is present (with presence of other people not in
the queue).
III. PROPOSED APPROACH
In this Section we describe in detail the idea behind our
proposed method for counting the number of people waiting
in line.
In the proposed monitoring system we use RF signals trans-
mitted by transceivers that are deployed in the monitored area.
Small Tx and Rx devices are placed along the line to be
monitored, at a distance of dmeters thus splitting any queue
into squares of d×dm2. The distance dwas selected on the
basis of an empirical evaluation of the space occupied by adults
considering an average shoulder width of 0.45 m. We also
assume that people in some cases may stay very close to each
other (e.g. family members), but we do not limit people to stay
in line close to each other since each module is not constrained
to be filled completely. We also limit to 4 the maximum number
of people within one module because a larger number means a
larger counting error over each module. Therefore, according to
real case of operation, limiting the maximum number of people
within one module to 4 and considering a small additional space
between people, the distance was set to 1 m. It is worth noting
that, the distance dmay be also customized according to the
specific scenario and shape of the queuing line that, in our
installation and possibly in real case installations, is delimited
by retractable belt stanchions.
In this scenario, where the objective is to understand what
happens just between the Tx and the Rx nodes, the changes on the
received signals are captured by statistical metrics calculated on
the RSSI values in a rolling window of Wpackets. In particular,
in this paper we propose and compare two different groups
of statistical features extracted from RSSI measurements, the
estimated p.d.f. and the set of statistical measures, as better
explained in detail in Section V. Then, the estimation of the
number of people is achieved through a classification process
based on a conventional learning approach. Therefore, the mon-
itoring system foresees a training phase that should be exhaustive
of all possible configurations of people in the queue. That would
make the system very complex in case the line to be monitored is
very long and the density of people is high. Therefore, one of the
key elements of the proposed approach is the use of a modular
approach. One module is a rectangular area (in our experimental
setting we assume 1m ×1m area) having WiFi transceivers at
the edges, as it is shown in Fig. 1. This setup is very similar to the
one proposed in [35], but the approach is completely different.
The proposed monitoring system counts the number of people
in each single module independently. In fact, we carried out the
training phase so that what happens in one module is independent
from what happens in the adjacent modules. To this end, for each
class (0, 1, 2, 3, 4 people) we have performed the training phase
including also people not in line and not in the module of interest,
but standing or walking outside the module. Specifically, in the
training and test experiments we included a variable number
of people located outside from the line and/or located outside
from the module but within adjacent modules. Therefore, the
estimation performed on one single module takes into account
both the interference from adjacent modules, thus allowing
us to assume that the estimations from different modules are
independent, and the interference from people outside the queue.
Thanks to this training approach, we may count the number
of people in a module independently from the other modules.
Finally, when more consecutive modules are needed to cover
the length of the line, we sum the estimated number of people
achieved in each single module of the overall waiting line.
It is also important to underline that thanks to the simplifica-
tion introduced by the modular approach, we do not need to make
Authorized licensed use limited to: Universita degli Studi di Roma Tor Vergata. Downloaded on December 10,2022 at 09:34:43 UTC from IEEE Xplore. Restrictions apply.

DE SANCTIS et al.: RF-BASED DEVICE-FREE COUNTING OF PEOPLE WAITING IN LINE: A MODULAR APPROACH 10475
Fig. 1. Breakdown of a line into RF sensing modules and identification of
the used WiFi links used in a single module, i.e. Tk→Rk,Tk+1→Rk,and
Tk→Rk+1.
Fig. 2. Example of a possible queueing scenario where people do not evenly
fill the line.
strong assumptions on the density and distribution of people
along the line. Fig. 2 shows the case of multiple modules each
having a certain number of people inside. The proposed system
works both in case of low people density and high people density
and also in case of non uniform density of people waiting in line:
in some sections of the queue there might be more people than in
others. Moreover, people inside one module can slightly move
and change their relative position inside the module, as it could
happen in a realistic scenario where people are not completely
static. The complexity of taking into account different densities
or different distribution of people in a long unique line is reduced
in case of a modular approach since for one single module the
number of possible different configurations to be considered is
lower with respect to the case of one single long line.
Therefore, the counting algorithm may be summarised as
follows:
1) Extract useful features from RSSI measurements of links
Tk→Rk,Tk+1→Rk, and Tk→Rk+1,k=1,2,...
2) For any module k, estimate the number of people in each
k-th module by applying a classification algorithm.
3) The estimated number of people waiting in line is the sum
of the number of people estimated in each k-th module,
for any k.
IV. SYSTEM SETUP FOR COUNTING OF PEOPLE WAITING
IN LINE
A. System Setup
Experiments have been carried out in an office room of size
5m×6m. Additional experiments have been carried out in a
corridor with maximum width of 3.5 m with the aim to confirm
the results achieved in the office room. As previously explained,
a long line is divided into square modules of 1m ×1m. A
module consists of 2 transmitters, Tk,T
k+1and 2 receivers
Rk,R
k+1, as they are denoted for the k-th module. Transmitters
and receivers are placed at the edges of the module as shown
in Fig. 1. The modular system of our experiment is based on
the use of retractable belt stanchions with small and low cost
RF transceivers positioned on top of each stanchion where each
transceiver is battery powered (see Fig. 3). The height of each
stanchion is 0.9 m.
The RF transceivers employed for the experiments are GL.inet
GL-MT300N-V2, which are low cost and small size (58 mm ×
58 mm ×25 mm) WiFi routers equipped with a dual antenna
2.4 GHz interface supporting IEEE802.11 b/g/n protocols and
running OpenWrt operating system. Custom applications for
RF sensing purposes have been developed and loaded on each
router of the module. Routers Tk,k =1,2,..., shown in Fig. 1,
are configured as WiFi APs and just send out Beacon messages
with a rate of about 65 packets/second. Routers Rk,k =1,2,...,
placed on the other side of the line and in front of Tk,k =
1,2,..., respectively, act as WiFi sniffers and capture Beacon
messages sent by target APs and extract the RSSI measurements
from each Beacon message.
For any k=1,2,...the link Tk→Rkis referred to as “direct
link” since transmitter and receiver are in front of each other,
while links Tk→Rk+1and Tk+1→Rkare so-called “cross
links” since the line of sight path between the devices diagonally
crosses the area of the module. In order not to consider two
times the same link in a multi-module counting system, the link
labelled Tk+1→Rk+1will not be used for counting purposes in
module k; this link is considered as part of module k+1 only.
Furthermore, in order to assume the independence of modules,
Authorized licensed use limited to: Universita degli Studi di Roma Tor Vergata. Downloaded on December 10,2022 at 09:34:43 UTC from IEEE Xplore. Restrictions apply.

10476 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 71, NO. 10, OCTOBER 2022
Fig. 3. Experimental configuration of the modular system using retractable
belt stanchions and battery-powered RF transceivers positioned on top of them.
the same features (i.e. the same links) cannot be reused over
different modules. In fact, if the same subset of features was
used by classifier to estimate the number of people over different
modules, we could not assume that the estimation of the number
of people in one module is independent from the estimation of
the number of people in another module.
RSSI measurements for each listed RF link are sent through
the Ethernet interface to a local server, which formats and stores
data in a non relational databases. Then, a Python module,
running on the local server, retrieves collected data from the
database and processes them to estimate the current number
of people inside each module. The overall number of people
waiting in line is simply the sum of the counts in each module.
B. Data Collection
The experiments have been carried out with the system setup
described in the previous Subsection. Considering the k-th
module, the RSSI measurements have been collected for links
Tk→Rk,Tk→Rk+1,Tk+1→Rk, and for the following 5
classes:
r0 persons inside the module.
r1 person inside the module.
r2 persons inside the module.
r3 persons inside the module.
r4 persons inside the module.
A person waiting in line is considered being inside module
kif most of its body is within the four sides of the k-th square
module shown in Fig. 1. Each phase of the data collection has
been carried out in a office room of size 5m ×6m having two
tables, 6 chairs and three cabinets. It is worth noting that for
each experiment the volunteers have been asked to take different
positions and spatial configurations inside the module during
the collection phase for the training data and test data. This
allows to test the system in multiple scenarios for the same
number of people and to prevent the overfitting problem due
to a specific configuration of the people inside the module. In
fact, the same set of people inside a module may be arranged
in different ways (different positions) and the system should be
able to effectively estimate the number of people independently
from their arrangement. The overfitting problem may come out
when the system fits exactly a specific arrangement while it is
not able to deal with a different arrangement. It is also important
to underline that each experiment has been carried out with
the presence of one or two persons moving outside the module
over each of the 4 sides with the aim to simulate an interfering
contribution coming from people out of the line and/or people
waiting in line but in adjacent modules. The distance between
the module and the interfering people outside of the module was
variable during the experiments (with a minimum distance of
0.5 m). Furthermore, the dataset was collected over multiple days
(specifically 5 days), so that the configuration of the surround-
ing environment changed during the overall dataset collection
(e.g. position of chairs, door opening, position and number of
notebooks, moving path of people outside of the module).
V. F EATURE EXTRACTION,FEATURE SELECTION,AND
CLASSIFICATION ALGORITHM
The overall set of features of the classifier has been extracted
from the time series of RSSI measurements for each link of
the k-th module, i.e. Tk→Rk,Tk→Rk+1,Tk+1→Rk.The
feature extraction process is based on a statistical analysis of the
RSSI time series over a time window of WRSSI measurements.
In each classification step, the sliding window moves forward
by one sample. The window size has been set to W=3,000
corresponding to about 15 seconds of continuous measurements,
as a result of a compromise between estimation accuracy (which
requires a large window size) and estimation delay (which
requires a small window size). However, we have also carried
out an analysis of the average accuracy as a function of the
window size, which is shown in the next Section. The window
size Wis an important parameter of the system which has to
be properly set considering the trade-off between the accuracy
and the acceptable counting delay. The minimum value of W
depends on the speed of movement of people in a queue; we
Authorized licensed use limited to: Universita degli Studi di Roma Tor Vergata. Downloaded on December 10,2022 at 09:34:43 UTC from IEEE Xplore. Restrictions apply.

DE SANCTIS et al.: RF-BASED DEVICE-FREE COUNTING OF PEOPLE WAITING IN LINE: A MODULAR APPROACH 10477
set the minimum window size to W=1000, corresponding to
5 seconds, so as to capture several slow people movements in a
single window. Furthermore, we should be able to handle users
crossing a module and since the crossing time is much lower than
the minimum window size, the effect of crossing is negligible
when analysed in the time scale of a single window. In addition,
the window size should be lower with respect to the time the
queue may be considered stationary and this is a function of
the departure rate of the queue model. However, in most of the
scenarios of interest for counting people waiting in line, the
considered window size is well below the queue stationary time.
A. Feature Extraction
Two alternative statistical approaches have been carried out
to extract features from RSSI measurements.
The first statistical approach uses an estimation of the prob-
ability density function (p.d.f.) of the WRSSI measurements
using the histogram method over Nbbins, where each bin is
taken as feature. Therefore, this approach has Nbfeatures over
each link Tk→Rk,Tk→Rk+1,Tk+1→Rk, for a total of 3Nb
features; the number of bins Nbshould be larger than or equal
to 4 to achieve an acceptable estimation of the p.d.f.
The second statistical approach uses common statistical mea-
sures, such as mean, variance, range, etc., computed on the
WRSSI measurements in each window; the list of features
considered for this second statistical approach is reported in
Table I with an associated ID.
B. Feature Selection and Classification
The classification step follows the modularity of the system,
hence, a single training set was built for a single counting
module, considering that the behavior of a certain number of
people within one module is independent from the ID of the
module. In fact, each module has the same shape and size, and
the propagation conditions provided by different modules are
equivalent, hence, there no factors that may affect a counting
method differently over different modules.
Then, a single classifier based on the naïve Bayes classifica-
tion method is built using the training set for a counting module.
The number of people within each module is estimated using
the features computed for that module and the unique trained
classifier. Finally, the total number of people waiting in line is
the sum of the estimated number of people over all the modules.
In the following, the performance of the two feature extraction
methods are compared in terms of classification accuracy.
Regarding the first approach, we start with a minimum number
of features equal to 12, that is 4 p.d.f. bins per link, and we
increase the number of features; a lower number of features
does not provide an acceptable estimation of the p.d.f.
For what concerns the second approach, we start with a
minimum number of features equal to 2. In order to carry out
a feature selection process we use a ranking list and a cross
correlation matrix. The ranking is evaluated through the mutual
information I(Xj;C)between the feature Xjand the class
feature C; the achieved ranking is reported in Table I. In our
specific case, we need to compute mutual information for a mix
TAB L E I
LIST OF FEATURES OF THE SECOND STATISTICAL APPROACH FOR FEATURE
EXTRACTION TOGETHER WITH THE RANKING EVALUATED THROUGH THE
MUTUAL INFORMATION WITH THE CLASS FEATURE
of discrete and continuous random variables, as Xjis continuous
and Cis discrete. Therefore, for the computation of the mutual
information, we applied the method described in [36], which
relies on a non-parametric method based on entropy estimation
from nearest neighbors distances.
A high mutual information means a high value of the feature
in the prediction of the class label. However, using a group
of features we should also discard redundant features, if they
exist [37]. The features are selected following the ranking order
starting with the feature with the highest rank (ID=16) and then
repeatedly adding a new feature with the highest rank but having
a correlation coefficient with previously selected features lower
than 0.9 (in absolute value); this value is a good trade-off so as to
consider a limited number of good features. The cross correlation
matrix is shown through Table II. As a result, features with ID
13 and 21 are not selected due to high correlation with features
8 and 16 which were selected first.
The average accuracy is computed using both: a) a Single Split
Train/Test (SSTT) validation, where the overall dataset is split
into 50% training data and 50% test data; b) a Repeated Stratified
K-Fold (RSKF) cross-validation with 5 folds and 20 repetitions.
A double method of computation of the classification accuracy
is used with the objective to increase the reliability of the results.
As we will show in the following, the classification accuracy of
Authorized licensed use limited to: Universita degli Studi di Roma Tor Vergata. Downloaded on December 10,2022 at 09:34:43 UTC from IEEE Xplore. Restrictions apply.

10478 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 71, NO. 10, OCTOBER 2022
TAB L E II
ABSOLUTE VALUE OF THE PEARSON CORRELATION COEFFICIENT BETWEEN PAIRS OF FEATURES OF THE SECOND STATISTICAL APPROACH
(IDENTIFIED BY THEIR ID FROM 0TO 23)
the two validation methods is very similar, proving very high
reliability of the results.
In order to compare the two statistical approaches for feature
extraction, we start with the minimum number of features and we
increase the number of features as long as the average accuracy
is increased by at least 0.5%. Finally, the best counting method
is achieved through the second feature extraction approach us-
ing 6 features having ID 16, 8, 11, 19, 1 and 0; this method
has achieved an average accuracy of 0.98 through the SSTT
validation and 0.97 through the RSKF validation. The high
accuracy achieved by this approach proves the effectiveness of
the proposed feature extraction and selection process.
In conclusion, such features allows to count the number of
people for two main reasons:
rStatistical measures of central tendency of RSSI measure-
ments (feature ID 0, 8, 16) allows to detect the presence of
static/non-static people obscuring the line of sight propa-
gation.
rStatistical measures of dispersion of RSSI measurements
(feature ID 1, 11, 19) allows to detect the presence of non-
static people which give rise to a variable link attenuation.
As a result, we estimate the presence and the number of
people within a module through the selected features which
quantify the effect of any small movement of people on the
propagation channel. It is worth noting that persons waiting in
line are considered non-static even when they only perform small
movements of any part of the body.
VI. EXPERIMENTAL RESULTS
In this Section, we analyze in more detail the results of the
counting method for the single module and then we extend the
results for the case of a long queue including several consecutive
modules.
A. Single Module Performance Results
Let us recall that, according to the feature selection method
applied to the two feature extraction approaches, the best
counting method is achieved through the second feature extrac-
tion approach using 6 features having ID 16, 8, 11, 19, 1 and 0. To
this regard, the first feature extraction approach requires a much
larger number of features to achieve similar performance with
respect to the second statistical approach to feature extraction.
As a matter of fact, statistical measures such as mean, variance,
etc. are efficient measures that extract descriptors of the p.d.f.,
hence, they are useful to compress with a single value some
specific information about the p.d.f. Specifically, the selected
statistical measures of RSSI measurements over the links, listed
according to their ranking order, are: mean over Tk→Rk+1
link, mean over Tk+1→Rklink, coefficient of variation over
Tk+1→Rklink, coefficient of variation over Tk→Rk+1link,
standard deviation over Tk→Rklink, mean over Tk→Rk
link. Firstly, we may conclude that the most important links are
the cross-links Tk→Rk+1and Tk+1→Rk, and these links are
not redundant in terms of information for the estimation process.
As a matter of fact, cross-links cover a different line-of-sight
direction of the module. Secondly, the mean of the RSSI is
very much sensitive to the number of people in the module as
the mean of the RSSI changes considerably from obstructed to
unobstructed conditions and this is very efficient to distinguish
between people waiting in line and people outside from the line.
Third, higher order statistics such as skewness and kurtosis do
not provide useful information; the reason is related to the low
stability of these statistics for the same class.
Previous classification results are achieved for a window size
W=3000; the average accuracy as a function of the window
size Wis shown in Fig. 4. As expected, the classification
accuracy increases as the window size increases. Furthermore,
Authorized licensed use limited to: Universita degli Studi di Roma Tor Vergata. Downloaded on December 10,2022 at 09:34:43 UTC from IEEE Xplore. Restrictions apply.

DE SANCTIS et al.: RF-BASED DEVICE-FREE COUNTING OF PEOPLE WAITING IN LINE: A MODULAR APPROACH 10479
Fig. 4. Average accuracy of the selected classification method as a function of the window size W.
TABLE III
CONFUSION MATR I X AS A RESULTOFTHECLASSIFICATION PROCESS TH E
SECOND STATISTICAL APPROACH TO FEATURE EXTRACTION
AND WITH 6FEATURES (SSTT VALIDATION)
Fig. 4 shows very similar results between the SSTT valida-
tion method and the RSKF validation method; this proves the
reliability of the achieved results in terms of classification
accuracy.
Given Ncclasses, where the actual class is denoted by cand
the predicted class is denoted by ˜c, we may evaluate the perfor-
mance of the classification step for a single module through
the confusion matrix, which is a Nc×Ncmatrix Pwhere
the element pi,j =P(˜c=j|c=i)represents the probability of
predicting class label jgiven the actual class label is i.The
confusion matrix for the selected estimation method is shown in
Table III. The analysis of confusion matrix allows to state that
counting a medium number of people (1 or 2) is more error-prone
than counting the largest or the lowest number of people (0 or 4).
B. Discussion About the Overfitting Problem
The very good results in terms of classification accuracy for a
single module counting system reported by the confusion matrix
in Table III may suggest that the learning process is affected by
overfitting, that is the prediction model is too strongly tailored
to the particularities of the training set and generalizes poorly to
new unknown data.
In order to prevent the overfitting phenomenon, we applied
the following techniques:
1) The number of features has been upper-bounded to limit
the complexity of the prediction model.
2) The selected naïve Bayes classifier is less prone to the
overfitting problem with respect to other classifiers be-
cause of its low complexity.
3) The dataset has been collected over multiple days for each
class. As a consequence, the surrounding environment
changed during the data collection over different days,
hence, this allows not to tailor the classifier to one partic-
ular data collection condition.
4) We estimated the classification accuracy using both SSTT
validation and RSKF validation. The dataset splitting
into training set and test set is 50%-50% for the SSTT
method. The accuracy achieved through the two methods
are largely coherent (see e.g. Fig. 4).
Consequently, while we may state that the good results are
not achieved through overfitting, we may state that the reasons
for such good results are the following:
rThe window size is large and, as shown in Fig. 4, the
accuracy largely depends on window size as the vari-
ance of statistical features for a single class decreases
as the number of measurements increases. This relatively
large window size, when measured in terms of time is
equal to 15 seconds, which is anyway acceptable for most
applications.
rThe size of a module is very small, hence, the received RF
signals are very sensitive to the changes of the propagation
environment inside the module.
Authorized licensed use limited to: Universita degli Studi di Roma Tor Vergata. Downloaded on December 10,2022 at 09:34:43 UTC from IEEE Xplore. Restrictions apply.

10480 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 71, NO. 10, OCTOBER 2022
rAs a consequence of small size of the module, also the
number of classes is low, i.e. 5 classes for a number of
people in one module ranging from 0 to 4.
C. Discussion About the Interference From People
Out-of-The-Modules
We recall that the classifier was trained taking into account the
presence of people in the surrounding of the module (interfering
people), both in adjacent modules or outside of the queue. As
a consequence of the concept of classification, the sensitivity
of the classifier is automatically tuned so that the classifier is
sensitive to people within the module but not to people outside
of the module. This is possible thanks to the different level of
impact of people not in the module with respect to people within
the module. In fact, people within the module are very close to
the transmitter and receiver and are also very close to the line
joining the transmitter and the receiver, while interfering people
not in the module can not be as close to the line joining the
transmitter and the receiver as people within the module.
The performance of the proposed solution in a real scenario, in
comparison with the performance achieved in the experimental
scenario, depends on the accuracy paid on the reproduction of
possible sources of error.
The principle of this system is based on the evaluation of
the extent of signal variation and link shadowing produced by
people waiting in line. Nevertheless, the system is sensitive to
other sources of signal variation which are not connected to
people waiting in line, hence, such other sources should be
considered as sources of error. In particular, as RF waves are not
constrained to propagate within the borders of the waiting line,
people moving in the areas out of the line, and specifically out
of the considered module, generate additional sources of signal
variations which, for our scope, can be considered as unwanted
sources of interference/error. However, people moving in the
area out of the considered module generate a level of signal
variation which is lower with respect to the level of signal
variation generated by people inside the module. This behavior
stems from the fact that people out of the considered module may
only generate an interfering effect to a subset of the three links
of the module or at least with a partial effect in terms of signal
variation. On the contrary, people within the module affect each
of the three links of the module and with a larger effect in terms
of signal variation.
In order to take into account this effect and prove the ro-
bustness of the proposed system to this kind of interference
coming from people out of the line and people waiting in
line but in a different/adjacent module, in any experiment we
considered a number of people from 0 to 4 within the consid-
ered module and also a variable number of people from 0 to
3 were present and moving around the considered module at
different distances. Therefore, the performance results in terms
of accuracy provided for a single module take into account the
effect of interference associated to people out of the module,
both waiting in different/adjacent modules and moving out of
the waiting line. As a consequence, in the following Section
we may extend the performance results in terms of accuracy to
the case of multiple modules considering the interference effect
from adjacent modules is intrinsically taken into account by the
confusion matrix of Table III
D. M-Modules Performance Results
In this Section we develop a method based on combinatorial
calculus for estimating the average counting accuracy achieved
by a M-module system, given the average counting accuracy
achieved by one single module.
Let us consider a counting system including Mmodules and
assume that the maximum number of people in a module is 4; the
results presented in this Section can be generalized for modules
with a different number of maximum people. Let nbe the number
of people waiting in line, ˜nthe estimated number of people, and
N=4Mthe maximum number of people waiting in line within
aM-modules counting system. Let us assume independence of
counting estimation over different modules. In order to meet this
assumption, the links for feature extraction must not be reused
over different modules and people inside one module should not
interfere with the counting over a different module. This is the
reason why link Tk+1→Rk+1is not used in module kbutitis
used only in module k+1 and the training phase was carried
out considering interference from people located also in adjacent
modules.
The average accuracy pcis the probability of counting the
correct number of people waiting in line:
pc=
N

n=0
P(n)P(˜n=n|n)(1)
It is worth noting that the modular approach allows to improve
the performance with respect to one single system that monitor
the whole queue for two reasons. First of all, each counting
system operates on a maximum number of people that is lower
than in case of one single system for the whole queue (in our case,
the maximum number of people in one module is 4) and hence,
the accuracy of the estimation in each single module is higher.
Moreover, given a certain accuracy, the errors can be assumed
randomly distributed over the different modules. Therefore, it
is statistically probable that in one module the system counts
one person more than the actual people inside the module and
in another module the system counts one person less. These two
errors cancel out. In case of one single monitoring system for
the whole queue, this cancelling effect does not occur.
Let pk
i,j be the elements of the confusion matrix for the k-th
module where index irefers to the actual number of people and
index jrefers to the estimated number of people. We assume that
pk
i,j =pl
i,j =pi,j ,∀k, l ∈N;pi,j is the element at the (i+1)-th
row and (j+1)-th column of the confusion matrix for 0 ≤i≤
4,0≤j≤4.
For a M-modules counting system (M≥2), the average
accuracy pc(M)can be computed as follows:
pc(M)= 1
4M+1
4M

n=0
B4(n, M )
C4(n, M )(2)
Authorized licensed use limited to: Universita degli Studi di Roma Tor Vergata. Downloaded on December 10,2022 at 09:34:43 UTC from IEEE Xplore. Restrictions apply.

DE SANCTIS et al.: RF-BASED DEVICE-FREE COUNTING OF PEOPLE WAITING IN LINE: A MODULAR APPROACH 10481
TAB L E IV
A
VERAGE ACCURACY pc(M)FOR A M-MODULES COUNTING SYSTEM,
COUNTING UP TO NPEOPLE WAITING IN LINE
where B4(n, M )is the probability of correct counting of n
people distributed over Mmodules of size 4 considering all
possible distributions of people, and C4(n, M )is equal to the
number of weak compositions of an integer ninto Mparts such
that each part is at most equal to 4. The term C4(n, M )is also
known as the Freund restricted occupancy model [38], [39]. The
computation of B4(n, M )and C4(n, M)are provided in (3)
shown at the bottom of this page, and (4) where p(i, j)=pi,j
are the elements of the confusion matrix P.
The average accuracy according to (2) may be computed
through a computer program using a recursive function; the
average accuracy pc(M)as a function of the number of modules
Mis shown in Table IV for M≤6. As expected, the average
accuracy in estimating the number of people decreases as the
number of modules Mincreases, that is the length of the line
increases. However, Table IV shows that the average accuracy
decreases rather slowly as the number of modules and the maxi-
mum number people waiting in line increases: 98% of accuracy
for counting up to 4 people and 93% of accuracy for counting
up to 24 people, and so on.
C4(n, M )=
n/5

k=0
(−1)kM
kn+M−1−5k
n−5k(4)
E. Comparative Evaluation of the Proposed Approach
Two different types of comparative analysis can be carried out:
1) comparison with systems using a completely different setup
or RF technology (different number of devices and different
technology); 2) comparison with systems using a different set
of features over the same experimental setup (same number of
devices and same configuration).
For what concerns the comparison with systems using a com-
pletely different setup or RF technology, Table V summarizes
the performance and the characteristics, such as the used RF
technology, number of devices, maximum number of people in
line, interference from people in the area but not waiting in line,
of relevant works on device-free RF systems for counting people
waiting in lines. In particular, we have considered the systems
that count the total number of people waiting in line at a par-
ticular instant and periodically refresh the counting estimation,
as they are more similar to our proposed system. It is worth
noting that systems performance of such device-free RF sensing
systems is influenced by several aspects of the experimental
setup: signal bandwidth, carrier frequency, distance between
transmitter and receiver, number of transmitters and receivers,
maximum number of people, room size, etc. Therefore, a fair
comparison of literature works is not possible. However, this
Section discusses the main characteristics and performance of
the mentioned works, outlining limits or advantages with respect
to our proposed system.
In [21], an IR-UWB radar approach is proposed to adequately
distinguish multipaths and count signals reflected by the people.
With a Random Forest classifier, they achieved an accuracy of
98.7% to count up to 15 people in a line of 5 meters. However,
they did not consider the effect of interference from people in
the environment which are not in line. Furthermore, they assume
that people cannot stand side by side, but they can only stand
one behind the other.
In [23], IR-UWB radar data is acquired by an NVA-R661 radar
module with a center frequency of 6.8 GHz and a 3 m detection
range. Three basic steps are applied to clean the raw input data
to obtain refined data for additional operations to determine
the number of people: 1) removal of the DC component, 2)
band pass filtering, 3) removal of the clutter signals. Finally,
a convolutional neural network is employed to estimate the
number of people waiting in line. The average accuracy reaches
99.88% for counting up to 3 people.
In [24] a device-free queuing head count detection using
commodity WiFi devices is proposed. This is the first example
of WiFi-based counting of people waiting in line which exploits
the CSI amplitude extracted form an Intel 5300 WiFi card.
In particular, this work exploits features computed through
the kurtosis and the skewness of CSI vectors over different
B4(n, M )=
min{4,n}

i1=max{0,n−4(M−1)}
min{4,n}

j1=max{0,n−4(M−1)}
min{4,n−i1}

i2=max{0,n−4(M−2)−i1}
min{4,n−j1}

j2=max{0,n−4(M−2)−j1}
...
...
min{4,n−k−1
l=1il}

ik=max{0,n−4(M−k)−k−1
l=1il}
min{4,n−k−1
l=1jl}

jk=max{0,n−4(M−k)−k−1
l=1jl}
...
...
min{4,n−M−2
l=1il}

iM−1=max{0,n−4−M−2
l=1il}
min{4,n−M−2
l=1jl}

jM−1=max{0,n−4−M−2
l=1jl}
pn−
M−1

l=1
il,n−
M−1

l=1
jlM−1

q=1
p(iq,j
q)(3)
Authorized licensed use limited to: Universita degli Studi di Roma Tor Vergata. Downloaded on December 10,2022 at 09:34:43 UTC from IEEE Xplore. Restrictions apply.

10482 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 71, NO. 10, OCTOBER 2022
TAB L E V
RELEVANT WORKS ON DEVICE-FREE RF COUNTING OF PEOPLE WAITING IN LINE
measurements and applies a Support Vector Machine (SVM)
classification algorithm. Experiments are carried out under the
following limitations: the distance between the first person and
the receiver was 1 m, the distance between persons was 0.6 m,
and the maximum number of queuing people was 4.
In [22], a WiFi-based scheme to count queuing people is
proposed using both CSI amplitude and phase and applying a
deep learning neural network. An average accuracy of 95% is
achieved using a deep learning network for counting up to 4
people (in total). However, as in [24], people are constrained
to stay at least 0.6 m apart which seems quite limiting and the
distance from the first person in the queue and the receiver was
kept to 1 m.
With respect to these works, we proposed a modular approach
exploiting RSSI measurements which are easily extracted from
any low cost and small size WiFi transceivers. First, we did not
assume any limitation to the distance between people (minimum
or maximum distance) and to the maximum number of people
waiting in line (results provided in Table IV may be extended
beyond 24 people). Second, we carried out experiments consid-
ering also the interfering effect from people around and far from
the module. The achieved average accuracy of our proposed
system was 98% for counting up to 4 people and 93% for
counting up to 24 people. As shown in Table IV, the accuracy
decreases very slowly as the number Mof modules increases
and this is a consequence of the error canceling effect of the
modular approach as discussed in Section VI-D.
It is worth outlining that among the counting systems based
on a different technology, there are the ones based on ultrasound
sensors [40], [41]. However, a direct performance comparison
is not possible as at the best of Authors’ knowledge, there are
no ultrasound systems specifically addressing the scenario of
counting the number of people waiting in line, which have the
peculiarities already explained in previous Sections. Moreover,
we do not expect the deployed ultrasound system to have the
same flexibility of our solution neither less cost. It is worth
outlining that our solution use transceivers with a battery and
the same WiFi router is used for sensing and for transmitted
the data to the server that performs the processing. In case of
ultrasound set up, each device should be equipped with a wireless
transmitter to keep the same flexibility.
For what concerns the comparison with systems using a
different set of features over the same experimental setup, we
implicitly carried out a comparison analysis with systems using a
different set of features over the same experimental setup (same
number of devices and same configuration). In fact, in Section V
we tested two of the most common methods of feature extraction
using a statistical characterization of the RSSI measurements.
The first statistical characterization uses an estimation of the
probability density function (p.d.f.) of the WRSSI measure-
ments using the histogram method over Nbbins, where each bin
is taken as feature. The second statistical characterization uses
common statistical measures, such as mean, variance, range,
etc., computed on the WRSSI measurements in each window.
VII. CONCLUSION
This paper presented a novel device-free RF sensing system
for estimating the number of people waiting in line. The pro-
posed method uses WiFi signals transmitted by transceivers de-
ployed along the queue line. RSSI is extracted from the received
signals and statistical features are calculated on a window of W
RSSI values. The chosen window size of 3000 samples corre-
sponds to 15 seconds. Therefore, we update the estimation every
15 seconds, which is a latency compatible with the considered
application scenario. One key feature of the proposed system is
the use of a modular approach that splits an indefinitely long
and complex waiting line into chunks of small size and fixed
shape. The monitoring system is trained over one single module
and the number of people inside each single module is estimated
through a naïve Bayes classifier. Then, the estimation of the total
number of people is achieved by summing up the estimates on
each single module. That can be done as the training phase is
performed in a way to consider the interference in the estimation
of people in one single module caused by the adjacent modules
(front and back). This modular approach can greatly reduce the
complexity of the training phase and improve the accuracy as
in each module the number of classes to be considered can be
kept rather low depending on the size of the module (we have
considered a size of 1m ×1m). Finally, the achieved results also
considered the interference effect caused by people surrounding
the modules but not waiting in line.
The modular system of our experiment is based on the use
of retractable belt stanchions with small and low cost RF
transceivers positioned on top of each stanchion where each
transceiver is battery powered. In particular, there is no need of
cables, hence, even our experimental setup offers a very flexible
installation. Therefore, our experiment represents an example of
feasibility of the proposed system.
REFERENCES
[1] F. Tofigh, M. Amiri, N. Shariati, J. Lipman, and M. Abolhasan, “Crowd
estimation using electromagnetic wave power-level measurements: A
proof of concept,” IEEE Trans. Veh. Technol., vol. 69, no. 1, pp. 784–792,
Jan. 2020.
Authorized licensed use limited to: Universita degli Studi di Roma Tor Vergata. Downloaded on December 10,2022 at 09:34:43 UTC from IEEE Xplore. Restrictions apply.

DE SANCTIS et al.: RF-BASED DEVICE-FREE COUNTING OF PEOPLE WAITING IN LINE: A MODULAR APPROACH 10483
[2] Q. Li, Q. Han, and L. Sun, “Collaborative recognition of queuing behavior
on mobile phones,”IEEE Trans. Mobile Comput., vol. 15, no. 1, pp. 60–73,
Jan. 2016.
[3] H. Shu et al., “Queuing time prediction using WiFi positioning data in an
indoor scenario,” Sensors, vol. 16, 2016, Art. no. 1958.
[4] K. Li et al., “An experimental study for tracking crowd in smart cities,”
IEEE Syst. J., vol. 13, no. 3, pp. 2966–2977, Sep. 2019.
[5] K. Li, C. Yuen, and S. Kanhere, SenseFlow: An Experimental Study
of People Tracking. New York, NY, USA: Association for Computing
Machinery, 2015, pp. 31–34.
[6] P. Kilambi, E. Ribnick, A. Joshi, O. Masoud, and N. Papanikolopoulos,
“Estimating pedestrian counts in groups,”Comput. Vis. Image Understand-
ing, vol. 110, pp. 43–59, Apr. 2008.
[7] V. Parameswaran, V. Shet, and V. Ramesh, “Design and validation of
a system for people queue statistics estimation,” Stud. Comput. Intell.,
vol. 409, pp. 355–373, Jan. 2012.
[8] J. Wang, Y. Zhao, X. Fan, Q. Gao, X. Ma, and H. Wang, “Device-free
identification using intrinsic CSI features,” IEEE Trans. Veh. Technol.,
vol. 67, no. 9, pp. 8571–8581, Sep. 2018.
[9] W. Xi et al., “Electronic frog eye: Counting crowd using WiFi,” in Proc.
IEEE Conf. Comput. Commun., Toronto, Canada, 2014, pp. 361–369.
[10] J. Yang, H. Zou, H. Jiang, and L. Xie, “CareFi: Sedentary behavior
monitoring system via commodity WiFi infrastructures,” IEEE Trans. Veh.
Technol., vol. 67, no. 8, pp. 7620–7629, Aug. 2018.
[11] S. Depatla, A. Muralidharan, and Y. Mostofi, “Occupancy estimation using
only WiFi power measurements,” IEEE J. Sel. Areas Commun., vol. 33,
no. 7, pp. 1381–1393, Jul. 2015.
[12] E. Cianca, M. De Sanctis, and S. Di Domenico, “Radios as sensors,” IEEE
Internet Things J., vol. 4, no. 2, pp. 363–373, Apr. 2017.
[13] M. De Sanctis et al., “WIBECAM: Device free human activity recognition
through WiFi beacon-enabled camera,” in Proc. 2nd Workshop Phys.
Analytics, New York, NY, USA, 2015, pp. 7–12.
[14] S. Di Domenico, M. De Sanctis, E. Cianca, P. Colucci, and G. Bianchi,
“LTE-based passive device-free crowd density estimation,” in Proc. IEEE
Int. Conf. Commun., 2017, pp. 1–6.
[15] J. Wang, Q. Gao, H. Wang, Y. Yu, and M. Jin, “Time-of-flight-based radio
tomography for device free localization,” IEEE Trans. Wireless Commun.,
vol. 12, no. 5, pp. 2355–2365, May 2013.
[16] C. Liu et al., “RSS distribution-based passive localization and its applica-
tion in sensor networks,” IEEE Trans. Wireless Commun., vol. 15, no. 4,
pp. 2883–2895, Apr. 2016.
[17] J. Wang, L. Zhang, Q. Gao, M. Pan, and H. Wang, “Device-free wireless
sensing in complex scenarios using spatial structural information,” IEEE
Trans. Wireless Commun., vol. 17, no. 4, pp. 2432–2442, Apr. 2018.
[18] S. Di Domenico, M. De Sanctis, E. Cianca, F. Giuliano, and G. Bianchi,
“Exploring training options for RF sensing using CSI,” IEEE Commun.
Mag., vol. 56, no. 5, pp. 116–123, May 2018.
[19] J.-H. Choi, J.-E. Kim, and K.-T. Kim, “People counting using IR-UWB
radar sensor in a wide area,” IEEE Internet Things J., vol. 8, no. 7,
pp. 5806–5821, Apr. 2021.
[20] M. Yusuf et al., “Human sensing in reverberant environments: RF-based
occupancy and fall detection in ships,” IEEE Trans. Veh. Technol., vol. 70,
no. 5, pp. 4512–4522, May 2021.
[21] X. Yang, W. Yin, L. Li, and L. Zhang, “Dense people counting using
IR-UWB radar with a hybrid feature extraction method,” IEEE Geosci.
Remote Sens. Lett., vol. 16, no. 1, pp. 30–34, Jan. 2019.
[22] H. Zhang et al., “Que-Fi: A Wi-Fi deep-learning-based queuing people
counting,” IEEE Syst. J., vol. 15, no. 2, pp. 2926–2937, Jun. 2021.
[23] X. Yang, W. Yin, and L. Zhang, “People counting based on CNN using
IR-UWB radar,” in Proc. IEEE/CIC Int. Conf. Commun. China, 2017,
pp. 1–5.
[24] F. Xiao, Z. Guo, Y. Ni, X. Xie, S. Maharjan, and Y. Zhang, “Artificial
intelligence empowered mobile sensing for human flow detection,” IEEE
Netw., vol. 33, no. 1, pp. 78–83, Jan./Feb. 2019.
[25] Y. Wang, J. Yang, H. Liu, Y. Chen, M. Gruteser, and R. P. Martin,
“Measuring human queues using WiFi signals,” in Proc. 19th Annu. Int.
Conf. Mobile Comput. Netw., New York, NY, USA, 2013, pp. 235–238.
[Online]. Available: https://doi.org/10.1145/ 2500423.2504584
[26] F. Wu and G. Solmaz, “Are you in the line? RSSI-based queue detection
in crowds,” in Proc. IEEE Int. Conf. Commun., 2017, pp. 1–7.
[27] Y. Wang, Y. Chen, J. Yang, H. Liul, M. Gruteser, and R. Martin, “A low-cost
Wi-Fi-based solution for measuring human queues,” GetMobile: Mobile
Comput. Commun., vol. 19, pp. 10–13, Jun. 2015.
[28] D. Budi´c, Z. Martinovic, and D. Simunic, “Cash registerlines optimization
system using RFID technology,” in Proc. 37th Int. Conv. Inf. Commun.
Technol., Electron. Microelectronics, 2014, pp. 459–462.
[29] Y. Zhuang, J. Yang, Y. Li, L. Qi, and N. El-Sheimy, “Smartphone-based
indoor localization with bluetooth low energy beacons,” Sensors, vol. 16,
Apr. 2016, Art. no. 596.
[30] R. Nishide, S. Yamamoto, and H. Takada, “Position estimation for people
waiting in line using bluetooth communication,” Jun. 2015.
[31] S. Di Domenico, G. Pecoraro, E. Cianca, and M. De Sanctis, “Trained-
once device-free crowd counting and occupancy estimation using WiFi: A
doppler spectrum based approach,” in Proc. IEEE 12th Int. Conf. Wireless
Mobile Comput., Netw. Commun., 2016, pp. 1–8.
[32] H.Zou, Y. Zhou, J. Yang, and C. Spanos, “Device-free occupancydetection
and crowd counting in smart buildings with WiFi-enabled IoT,” Energy
Buildings, vol. 174, pp. 309–322, Sep. 2018.
[33] F. Brockmann, M. Handte, and P. J. Marrón, “CutiQueue: People count-
ing in waiting lines using bluetooth low energy based passive presence
detection,” in Proc. 14th Int. Conf. Intell. Environments, 2018, pp. 1–8.
[34] J. W. Choi, D. H. Yim, and S. H. Cho, “People counting based on an
IR-UWB radar sensor,” IEEE Sensors J., vol. 17, no. 17, pp. 5717–5727,
Sep. 2017.
[35] F. Brockmann, R. Figura, M. Handte, and P. J. Marrón, “RSSI based passive
detection of persons for waiting lines using bluetooth low energy,” in Proc.
Int. Conf. Embedded Wireless Syst. Netw., 2018, pp. 102–113.
[36] B. C. Ross, “Mutual information between discrete and continuous data
sets,” PLoS One, vol. 9, no. 2, pp. 1–5, Feb. 2014.
[37] M. De Sanctis, I. Bisio, and G. Araniti, “Data mining algorithms for
communication networks control: Concepts, survey and guidelines,” IEEE
Netw., vol. 30, no. 1, pp. 24–29, Jan./Feb. 2016.
[38] J. E. Freund, “Restricted occupancy theory–A generalization of pascal’s
triangle,” Amer. Math. Monthly, vol. 63, no. 1, pp. 20–27, 1956.
[39] J. Ratsaby, “Estimate of the number of restricted integer-partitions,” Ap-
plicable Anal. Discrete Math., vol. 2, no. 2, pp. 222–233, 2008.
[40] A. Dragulinescu, I. Marcu, S. Halunga, and O. Fratu, “Persons counting
and monitoring system based on passive infrared sensors and ultrasonic
sensors (PIRUS),” in Pervasive Computing Paradigms for Mental Health.
Cham, Switzerland: Springer, 2018, pp. 100–106.
[41] A. Lesani, L. Miranda-Moreno, T. Fu, and T. Romancyshyn, “Devel-
opment and testing of ultrasonic-based pedestrian counting system,” in
Proc. 94th Annu. Meeting Transp. Res. Board, Washington, DC, USA,
Jan. 11–15, 2015, vol. 15-5136.
Mauro De Sanctis received the Laurea degree in
telecommunications engineering in 2002 and the
Ph.D. degree in telecommunications and microelec-
tronics engineering in 2006 from the University of
Roma “Tor Vergata” Rome, Italy. Since 2008, he
has been an Assistant Professor with the Depart-
ment of Electronics Engineering, University of Roma
“Tor Vergata’, teaching Information Theory and Data
Science. In April 2017, he received the Associate
Professor habilitation (Italian National Scientific Ha-
bilitation - ASN 2016) from the Italian Ministry of
University and Research for the scientific sector of telecommunications. From
January 2004 to December 2005, he has been involved in the MAGNET (My
personal Adaptive Global NET) European FP6 Integrated Project and in the
SatNEx European network of excellence. From January 2006 to June 2008, he
has been involved in the MAGNET Beyond European FP6 integrated Project as
scientific responsible of WP3/Task3. He has authored or coauthored more than
100 papers on journals and conference proceedings, seven book chapters, one
book, and one patent. His main research interests include wireless terrestrial and
satellite communication networks, data science, and information theory. He is
an Associate Editor for the Signal Processing and Communication in Aerospace
Systems area of the IEEE Aerospace and Electronic Systems Magazineandan
Associate Editor for the Command, Control, and Communications Systems area
of the IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS.
Authorized licensed use limited to: Universita degli Studi di Roma Tor Vergata. Downloaded on December 10,2022 at 09:34:43 UTC from IEEE Xplore. Restrictions apply.

10484 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 71, NO. 10, OCTOBER 2022
Simone Di Domenico received the bachelor’s and
master’s degrees in Internet technology engineering
and the Ph.D. degree in electronic engineering from
the University of Roma “Tor Vergata,” Rome, Italy, in
2012, 2014, and 2018, respectively. He is currently a
Postdoctoral Researcher with the University of Roma
“Tor Vergata”. His main research interests include the
RF device-free human activity recognition and people
counting.
Davide Fioravanti received bachelor’s degree in In-
ternet technology engineering from the University
of Roma “Tor Vergata,” Rome, Italy, in 2019. His
academic path was focused on networking (mostly
network protocols and wireless communications) and
mobile devices. He is currently CTO at Made In
Tomorrow srl.
Eduardo Ballesteros Abellán received the bache-
lor’s degree in industrial engineering, specialized in
automation and electronics, from the Escuela Téc-
nica Superior de Ingenieros Industriales (Universi-
dad Politécnica de Madrid, Spain), in 2018. He is
currently working toward the master’s degree (last
year) in industrial engineering with the University of
Rome “Tor Vergata” Rome, Italy, on an international
mobility programme, working on his master’s thesis
related to people counting based on WiFi RSSI mea-
surements. In 2018, he started the master’s degree in
industrial engineering at the Escuela TécnicaSuperior de Ingenieros Industriales.
Tommaso Rossi is currently an Assistant Professor
with the Department of Electronics Engineering, Uni-
versity of Rome “Tor Vergata,” Rome, Italy, where he
teaches digital signal processing, signals and multi-
media processing and communication. He is currently
a Co-Investigator of the Italian Space Agency Q/V-
band satellite communication experimental campaign
carried out through the Alphasat “Aldo Paraboni”
Payload. He has authored or coauthored more than 90
papers on journals and conference proceedings. His
research interests include space systems, extremely
high frequency satellite and terrestrial telecommunications, satellite and inertial
navigation systems, digital signal processing for radar and telecommunications
applications. He is an Associate Editor for the Space Systems area of the IEEE
TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS.
Ernestina Cianca (Member, IEEE) is currently an
Associate Professor with the Department of Elec-
tronic Engineering of the University of Rome “Tor
Vergata,” Rome, Italy, where she teaches digital com-
munications and ICT infrastructure and applications
(WSN, Smart Grid, ITS). She is currently the Director
of the II Level Master in Engineering and Interna-
tional Space Law in satellite systems for communi-
cation, navigation and sensing. She is currently the
Vice-Director of the Interdepartmental Center CTIF-
Italy. She has worked on wireless access technolo-
gies (CDMA, OFDM) and in particular in the waveform design, optimization
and performance analysis of radio interfaces both for terrestrial and satellite
communications. An important part of her research has focused on the use
of EHF bands (Q/V band, W band) for satellite communications and on the
integration of satellite/terrestrial/HAP (High altitude Platforms) systems. She is
author/co-author of 140 papers in international journals and conferences. Her
main research interests include the use of radio-frequency signals (opportunistic
signals such as WiFi or specifically designed signals) for sensing purposes, and
in particular Device-free RF-based activity recognition/crowd counting/density
estimation and localization, UWB radar imaging ( stroke detection).
Authorized licensed use limited to: Universita degli Studi di Roma Tor Vergata. Downloaded on December 10,2022 at 09:34:43 UTC from IEEE Xplore. Restrictions apply.