Enabling and Emerging Sensing Technologies for Crowd Avoidance in Public Transportation: A Review

,


I. INTRODUCTION
S INCE 2002, the European Commission has promoted across Europe and beyond a campaign supporting the use of public transportation (PT) systems as "a safe, efficient, affordable, and low-emission mobility solution for everyone" [1].To cope particularly with PT limitations in urban areas, a key role is expected to be played by intelligent transportation to passengers and for medium-to-long term management and monitoring of the service.
With the advent of Internet of Things (IoT) technologies, fine-grained and real-time passenger data collection is becoming a feasible task, especially in smart cities [7].Indeed, the pervasive use of mobile and portable devices, equipped with different sensors, allows one to gather huge quantities of data in urban scenarios [8,9,10,11], which can be used for different applications and tasks [12].In particular, the fifth generation (5G) of cellular networks is potentially able to support massive IoT connections [13], where billions of smart devices are connected to the Internet and can be easily located and tracked; these features will be further extended by forthcoming sixth generation (6G) networks [14].
Recently, some researchers have argued that the quality of service (QoS) perceived by PT users, as well as their travel satisfaction/quality of experience (QoE), are seriously affected by crowding [15,16,17].To cope with this issue, it is required to acquire in real-time reliable and capillary information about the crowding status of PT rail or road vehicles (e.g., buses, trams, and trains) and the related access infrastructures (e.g., bus/tram stops and metro/railway stations).Indeed, it is stated in [18] that "the availability of real-time passenger demand data can significantly improve the performance of control models in case of overcrowding".
Motivated by the previous needs, new modeling, planning and management strategies that collect real-time crowding data and use them to improve QoS/QoE in PT systems are appearing in the literature [17,19,20,21,22], and will be referred in the following as crowd management [23,24].Crowd management systems are composed [24] by crowd analysis/monitoring and crowd control components: the former includes a network of physical sensors aimed at detecting crowds and estimating their parameters, whereas the latter includes prediction, decision making and control strategies aimed at managing the crowd events.Crowd management should not be confused with crowdsourcing/crowdsensing [25].In PT systems, crowdsourcing (also referred to as "sensing by the crowd" approach) is based on reporting activities by which passengers provide their suggestions and feedback, as well as announce problems to a PT company or, even, provide a resource or create a product, e.g., peer-to-peer services.Relying on crowdsourced data, PT companies work together with passengers to solve a problem or jointly plan public transit by finding and developing solutions aligned with PT user preferences.Crowd management is instead based on the different principle of "sensing the crowd", according to which environmental data collected through networks of IoT sensor devices and/or user terminals are shared with PT companies, which analyze such data for forecasting, choosing among several possible alternative options, as well as ensuring robust and safe operation of PT systems.
The problem of crowd management in PT systems has emerged dramatically during coronavirus disease (COVID- 19) pandemic, which have first spurred our interest to this topic (see Section II for a discussion).During the acute outbreak phase, overcrowding of buses and trains needed to be strictly avoided to protect people from contagion, which required emergency measures, such as, e.g., limiting to 50% the service capacity.These measures are generally not sustainable in the long-term, since they shift a significant portion of passengers to private transportation.Moreover, the problem of overcrowding in PT systems must be tackled in a smarter and more structural way, since many experts predict that, in the future, recurrent waves of pandemic outbreaks could become the norm rather than the exception.
Although state-of-the-art AVL/APC/AFC systems collect a large amount of data about vehicles and passengers in ITS systems, often they are not suitable to perform real-time crowd monitoring and control.However, many ICT/IoT sensing technologies for crowd monitoring are already available or will be available soon in smart cities [7].A recent special issue of this journal [26] reports the cutting-edge advances and ICT technologies pertaining to the seamless integration of sensors with the transportation infrastructure.The main focus of the contributions in [26] is on sensing technologies for private transportation systems, oriented to autonomous driving [27,28], intelligent fault detection [29,30], electric charging optimization [31], road condition monitoring [32,33], precise fleet management [34], speed detection and accident avoidance [35,36]: in the same issue, less attention is dedicated to the integration of sensing technologies within PT systems.Moreover, sensing techniques for crowd monitoring are not specifically discussed.

A. Contribution
In this paper, we argue that, although many sensing technologies for crowd monitoring are already available, the diffusion of crowd management techniques in modern PT systems is hindered by the lack of a structured framework and reference architecture.Motivated by this observation, we pursue three main goals in this paper: (i) to present a survey and taxonomy of crowd monitoring technologies for PT systems based on ICT/IoT technologies; (ii) to discuss their adoption into a reference architecture, aimed at integrating state-of-the-art ITSs (already available in many PT systems in urban areas) with the new crowd management functionalities; (iii) to highlight a series of challenges opened up by the proposed reference architecture that need to be investigated, by also indicating what could be the best way to address them.Although some reviews regarding techniques for crowd monitoring inside PT vehicles exist (see [4,37]), to the best of our knowledge, this is the first review addressing sensing technologies for crowd management in all the different segments of the PT systems.In this respect, Table I enlists all the related surveys in crowd management in PT systems, by comparing them with the review paper at hand.
In our vision, the crowd management functionalities are built upon a distributed IoT subsystem, composed by a capillary network of heterogenous active/passive sensors, aimed at monitoring passenger crowding inside buses, trams, and trains, at bus/tram stops, and in railway/metro stations.By means of a communication infrastructure, the acquired measures are transmitted in real-time to a smart subsystem, which performs crowd control functionalities.Crowding information can also In the proposed reference architecture, real-time knowledge of crowding data can be used by PT operators for fast or even proactive adaptation of some service features (e.g., vehicle holding, stop-skipping, overtaking, limited boarding, speed changing, short turning), in order to cope with spatially and/or temporally localized crowding situations, which cannot be tackled by conventional (statistical) tools used in transportation system design, such as, e.g., origin-destination (O-D) flow analysis.The new crowd management functionalities essentially achieve two scopes: (i) to improve the QoS/QoE of passengers, thereby fostering PT system usage; (ii) to allow for safe PT usage during exceptional events like a pandemic outbreak, such as COVID-19.

B. Paper organization
Section II highlights the adverse impact of COVID-19 pandemic outbreak on sustainable mobility and PT systems.In Section III, the main aspects of crowd management in PT systems are discussed.In Section IV, a crowd management reference architecture is proposed.A taxonomy and review of the sensing solutions for crowd management and their application in different PT scenarios is presented in Section V. Innovations and advantages provided by the adoption of the new crowd management functionalities are highlighted in Section VI.The main challenges and gaps related to the introduction of crowd monitoring management in ITS systems are discussed in Section VII.Finally, conclusions are drawn in Section VIII.

II. SUSTAINABLE MOBILITY IN THE COVID-19 ERA
During the last years, the transport sector and mobilityin particular PT systems in urban areas -have been seriously affected by COVID-19 pandemic.A survey [44] carried out in China in 2020 estimated that, as a consequence of the outbreak, the use of private cars will be roughly doubled, increasing from 34% to 66%, whereas the use of public transports (buses/metros) will be more than halved, dropping down from 56% to 24%.Furthermore, due to the lack of trust in PT systems, more than 70% of the surveyed people not owning a car declared their intention to buy a new one, with negative consequences on the environment (landscape and air pollution) in urban areas.Other recent studies (see, e.g., [45,46]) have highlighted that COVID-19 pandemic has seriously discouraged the use of PT systems.
To counteract the shift to private car usage during COVID-19 pandemic, national governments have implemented different strategies.A widespread measure [47] has been to favor the use of individual sustainable mobility and micromobility means, such as bikes, electrical scooters, and segways, by deploying the related infrastructures (bike lanes) or empowering vehicle sharing services, which can shift to this transport mode a certain percentage of short and medium-distance trips.
However, owing to the large number of passengers carried by PT systems in urban areas, it is of utmost importance to adopt measures aimed at safe and reliable PT usage in such a scenario.During the acute phase of the outbreak, severe anti-COVID-19 measures were adopted [48] to minimize the contagion risk, such as back-door boarding, cashless operations, frequent sanitization of vehicles and stations, enforcing social distances, limiting the service capacity, and requiring the passengers to wear face masks.Other anti-COVID-19 measures were applied [48,49] to PT system operations, such as modifying timetables, frequencies, paths, leveraging modal integration, and so on.Unlike other countries, the Land Transport Authority (LTA) of Singapore has made adjustments to train frequencies to reduce crowding on commuters' lines, increasing the peak-period frequency for trains from once every five minutes to once every three minutes [50].Some of these measures, like increasing PT service frequency or introducing extraordinary trips to compensate for the reduced vehicle capacity, are seen by PT companies as effective, but not sustainable in the long term, due to the limited number of drivers and vehicles and the increased operational costs [51].
Generally speaking, the COVID-19 pandemic has pushed towards a critical rethinking of many economical, social, and cultural habits, not only those related to sustainable mobility.A plethora of innovative solutions have been proposed to cope with this new challenge, many of them employing ICT/IoT technologies.In [52], the use of new technologies, such as IoT, unmanned aerial vehicles (UAVs), artificial intelligence (AI), blockchain, and 5G, has been considered for managing the impact of COVID-19 in health applications.In [53,54] a review of technologies for social distancing has been provided, with emphasis on wireless technologies for positioning, including crowd detection and people density estimation.
As far as sustainable mobility is concerned, a review of the PT planning literature can be found in [49] from the perspective of the changes in demand patterns and limited capacity requirements associated with the COVID-19 pandemic crisis.It is evidenced that, besides reducing the service capacity to adhere to physical distancing measures, PT service providers worldwide have resorted to limiting service span in order to reduce operational costs as a consequence of the reduction in catchment area, by canceling certain services or closing some stations.Moreover, it is pointed out the importance to develop and deploy methods that are able to maintain the functionality of PT systems, while minimizing the public health risks; it is suggested that some changes in service provision can be made at the tactical planning phase, by modifying timetables and/or service frequencies.
Some recent studies have explored the possibility to rely on ICT to organize and facilitate human mobility during the pandemic.In [55], the Authors propose to use a machinelearning (ML) based approach to trace daily train travelers in different age cohorts of 16-59 years (i.e., the less vulnerable age-group) and over 60 (i.e., the more vulnerable age-group) in order to recommend certain times and routes for safe traveling.In this work, many ICT technologies, such as WiFi, Radio-Frequency IDentification (RFID), Bluetooth, and Ultra WideBand (UWB), are employed.Using a dataset of the London underground and overground network, different ML algorithms are compared in [55] to properly classify different age group travelers, showing that the Support Vector Machine (SVM) approach performs better to predict the mobility of travelers and achieves high accuracy (more than 80%).
In [56], a comprehensive review on human mobility research using big data is carried out: big data collected thanks to the pervasive use of ICT/IoT can help, indeed, to discover the relationships between human mobility and resource use, thus entailing great opportunities for smart city development.In [57], instead, the Authors pursue the objective of identifying data sources and ML approaches to properly estimate the impact of COVID-19 on human mobility reduction.In particular, the consequences of the pandemic on mobility patterns of urban populations are investigated in [57], by quantifying even the impact of mobility reduction on improving air quality in urban areas.

III. CROWD MANAGEMENT IN PT SYSTEMS
The main architecture of a crowd management system for PT applications is schematically depicted in Fig. 1.A certain number of physical sensors S 1 , S 2 , . . ., S n are used to collect all the relevant information regarding crowd events.After a suitable pre-processing of the collected information (e.g., dimensionality reduction, normalization, interpolation, denoising, and so forth), the output data are processed by the data-fusion system.Data-fusion can conceptually be divided into three layers [58]: sensor fusion, feature-based data fusion, and decision fusion.Sensor fusion includes data classification, object refinement (e.g., spatio-temporal information alignment, correlation, clustering, grouping, state estimation, error elimination, and reduction), positioning, and recognition.The output of the sensor-function layer has a consistent data structure.Data feature fusion helps to reduce the requirements of application services for system storage resources and computing performance, and it can provide additional complete and in-depth features.Decision-level fusion is aimed at detecting crowds and estimating their parameters in realtime.Prediction of crowd events and their manage is obtained by elaborating not only the output data of the decision-fusion layer, but also other information arising from the Internet and historical databases.Control actions are finally implemented through a certain number of actuators A 1 , A 2 , . . ., A m .
A general introduction to crowd management in transportation systems is provided in [24], where some limitations of state-of-the-art ITSs are highlighted, and the potentials of the new approach are discussed, together with a brief introduction to crowd analysis/monitoring techniques.In [16], different aspects of passenger crowding in PT systems are discussed, related to demand, supply, and operations, including effects on route and bus choice, as well as passengers' wellbeing.Table II subsumes some relevant works on crowd management discussed in the following, focusing in particular to those reporting numerical performance achievements in realistic scenarios.The benefits of real-time crowding information (RTCI) dissemination on passenger travel choices have been discussed and assessed by simulations in [17,59,60].In particular, it is shown in [59,60] that providing RTCI at bus stops might help reduce the deleterious "bus bunching" effect [61].In [17] a complete framework for RTCI modeling in PT systems is introduced, which incorporates RTCI in a dynamic path choice model: the new methodology is tested on a simplified model of the urban PT system of Kraków, Poland, showing that RTCI dissemination contributes to a more efficient distribution of passenger loads in the PT network, improving travel confort and reducing waiting time by about 30%.
A field application of crowd management strategies to PT systems is considered in [19,21].The proposed solution, tested in the municipal bus infrastructure of Madrid, Spain, estimates the current number of passengers in a bus by exploiting the properties of the existing WiFi connections (see Section V-H) and incorporates such information in a bus navigation system, which is capable of giving crowd-aware route recommendations.Well before the COVID-19 outbreak, the LTA in Singapore has been collecting AFC data [62] to help identify commuter hotspots, which enables them to better manage bus fleets and commuter demand, achieving remarkable service improvements, such as 92% reduction in the number of bus services with crowding issues, and 3-to 7-minute reduction in the average waiting time on popular bus services [63].It is worth noting that recently also Google started acquiring and disseminating crowding information to the users [64,65,66].

A. Crowd prediction
Estimating and predicting (in space and/or time) some features of a human "crowd" in indoor and outdoor locations is an active research topic, with many applications, including surveillance and security, situation awareness, emergencies, and crowd management [38].In transport applications, the feature of interest is the number of components of the crowd and/or its density.
Crowd prediction algorithms for transport applications can be classified on the basis of the prediction horizon into short-term (less than 60 minutes) and long-term ones [69].Moreover, they can be classified, on the basis of the prediction methodology, into model-based methods or data-driven ones.
Model-based methods include time-series analysis, regression modeling, hidden Markov models [73], and Kalman filtering models [74].Due to highly nonlinear and random nature of crowds, data-driven approaches have recently gained significant attention, including ML and deep learning (DL) techniques.Several works consider ML-based approaches for crowd management (see, for example, [39,75]).In [40] the authors provide a thorough survey on ML techniques for intelligent transportation systems.
Although ML-based approaches can achieve good results in crowd flow prediction, several reasons pushed researchers to adopt DL-based methods, such as, above all, the ability to automatically extract relevant patterns from unstructured and heterogeneous data.Differently from ML ones, indeed, DL approaches do not require manually extracted or handcrafted features, but can automatically extract the relevant features from the raw data collected by the sensors, process them and make the subsequent decision.In [41] the reference structure of a crowd counting technique based on a convolutional neural network (CNN) is reported, whereas in [42] crowd counting techniques are classified, according to the network property, into basic, scale aware models, context aware models, and multitask models.In some cases, simulation tools to predict traffic flow, like BusMezzo [76] or SUMO [77], have been considered as well.Although crowd/traffic prediction and mobility forecasting are considered in several papers, mostly devoted to road congestion management in urban transportation, their application to PT systems is a relatively new topic: some recent studies are [74,78,70,67,68].In [70], a predictive model for bus crowding is proposed and tested on a AVL/APC dataset taken from the Pittsburgh-area bus using ML techniques.In [67] a data-driven approach is considered to perform car-specific, metro, and train crowding prediction, aimed at providing accurate in-vehicle or station RTCI.
Real-time crowd estimation in [67] is based on load sensors (see Section V-B), which are used, together with historical data, to perform crowd prediction, whose accuracy is tested with data gathered on a metro line in Stockholm, Sweden.It is shown that real-time load data significantly improve prediction accuracy (from between 70% to 90% to between 80% to 95%).In [68] a framework for personalized (i.e., user-specific) crowd prediction is proposed, which takes into account not only loading data, but also other parameters that affect user confort, such as seat availability, expected travel time standing, and excess perceived travel time (compared to uncrowded conditions).In [79] a system providing real-time short-term crowd predictions on trains and platforms is proposed, which uses both real-time AFC and O-D historical data.

B. Crowd control
Crowd information can be incorporated in PT design, optimization, and control at different levels.Besides some approaches showing the benefits of RTCI dissemination to the passengers [17,59,60,67], operator-based crowd control in PT systems is still at an early stage of development.Crowding data can be used at the strategic and tactical planning level in several fields, such as increased services, vehicle capacities, networks expansion, headway and timetable optimization (see e.g.[80,71] for solutions applied to the metro/subway scenario).In [81] a sequential heuristic method is introduced for re-scheduling the timetables of demand-responsive public transport modes in near-real time.The approach was tested on data of Stockholm bus service, with reference to a figure of merit called EWT (Excess Waiting Time).Due to the schedule changes, the operational performance of bus services demonstrated a remarkable service-wide EWT improvement (e.g., the service-wide EWT is reduced from 1.13 to 0.75 mins for the line 1).
In [69] a timetable optimization method aimed at reducing the passenger waiting time (PWT) in metro scenarios is proposed, employing a Genetic Algorithm (GA) integrated with the Interior-Point Algorithm (IPA).The proposed method is tested by simulation on data of the Bejing Metro, showing that the PWT under the optimized timetable is reduced at best by 17.18 % in off-peak hour and by 3.22 % in peak hour in comparison with the standard timetable.Moreover, the method reduces the peak number of passengers on platforms by 44.5 % during off-peak hours and by 9.5 % during peak hours.
In many studies, the effects of crowding are represented by an additional in-vehicle travel time multiplier, so-called crowding penalty [16], which is typically estimated from historical data.A recent discussion of crowd-related studies in PT railway systems is provided in [72], which also proposes a control model for joint optimization of train scheduling and passenger routes taking into account in-vehicle crowding on the basis of O-D data.Real-time crowd control techniques are less explored, mainly due to the lack of availability of finegrained crowding data.Common measures to cope with service irregularities in scheduled PT systems are vehicle holding, stop-skipping, overtaking, limited boarding, speed changing and short turning [82,43].All these strategies can further benefit from the availability of real-time crowding information, both inside vehicles and at the access infrastructure (e.g., bus/tram stops and railway/metro stations).

IV. A REFERENCE ARCHITECTURE FOR CROWD MANAGEMENT
From the previous discussion, it is apparent that integration of ICT/IoT sensing technologies into PT systems for crowd management is fragmentary and their potentials are not fully exploited to date.To bring together these technologies in a systematic and common scenario, we introduce a crowd management reference architecture for ITS, whose scope is to integrate/augment the ITS system already available in a urban PT system with the new crowd management functionalities, aimed at smart and proactive control and reduction of passenger crowding.
The key idea is to integrate heterogenous sensing and communication technologies, in dependence on the operation scenario and the ICT infrastructure available in the urban area where the system must be implemented.To achieve such a goal in practice, strong interdisciplinary design skills are needed, including transportation engineering, telecommunications, computer science, electronics, data analysis, and AI.
In our vision, the reference architecture encompasses (Fig. 2) three subsystems: 1) the sensing and actuator subsystem (SASS); 2) the communication subsystem (CSS); 3) the monitoring, prediction and control subsystem (MPCSS).The architecture involves new data flows (marked as green arrows in Fig. 2) for crowd management, in addition to existing data exchanges (red arrows) commonly used in state-of-the-art ITS systems for PT services.The core and most innovative part of the system is the SASS, with particular reference to sensors for crowd monitoring: in Section V, we provide a taxonomy and discussion of the different sensing technologies than can be utilized to this aim.The actuator component mainly encompasses two crowd-related flows of information: (i) toward the users, carried out by audio speakers, variable-message panels, or displays, which are typically already present in the PT system, or can be readily installed at the bus/tram stops as well as in railway/metro stations, and also inside vehicles; (ii) toward the PT operators (drivers and staff), e.g., by means of evolved driver displays that communicate to the drivers decisions regarding PT adjustments (e.g., holding, stopskipping, or even route changing) to be carried out in real-time.It is worthwhile to note that, with the advent of autonomousvehicle PT systems [83] (based on driverless buses or metro trains), the actuators can directly modify the vehicle behavior, without the need for human intervention.
As regards the CSS, its characteristics are strongly dependent on the communications infrastructure available in the urban area.In general, this subsystem might encompass public wireless networks (such as cellular networks) and/or private wired and/or wireless networks owned by the operator, such as, e.g., Global System for Mobile Railway (GSM-R) or Long Term Evolution for Railway (LTE-R).To cope with this heterogeneity, it is envisioned that, at the protocol level stack, the CSS can be readily interfaced with the other subsystems by means of standard or open interfaces and/or using simple adaptation layers.
The MPCSS performs data collection and real-time crowd prediction, possibly employing AI and ML/DL techniques.Based on such predictions, modifications to the transport services can be implemented in real-time (e.g., vehicle holding, stop-skipping, overtaking, limited boarding, speed changing, short turning) as well as at the strategical/tactical planning level (e.g., optimization of timetables and routes).The related control data are sent to service operators (drivers, supervisors, etc.), whereas service information, including RTCI, is sent to the passengers by means of displays and/or mobile transport apps.This information could be notified by the same applications to all the users of the PT, so as to discourage the access to overcrowded stations and/or bus/tram stops, and propose alternative travel solutions.The MPCSS can be strongly integrated (and typically colocated) with the ITS control system of the PT service operator.

V. SENSING TECHNOLOGIES FOR CROWD MANAGEMENT
From the transportation perspective, our taxonomy of sensing technologies for PT systems considers (see also Fig. 2) the following scenarios: 1) train and bus/tram vehicles (indoor scenarios); 2) railway/metro stations (indoor/outdoor scenarios); 3) bus/tram stops (mainly outdoor scenarios).
The sensing solutions to be adopted in these scenarios belong to the crowd analysis/monitoring [84] family.Detecting and estimating some features of a "crowd" in indoor and outdoor locations is an active research topic, with many applications, including surveillance and security, situational awareness, and emergencies (for a recent review, see [85] and references therein).In PT applications, the feature of interest is the number of components of the crowd (i.e., crowd counting) and/or its density (i.e., crowd density estimation); moreover, tracking of individuals in a crowd could be needed to build O-D matrices, useful for long-term planning.A review of crowd analysis techniques for urban applications, including transportation systems, is provided in [10], where approaches based on different data sources, including information fusion techniques, are discussed and compared.
Sensing technologies for crowd analysis can be classified [86] as visual-based (VB) solutions, based on still or moving images/videos mainly acquired by optical, thermal, or laser cameras, or non-visual based (NVB) ones, which do not rely on images to estimate crowd parameters, but resort to other physical quantities or features that can be related to crowd parameters, such as, e.g., those of radio signals, temperature, or sound.
To perform crowd analysis, VB techniques resort to sophisticated image processing, pattern recognition, or computer vision techniques [84,87,88].Indeed, thanks to recent advances in AI, traditional camera sensors are becoming "smart" and can detect, recognize, and even identify persons.VB technologies perform passive sensing, relying on a network of dedicated sensors, without requiring active cooperation/participation of the users.However, VB data are subject to stringent protection regulations enforced by international data privacy laws, such as, e.g., general data protection regulation (GDPR) [89] in the European Union.
Among NVB technologies (see [90] for a recent review), sensing solutions based on mobile RF devices represent an interesting approach, due to the diffusion of smartphones and other portable/wearable devices, such as, e.g., pedometers, smart watches, or biometrical sensors.This approach to sensing is known as mobile sensing [91], opportunistic sensing [92] or participatory sensing [93].Data collected by means of such device-aided NVB systems can be used not only to count people in a crowd, but also to gather additional information about individuals (e.g., planned routes, O-D flows, passengers using off-peak hours group ticket, and so on).However, a problem inherent to device-aided systems is that they usually require user cooperation/participation. Another important aspect of NVB systems is that they potentially collect sensitive data pertaining to individuals, such as, e.g., daily movements as well as home and work locations.To motivate participation, it is sometimes needed to introduce incentive or reward mechanisms, or apply radical modifications to the procedures to access the PT service, such as an authentication phase to use the service.Although this issue could enhance the aforementioned privacy concerns, it could also be useful as a means to increase the overall safety of the PT systems during a pandemic outbreak, by reducing the risk that infected people can access the system.
When device-aided NVB techniques cannot be used for crowd characterization, due to, e.g., lack of user cooperation and/or security/privacy issues, RF-based non-device aided or device-free approaches can be pursued, which operate by analyzing the propagation channel variations of wireless signals induced by the people present in a given spatial area.In limited cases, NVB sensing techniques relying on physical properties different from those of RF signals (such as audio or ultrasound signals) can also be employed.From the user point of view, information collected through device-free solutions is less critical, as it may not affect users' privacy.
The main sensing technologies for crowd management in PT systems are summarized in Table III and will be discussed in the forthcoming subsections.Table III also provides a preliminary classification of sensing technologies on the basis of their degree of privacy.We defer to Section VII for a more detailed discussion about privacy issues.

A. Infrared sensors
Infrared sensors (IR) are commonly used in traditional APC systems for counting the number of passengers boarding or alighting a vehicle (usually a bus or a tram).Commerciallyavailable solutions employ a couple of IR sensors (acting as a transmitter/receiver) forming a "light barrier" aimed at detecting the passage of people at the input/output gates of the vehicle.In alternative, a single pyrolectric IR (PIR) sensor can be used, which detects the IR radiation emitted by the human body in the wavelength range 2-14 µm.To enhance reliability, combined installation of IR/PIR sensors can also be conceived.Although IR sensors are very frequently used in PT systems, with counting accuracy generally well above 90% [94], their performances worsen when multiple people board through the same door simultaneously [95,96].Moreover, installation could be expensive, since typically more than one barrier per door is needed to detect the passenger flow.

B. Pressure/load sensors
Another solution commonly employed in traditional APC systems counts the number of passengers boarding or alighting buses or trams by means of pressure-sensitive switches ("treadle mats") placed on the vehicle steps, which are activated under the effect of the passenger weight.These solutions are simple, accurate (above 95%) and rugged, ensuring long operational life: installation of multiple treadle mats over different steps allows one to discriminate between infeed/outfeed motion [4].An innovative pressure sensor is the Velostat/Linqstat one, which is a carbon-impregnated conductive polymeric foil that can be used as a low-power inexpensive pressure sensor.Such a sensor has been adopted in [97] to implement a system aimed at monitoring seat occupancy in a bus.
Another application of pressure/load sensors, placed on the ground or on the suspensions, is in weigh-in-motion (WIM) systems, which estimate the number of passengers by the loading of the vehicle detected before and after the stops.Since most of the modern trains are equipped with electronic weighing sensors providing information to the braking system, a WIM solution exploiting such sensors has been proposed and implemented in the Copenaghen metro system [98].In [99] two algorithms for passenger counting are proposed, which estimate the passenger load on the basis of the pressure variations of vehicle air ride suspensions, which are commonly employed in almost all modern transit buses.WIM systems represent a convenient solution to measure crowdedness inside vehicles, even though it can be difficult to infer the actual number of passengers boarding or alighting [4].

C. Optical cameras
Optical cameras are widely used in private and public spaces for surveillance and security, like in closed circuit television (CCTV) systems, and are routinely installed inside PT vehicles and stations to this aim.Since often they can support crowd monitoring functions with firmware/software upgrades, optical cameras are among the most versatile and used techniques for crowd analysis [87,100].Most works on crowd counting and detection based on optical cameras rely on computer-vision technologies [95,101,102], wherein crowds are detected using specific features, e.g., facial recognition or motion tracking, extracted from images/videos.Moreover, the majority of recent works employ AI or ML/DL algorithms [103].
Camera-based solutions can be applied in all transportation scenarios, both indoor and outdoor ones.Video cameras mounted inside road or rail vehicles can be used to estimate the number of passengers and their flow (i.e., whether they are boarding or alighting) [104,95,105,106].An imageprocessing technique based on a modified Hough transform is proposed in [104], aimed at detecting the contour features of heads and estimating accordingly the number of passengers and their flow in a bus.Many recent approaches are based on CNNs, such as the passenger counting system proposed in [107], which also exploits the spatio-temporal properties of video sequences acquired on a PT bus in China, or the solution proposed in [108], where crowd density inside a bus is detected and classified in 5 different levels (from very low to very high), to be displayed by LCD screens installed at the bus stops.A deeply-recursive CNN-based solution is proposed in [109] and tested on a dataset of images taken at the Bus Rapid Transit (BRT) in Bejing, China.A neural-network based crowd density estimation algorithm is described in [110], targeted at underground station platforms, which has been experimentally tested on a sequence of CCTV images acquired at a metro station in Hong Kong.Finally, VB techniques for counting people at bus stops are proposed in [111,112], based on computervision techniques, which process measurements of foreground areas corrected by suitable perspective transformations.
In summary, crowd analysis based on optical sensors is versatile and powerful, but has several limitations: camera sensors, indeed, are expensive, and each camera can only cover a small area, resulting in high deployment costs in complex and/or large environments.In some scenarios, optical camerabased systems do not allow to estimate the number of people with sufficient accuracy, due to possible obstructions, clutter, and poor light/weather conditions.Moreover, crowd analysis techniques based on images require a high computing power.Finally, management of optical camera-based systems might be cumbersome, due to privacy-related restrictions.

D. Thermal cameras
Thermal cameras can detect people in low-light environments, complete darkness, or other challenging conditions, such as smoke-filled and dusty environments [113].Thermal imaging cameras are currently used in some countries to prevent accidents and infrastructure damage in PT systems, detecting, for example, people walking on the tracks or fire events.This technology, however, can be employed also to monitor crowding situations both in indoor and outdoor scenarios.
Some people counters based on thermal cameras are proposed in [114,115,116].Different from optical cameras, people counters based on thermal cameras are less sensitive to the level of ambient lighting or background colour contrasts, but their performance can be adversely affected by heat sources and weather conditions.In addition, real-time image processing can be computationally intensive.The high costs of instrumentation still limit the widespread use of thermography for crowd monitoring.
Recently, hybrid approaches, combining thermal and optical imaging sensors, and intelligent processing (based on DL and CNNs), have been implemented to improve the accuracy and real-time processing of camera-based systems [117,118].In [119], a people counting algorithm is developed, which uses low-resolution thermal images and employs small-size CNNs, being able to run on a limited-memory and low-power platform.

E. LiDAR
Another option to detect and track persons is represented by light detection and ranging (LiDAR) sensors, especially in environments where there are several interacting people [120].A LiDAR sensor is a distance-measuring system that works by illuminating the target with a laser beam and sensing the reflected laser light.There are different implementations of LiDAR sensors based on their coverage area and wavelengths.
Compared to traditional VB approaches, LiDAR is less sensitive to varying lighting conditions and requires, in general, lower data-processing times.Thanks to these features, LiDARbased counting systems are suitable to all transportation scenarios, both indoor and outdoor.
A real-time 2D LiDAR monitoring system for people counting is proposed in [121], which turns out to be useful in monitoring wide areas and dense groups of persons.The solution proposed in [122], instead, employs two LiDAR sensors set at different heights, aimed at detecting people's heads and knees, to improve tracking performance.To increase system accuracy, other solutions use three-dimensional (3D) LiDARs: in [123], for example, a technique for people counting is presented, which works well even if two or three persons pass at the same time.However, an important limitation of 3D LiDAR is represented by its computational cost, which can be even higher than optical camera-based solutions.

F. Acoustic/ultrasound sensors
Acoustic sensor-based approaches perform people counting by relying on audio signals transmitted by smartphones or produced by speaking people [90].A crowd counting solution based on audio tones is presented in [124], leveraging the microphones and speaker phones available in most mobile phones.Effectiveness is proven through several experiments at bus stops or aboard, which show, however, that counting latency can significantly grow as the number of devices increases; as a consequence, this technique may be appropriate only for low-density scenarios.
Despite their simplicity, applicability of purely acousticbased solutions in crowd counting is largely limited by their high sensitivity to environmental noise level.For this reason, hybrid solutions, employing not only acoustic sensors, have been proposed.In [125], for example, a multimodal sensor network is built, which exploits sound intensity in conjunction with additional information sources, like carbon-dioxide level and RF link strength, in order to increase estimation accuracy.Another hybrid solution is presented in [126], which proposes an opportunistic collaborative sensing system, based on acoustic and motion sensors integrated in smartphones.
A different option to count people is represented by acoustic-based solutions using ultrasound sensors, which perform well for indoor spaces and small crowds [90].When the reverberation of the transmitted waves is received, the number of people can be estimated by exploiting information from the receive time or the signal decay.Based on this approach, an ultrasonic sensing technique for estimating the number of people is presented in [127], which exhibits satisfactory performance when the occupancy of the space does not reach its maximum.

G. Device-free RF sensing
Device-free crowd analysis based on RF signals is an emerging technique, which does not require installation of expensive cameras nor it suffers from privacy related concerns.It exploits the impact of the monitored crowd on RF communications to infer information regarding the size and/or density of the crowd, either using traditional radar methodologies (range and Doppler analysis) [128,129,130], mainly with impulse-radio (IR) ultrawideband (UWB) signals [96,131], or by analyzing features extracted by channel quality measurements, such as received signal strength indicator (RSSI) [132,133,134] or channel state information (CSI) [135,136,137].
People counting using RF signals can be used in dimly lit places, and in smoky and dusty environments: hence, it represents an interesting solution for metro/subway stations and/or at bus/tram stops.In [96] a people counting algorithm using a couple of IR-UWB radar sensors is proposed, which was tested in a subway platform in Seoul, South Korea, showing accuracy values higher than 90%.A solution to count the people in a queue, based on RSSI measurements carried out by Bluetooth low-energy (BLE)1 devices, is described in [138]: the system is completely passive and estimates the number of persons in the queue by analyzing the mean and variance of the RSSI values between a BLE beacon and a receiver covering a certain area.
In general, RSSI-based approaches for crowd analysis exhibit good performances in small monitored environments, where the propagation channel variations are dominated by the attenuation caused by the people actually present in the environment.On the other hand, in a rich scattering environment, crowd analysis approaches based on CSI provide a more reliable people counting, but are considerably more complex.Belonging to this class, in [139] and [136] the received WiFi and LTE downlink control signals are processed, respectively, to extract the changes of the propagation channel induced by the presence of different number of people: the related counting algorithms exhibit variable levels of accuracy in different scenarios.
Finally, more sophisticated microwave SAR tomography techniques [56,140] can also be used, which provide specific RF images from which more detailed crowd information can be extracted, by using complex image classification algorithms (i.e., count, distribution, mobility, etc.).Modeling issues induced by the intrinsic near-field scenario (e.g., typical for bus/tram stops) could be overcome using specialized algorithms [141].
In summary, excluding SAR-based techniques, device-free RF sensing is a moderate-complexity crowd analysis technique with low installation costs.Moreover, RF signals can penetrate obstacles to a certain extent and are not affected by weather/illumination.A moderate accuracy (around 80% for RSSI-based techniques, up to 90% and larger for CSIbased ones) can be expected in simple scenarios, but it is questionable whether this approach can be scalable to large crowds, especially in complex propagation scenarios [90].Moreover, counting algorithms based on this approach usually require a site-specific training phase, which complicates practical installation and maintenance.

H. Device-aided RF sensing
In device-aided approaches for crowd analysis, users are expected to carry RF devices, which must be switched on to enable people counting.Since modern smartphones are ubiquitous and, moreover, are equipped with several sensors and multiple RF interfaces, device-aided solutions are commonly used to gather different types of information for many different purposes and applications (see [90] for a recent review).
1) RFID and NFC: RFID-based solutions require that the passengers carry passive RFID tags, whose presence can be detected by a reader.The Authors in [142] propose an APC system for bus vehicles employing commercial EPC Gen2 tags, which are recognized by a reader located in correspondence of the bus gate.A similar short-range communication system is Near Field Communication (NFC) [143] technology, which is currently supported by many modern smartphones and tablets.When a device with NFC functionalities appears in the reader's working range, which can be placed at the station gates or at any other fixed access point, it "wakes up" and sends a signal containing encoded data.Finally, it should be observed that RFID and NFC technologies are used in traditional and emerging AFC systems for electronic ticketing, such as MIFARE contactless cards or mobile-based payment systems [144], which can also be employed for passenger counting.For these applications, privacy is characterized by the ability of unauthorized users to trace RFID and NFC devices using their responses to readers' interrogations.Since RFID and NFC devices are typically not tamper-resistant, an adversary can capture them and expose their secret parameters.
2) Bluetooth: Bluetooth is a consolidated short-range RF technology, supported by almost all smartphones on the market: crowd monitoring algorithms using Bluetooth have been proposed in several papers (see e.g.[145] and references therein).An algorithm based on off-the-shelf Bluetooth hardware for counting bus passengers has been proposed in [146] and tested in the city of Funchal, Portugal.The system consists of a Bluetooth scanner mounted on the bus ceiling, which periodically scans for discoverable Bluetooth devices in its range, and is aimed at discovering O-D relations by postprocessing data and correlating them with information related to bus location and tickets issued by fare machines.
Bluetooth can also represent an efficient solution for crowd counting at bus/tram stops.A crowd analysis solution based on BLE is proposed in [147], where a large population carry BLE proximity tags, acting as beacons, whose presence is sensed by smartphone carried by few volunteers.A reciprocal solution can be used at bus stops, where BLE beacons are installed at the bus stations, and are detected by passenger smartphones in close proximity of the stops. 2ompared to WiFi-based sensing (discussed later), Bluetooth devices are cheaper, less power-hungry and are characterized by increased flexibility and simplified installation.
3) WiFi: Many crowd analysis solutions exploit the characteristics of the IEEE 802.11 (WiFi) protocol, which is widely used by passengers during their trips.The technique adopted in [19,21] estimates the current number of passengers in a bus by counting the number of probe requests, i.e., medium access control (MAC) addresses, sent by WiFi-equipped devices in the vehicle.A similar approach is followed in [149], where a de-randomization mechanism is introduced to counteract software randomization of MAC addresses, recently introduced in many operating systems.
One of the problem inherent to the use of WiFi-based techniques for crowd counting inside vehicles is the ability to distinguish between people outside the vehicle and actual passengers.This issue was tackled in [19,21] by filtering the probes with a sliding window, aimed at removing MAC addresses that were not detected over a longer period of time.In addition, in [149] the received power is also used to discriminate devices that are likely to be outside the bus.The system in [19,21] was able to detect only around 20% of the passengers in a real setting, since several users may have turned off the WiFi interface.Underestimation of the number of passengers is a common problem for these techniques, which can be compensated by a proper calibration of the procedure in each scenario of interest.
The main advantage of WiFi-based techniques is that they allow to track passengers also when they alight the bus, allowing to estimate O-D flows.WiFi-based counting can also be employed in metro/railway stations, since access points are typically available in such scenarios.
4) Cellular: Similarly to WiFi, cellular signals such as LTE and 5G ones can be used for device-aided crowd counting, thanks to their ubiquitous availability and good penetration in indoor environments.Cellular signals could be available in areas where the WiFi coverage is not present, such as bus/tram stops, remote and small railways/metro stations.A cellular-based crowd density estimation method is proposed in [150], which measures the signal strength emitted in uplink by the smartphones of the crowd components, and classifies the crowd density in different levels using DL techniques, with an accuracy of 78% when three levels are considered.
In principle, the position of users in a cellular network can be obtained with satisfactory precision by combining knowledge of the serving base station, RSSI values, and triangulation principles [151], which can be at the basis for large-scale crowd analysis.However, this approach is difficult to be used in real-time, whereas it is more suited for long-term travel demand estimation [152].Real-time passenger counting using cellular data is rarely performed, due to several drawbacks: (i) it requires gathering data from different mobile operators; (ii) it raises significant privacy concerns; (iii) it does not allow one to precisely discriminate passengers from general public in open areas.A breakthrough could be the planned introduction in 5G systems of the millimeter-wave (MMW) band, which will require very small cells: from the viewpoint of crowd analysis, the placement of small cells allows one to more precisely monitor spatially limited areas, like bus/tram stops.

I. Discussion
In this Section, we provide a discussion regarding the applications of the above-mentioned sensing techniques in the different transportation scenarios. 3) Trains/buses/trams: Such vehicular scenarios are characterized by a well delimited indoor space with a limited number of accesses (gates).Traditional APC systems [4,37,149] count the number of passengers inside vehicles on the basis of various onboard sensors, mainly IR or pressure-sensitive ones.The number of passengers can also be estimated by the number of validated/issued tickets, as in AFC systems, which however requires user cooperation and can provide underestimated results in case of diffuse fare evasion.
Although solutions based on sensors installed on the vehicles are simple, they require a large initial investment.Therefore, the diffusion of portable and mobile devices between passengers have pushed many researchers to study solutions based on RF techniques (both device-aided and device-free).Summarizing, as also indicated in Table III, even though more sophisticated VB/NVB sensing technologies may be used as well, their usage is not expected to lead to significant innovations in this scenario, compared to traditional solutions adopted in commercial APC systems.
2) Railway/metro stations: Many crowd monitoring options are available in this scenario, since the access to the stations occurs through a limited number of gates.Morover, CCTV surveillance systems are typically present inside stations and can be used for VB crowd analysis.The access to train platforms is usually governed by turnstiles where tickets/passes must necessarily be validated: thus, IR-based APC or AFC systems could be a reasonable option to count passengers in this scenario.However, this solution only counts people trying to access the platforms, disregarding other people which could walk inside the station for different purposes (e.g., for shopping or leisure).RF-based techniques could suffer from coverage problems, especially within metro/subway stations.
3) Bus/tram stops: This scenario is by far the most difficult to manage, since bus/tram stops are usually located in outdoor spaces, not well delimited by fixed gates.In this case, both VB and NVB sensing technologies can be used, but it is imperative to adopt cost-effective, rugged, and low-power solutions, in order to reduce the maintenance cost.Moreover, solutions that do not require significant infrastructures are preferred, since in many cases the stops are not equipped with shelters and are indicated by simple poles.

VI. MAIN INNOVATIONS AND ADVANTAGES
The crowd management functionalities of the new framework can provide several innovations and advantages that are not present in state-of-the-art ITSs:

Actor Advantages
PT service operator (planner, manager) Having real-time crowding data of the different segments of the PT system allows one to plan services more efficiently and to quickly readapt them to tackle critical situations, localized in space and time PT operators (drivers, supervisors, staff) Knowing in advance crowding situations at the stops or stations allows one to tackle in real-time critical situations (e.g., implementing holding or stopskipping strategies, or following alternative routes)

PT users
Knowing crowding situations allows one to use alternative transportation modes or make the trip in another hours, if not strictly necessary; by the reservation system, if available, the users can access vehicles without unnecessary crowding at the stops or stations Police forces Knowing in real-time and/or predicting possible crowds -potentially dangerous for public health and/or for public order -allows a more timely and targeted intervention Sanitary system Smart reduction of crowding resulting from an agile management of the PT system allows one to reduce the diffusion of infection and prevent further outbreaks

General population
An improved QoE of the PT system entails a reduction of private car usage and of pollution in urban areas, with benefits on energy consumption and climate change 1) Proactive control of station access: In railway/metro applications, on the basis of the knowledge of the number of passengers aboard the arriving trains and the prediction of those alighting at the station, it will be possible to predict the number of accesses to stations/platforms with a low error margin and in real-time, so as to avoid crowding.This number can be communicated to the users (by displays at the station gates or by the mobile transport apps) and can be used by the security operators to filter passengers at the turnstiles.Priority policies can be envisioned, such as taking into account the time already spent in queue, or the trip motivations (e.g., a priority could be assured to health workers, disabled or elder users, law enforcement operators, and teachers/students).
2) Vehicle access reservation: In bus/tram trips, a vehicle access reservation system can be implemented.A sensor at the bus/tram stop detects the user presence and exchanges information with his/her device (i.e., the smartphone), so as to grant him/her the access to board the first arriving vehicle (a virtual queuing system) or putting him/her in an overbooking list (with priority) to allow him/her to board the next one.The application can generate an e-ticket with the access grant (e.g., a QR-code) that can be validated on board at the ticket machine.
3) Crowding information dissemination: Users receive realtime information related to available capacity (in terms of number of seats or in percentage) of the bus/trams/trains in arrival and/or crowding at the stops/stations, so as to avoid unnecessary waiting or crowding, and possibly reschedule their trips.Such RTCI can be provided by means of displays installed in correspondence of the stops or at the entrance of the stations, or by messages/alerts issued by mobile transport apps.
4) Crowd-based route planning: Users can plan their trips on the mobile transport app, by taking into account not only geographical information and traveling times (static data), but also traffic and crowding information about vehicles and stops/stations during the trip (dynamic data).The app may not necessarily suggest the shortest route, but the least crowded one, taking into account also crowding levels measured during the trip.Such a feature not only helps reduce crowds (and the consequent infection risk in case of a pandemic), but also improves the passenger QoE, by distributing more efficiently the load on the transportation network.
5) Crowd-aware real-time control: Some of the typical functions of PT planning and operations can benefit from the availability of the new crowd information.At the strategic and tactical levels, average long-term crowding data can be used to plan the services.At the real-time and operational level, critical situations (service disruptions, unusual crowding) or even random spatio-temporal demand variations can be tackled by implementing crowd-aware rescheduling solutions, such as holding, stop-skipping, or similar ones, which can be communicated to the drivers and staff in real-time, as well as to the passengers via displays and/or alerts.
The main advantages of the crowd management framework are summarized in Table IV.Compared to static methods, like traditional survey-based compilation of O-D matrix flows, more efficient planning and real-time control of the operation of the PT system is allowed.The large amount of generated data can be used by AI and ML/DL algorithms to better understand and plan a series of aspects generally associated with improvements of the quality of life in urban areas and smart cities.
Compared to the other anti-COVID-19 solutions discussed in Section II, the new framework is not aimed at enforcing social distancing measures only.Its scope is wider, since it tries to incorporate crowding information to adaptively optimize the overall performance of the PT system.As a by-product, it also allows to partially recover the drawbacks and inefficiencies of PT systems due to the adoption of rigid social distancing measures during both pandemic and post-pandemic phases.

VII. MAIN CHALLENGES
In this Section, in relation to the proposed reference architecture, we enlighten some of the main challenges to be addressed for the introduction of crowd management functionalities in existing or forthcoming ITS systems.The impact of such challenges on the subsystems of the introduced reference architecture is summarized in Table V.
1) Cooperative sensing: In NVB cooperative sensing approaches, a large number of users must be engaged to obtain significant amounts of data.Several works [153,154] have proposed to introduce incentive mechanisms aimed at stimulating user motivation and encouraging participation.A common solution is to let participants earn credits in exchange of their data.However, such a strategy might present some privacy risks, by exposing sensitive data and linking them to users' identities.Several efforts have been made to propose privacy-preserving cooperative sensing approaches (based, for example, on data anonymization, randomisation, and aggregation).An interesting solution is discussed in [155], where a rewarding platform based on a voucher exchange system is proposed, which decouples mobile crowd sensing instruments from participation incentives.Each voucher is produced as a compensation for user's participation, and it is designed to be fully anonymous and not exclusive.
2) Security: As an ICT-based evolution of the state-of-theart ITSs, the proposed reference architecture strictly relies on ICT infrastructures too, i.e., most of the functions accomplished by the MPCSS subsystem could be implemented using modern cloud technologies.Since every ICT infrastructure can be prone to well-known cyber-attacks, security issues are particularly serious in all portions of our reference architecture, mainly due to the large volume of data exchanged among subsystems.A malicious user, for example, could launch a man-in-the-middle attack to the system, by intercepting and altering the content of the exchanged messages between crowd monitoring tools and vehicles, which could entail catastrophic effects in the decision-making phase, especially when selfdriving vehicles are involved.Furthermore, denial-of-service attacks could be used to saturate the resources of the infrastructure to completely interrupt the decision-making process.
Security issues related to the IoT paradigm are well-known in the literature, but there are some aspects peculiar to ITSs [156].In particular, one of the most limiting factor is represented by the scarce computational resources of many sensing devices used in ITSs, which may render the commonly used countermeasures (i.e., cryptography and authentication schemes) infeasible for providing secure communications.The proposed architecture, indeed, is based on a plethora of heterogeneous sensors and devices, which are required to be inexpensive, low-energy, and, in some cases, have small form-factor.However, energy-constrained sensors are resource-limited in terms of memory, computational capabilities, and communication range.Such constraints greatly limit the use of complex algorithms, useful even to guarantee security and privacy.Therefore, it becomes paramount to adopt energy-aware solutions to prolong the lifetime of the sensing subsystem.In this sense, a promising solution can be ambient backscattering [157], where small passive devices are able to transmit data by reflecting electromagnetic waves transmitted by existing RF transmitters.
3) Data fusion: The availability of large amounts of data acquired by heterogeneous sensors can pose a challenging problem for crowd monitoring, prediction, and control tools.In the proposed architecture, indeed, the MPCSS subsytem must analyze in real-time a large amount of multi-source and multi-modal data, which are characterized by different levels of resolution, accuracy, reliability, and redundancy.Various data fusion (DF) algorithms, aimed at associating, correlating, and combining information from multiple sensors, are commonly adopted to provide accurate and timely decision-making support.Among the commonly adopted approaches (i.e., statistical, probabilistic, and data-driven ones [58]), probabilistic-based methods seem to be more suitable in our scenario.In particular, Bayesian approaches, maximum likelihood methods, and Kalman filter-based DF techniques can be utilised for multi-sensor data fusion [158].VB solutions must be designed to implement data protection principles, collectively referred to as privacy-by-design, set out in international data privacy laws.For crowd counting and crowd density estimation, identification is not necessary.Therefore, for such applications, privacy-by-design can be achieved by suitably choosing resolution and other modifiable factors of VB devices to ensure that no recognizable facial images are captured.In addition to privacy-by-design, some data privacy laws also mandate privacy-by-default.Under such an obligation, personal data collected through VB approaches must be used only for the specific purpose of crowd management.This means that the minimum required amount of personal data should be collected, their processing should be minimized, and their storage and accessibility should be controlled.A viable approach to ensure privacy-by-default consists of turning VB feeds featuring individuals into numbers and heatmaps in such a manner that the data subject is not or no longer identifiable.Within the European Community, such countermeasures allow VB solutions to comply with the GDPR, which is one of the most restrictive legislations in the world in terms of safeguarding citizens' privacy.
Regarding NVB solutions, sensing solutions based on mobile RF devices pose important privacy issues, especially for device-aided crowd monitoring.Apart from the issues regarding participatory sensing, which have been discussed previously, many device-aided RF-based NVB techniques basically perform monitoring of over-the-air beacon signals sent by individuals' smartphones when they connect to a network or install an application.In this case, to extract relevant features for crowding information, processing of devices' IDs is required, e.g., MAC addresses in WiFi networks.Most RFID tags emit unique identifiers when they respond to reader interrogation, even tags that protect data with cryptographic algorithm [159].A similar problem arises for Bluetoothenabled wireless devices [160], as well as for the NFC protocol that is based on the transfer of an ID for anti-collision during the process of contactless reading of transponders [161].Although cryptographic techniques can be used to transform ID data to preserve privacy, thereby complying with the GDPR, they are computationally-intensive and require the generation and maintenance of multiple keys, which also leads to higher energy consumption.Alternatively, devices' IDs can be aggregated or perturbed in such a way that individual data privacy is preserved but, at the same time, useful information for crowding management is not destroyed.In the case of WiFi networks, a viable solution is represented by MAC address randomization [162], which is an available option in different operating systems, such as iOS, Android, and Windows.However, although it is compliant with the GDPR, some crowd counting methods are adversely affected by MAC address randomization.A cutting-edge future research direction might be exploiting the mathematical concept of differential privacy [163], which addresses the challenging goal of learning nothing about an individual while learning useful information about a population.This is in accordance with existing data protection laws, including the GDPR.However, the feasibility of applying differential privacy approaches to crowd management remains an open topic of research.Device-free NVB solutions, both based on RF or other physical properties (such as audio or ultrasound signals), are less harmful to an individual's privacy since they do not require processing of devices' IDs and, thus, according to the GDPR, they do not collect personally identifiable information that could potentially be used to identify a particular person during people counting activities.This pushes towards emerging NVB techniques using reflectedpower approaches [164], also referred to as passive NVB solutions, which reuse existing over-the-air signals to count people or estimate their flow, by viewing the signals as power carriers rather than information ones.

5) Communications:
A constant stream of up-to-the-minute data can help PT system operators stay one step ahead of crowd situations.However, with limited budgets and compressed time frames, implementing a state-of-the-art CSS can present fundamental challenges.There are three basic challenges regarding ITSs based upon acquisition and processing of large amounts of sensor data: bandwidth, latency, and heterogeneity of data and infrastructures.
A first critical issue is represented by the significant bandwidth demand to connect thousands of devices and support hundreds of real-time video feeds, along with data gathered from ubiquitous sensors.A potential approach to fulfill such a bandwidth demand is offered by the unlicensed band technology, which has attracted significant effort during the last decade.In particular, the new radio unlicensed (NR-U) technology appears to be particularly suited for ITS applications [165], since it considers multiple bands and other deployment scenarios, such as dual connectivity and standalone operation in unlicensed bands.
In addition to bandwidth, since some crowd management decisions are time-sensitive in nature, latency is another key performance indicator of the CSS.Cloud computing at the edge of the network [166], namely, close to railway/metro stations, bus stops, and ITS sensors, can provide a solution for satisfying latency and bandwidth constraints, thus avoiding unacceptable upload delays as well as energy consumption of IoT sensors.In those scenarios, when there is no edge server nearby that can offload the tasks, employing UAVs is a promising solution by serving as computing-communications edge server for resource-constrained IoT devices [167].One can imagine a hybrid communication architecture for future ITSs where edge nodes locally process the time-critical raw sensed data, while non-time-sensitive data are transmitted to the cloud.
From a communication viewpoint, an ITS is a heterogeneous network, where a huge number of devices are connected to global networks using multiple technologies and platforms (i.e., cloud, edge, wireless, etc.), with various intelligence levels.Traditionally, communication between different wireless technologies is achieved indirectly via gateways equipped with multiple radio interfaces, which will become a bottleneck when many heterogeneous IoT devices are deployed.Crosstechnology communications (CTC) opens a new direction of direct communication among different wireless technologies, when they operate in the same spectrum band [168].These technologies are purely a software-based solution, requiring no hardware modification.6) Crowd control: A significant research effort is studying how to optimally employ (possible heterogenous) crowding data to optimize (at the various levels) the PT system behavior.Suitable performance metrics must be defined, as well as optimization methodologies, especially when applied to largescale complex PT networks involving thousands of vehicles and stations/stops.Indeed, in many papers crowd control strategies are localized, i.e., they are applied only to single lines or to a small portion of the network, whereas obviously crowd control must be applied holistically to the whole PT network.As an example, stop-skipping can solve the problem of a single vehicle overcrowding, but can negatively affect crowding at the stops or metro platforms.To perform complex system-wide optimization, it seems difficult to apply modelbased approaches.Thus, the preferred choice should be resorting to techniques such as data mining, AI, or ML/DL, along the vision of data-driven PT systems [169].An interesting issue is to assess the impact of inaccurate crowd monitoring/prediction measurements due to sensing on optimization strategies, in order to devise robust techniques.Another interesting development is how to incorporate crowding data in existing simulation tools for transportation planning, like BusMezzo [76] or SUMO [77].

VIII. CONCLUSIONS AND DIRECTIONS FOR FUTURE WORK
A clean, smart, and resilient PT system is at the core of worldwide economies and is central to people's lives: this is why there has been an increased number of research articles that discuss the feasibility of integrating ICT/IoT sensing solutions in ITSs.In this paper, we have reviewed the main ICT/IoT sensing technologies for crowd analysis, showing also how they can be adopted in a reference architecture aimed at introducing innovative crowd management functionalities in legacy ITS systems.The new framework is based on some basic components and subsystems, which can be used as building blocks to implement an evolved ITS, capable of real-time monitoring and predicting crowd situations, as well as disseminating useful information to users at the bus stops/stations and/or through mobile transport apps.A series of challenges arise from the proposed reference architecture, such as incentive mechanisms for user cooperation, data fusion and processing of heterogeneous data, security, privacy, broadband and ultra-low latency communications, system-wide optimization of crowd control strategies.Some features of the new framework are similar to those of a contact tracing system, which can be implemented more easily by resorting to user cooperation.In this sense, the crowd detection functionalities can be incorporated in a more complex system, which can implement, besides the typical mobile transport app functionalities (like Moovit), also the possibility to buy tickets and/or to reserve the access to the vehicles, in conditions of particular crowding.This could represent a decisive incentive to the use of a PT system.However, the more appropriate crowd monitoring solution must be singled out case-by-case, in dependence on the scenario, the ICT infrastructure owned or leased by the PT operator, the socioeconomic context, and the cost-benefits ratio.The potentials of crowd management go beyond the scope of dealing with typical social distancing problems, by also allowing real-time optimization and adaptive management of PT systems.
Finally, the study of the open literature makes it clear that an in-depth performance analysis of the sensing techniques for PT systems reported in Table III has not yet been carried out in each considered scenario.Therefore, a first interesting research development consists of performing a comparative statistical performance evaluation (numerical and/or experimental) of such sensing techniques for each outlined indoor/outdoor scenarios.An additional research issue is to carry out a performance analysis of the DF algorithms for associating, correlating, and combining information from different sensors, whose accuracy is crucial to provide effective actions.

Fig. 1 .
Fig. 1.Main architecture of the crowd analysis/monitoring and crowd control mechanisms.

Fig. 2 .
Fig. 2. A crowd management reference architecture for ITS.

4 )
Privacy: Privacy is one of the major issues related to sensing technologies for crowd management.In what follows, VB and NVB techniques are discussed separately, since they involve different privacy concerns that can be overcome by resorting to radically different technical solutions.

TABLE I COMPARISON
AMONG SURVEYS RELATED TO CROWD MANAGEMENT IN PT SYSTEMS way/metro stations, and bus stops Infrared sensors, pressure/load sensors, optical and thermal cameras, LiDAR, acoustic/ultrasound sensors, RFID and NFC devices, Bluetooth, WiFi, and cellular systems (4G/5G) be reported (in aggregated or anonymized form, for privacy concerns) to PT users, by means of displays installed inside vehicles or at stations/bus stops, or through mobile transport apps (e.g., Moovit or proprietary operator applications).

TABLE II REPRESENTATIVE
PERFORMANCES OF CROWD MANAGEMENT SOLUTIONS FOR PT SYSTEMS Reduction of the passenger in-vehicle travel time by 39% PWT = Passenger Waiting Time, EWT = Extended Waiting Time

TABLE III A
TAXONOMY OF SENSING TECHNOLOGIES FOR CROWD MANAGEMENT IN PT SYSTEMS.
a Power consumption refers to a single sensor.

TABLE IV MAIN
ADVANTAGES RELATED TO ADOPTION OF CROWD MANAGEMENT IN PT SYSTEMS.

TABLE V A
COLORTABLE SHOWING THE IMPACT OF THE MAIN CHALLENGES ON THE SUBSYSTEMS OF THE INTRODUCED REFERENCE ARCHITECTURE: THE INTENSITY OF THE COLOR VARIES FROM DARKER (MAPPING AN HIGHER IMPACT) TO LIGHTER (MAPPING A LOWER IMPACT).