A Comprehensive Survey of Enabling and Emerging Technologies … · 2020. 12. 15. · C. T. Nguyen et al.: Comprehensive Survey of Enabling and Emerging Technologies for Social Distancing

Received August 11, 2020, accepted August 13, 2020, date of publication August 20, 2020, date of current version September 1, 2020.

Digital Object Identifier 10.1109/ACCESS.2020.3018124

A Comprehensive Survey of Enabling andEmerging Technologies for SocialDistancing—Part II: EmergingTechnologies and Open IssuesCONG T. NGUYEN 1,2, YURIS MULYA SAPUTRA 3,4, (Graduate Student Member, IEEE),NGUYEN VAN HUYNH 3, (Graduate Student Member, IEEE),NGOC-TAN NGUYEN 3,5, (Graduate Student Member, IEEE), TRAN VIET KHOA 5,BUI MINH TUAN 5, DIEP N. NGUYEN 3, (Senior Member, IEEE),DINH THAI HOANG 3, (Member, IEEE), THANG X. VU 6, (Member, IEEE),ERYK DUTKIEWICZ 3, (Senior Member, IEEE),SYMEON CHATZINOTAS 6, (Senior Member, IEEE), AND BJÖRN OTTERSTEN 6, (Fellow, IEEE)1Department of Computer Science and Engineering, Ho Chi Minh City University of Technology, Ho Chi Minh City 700000, Vietnam2Department of Computer Science and Engineering, Vietnam National University-Ho Chi Minh City, Ho Chi Minh City 700000, Vietnam3School of Electrical and Data Engineering, University of Technology Sydney, Sydney, NSW 2007, Australia4Department of Electrical Engineering and Informatics, Vocational College, Universitas Gadjah Mada, Yogyakarta 55281, Indonesia5VNU University of Engineering and Technology, Vietnam National University, Hanoi 711000, Vietnam6Interdisciplinary Centre for Security, Reliability and Trust, University of Luxembourg, 4365 Luxembourg City, Luxembourg

Corresponding author: Cong T. Nguyen ([email protected])

This work was supported in part by the Joint Technology and Innovation Research Centre—a partnership between the University ofTechnology Sydney (UTS), in part by the Vietnam National University, University of Engineering and Technology Hanoi (VNU UET), andin part by the Vietnam National University, Ho Chi Minh City University of Technology (VNU HCMUT).

ABSTRACT This two-part paper aims to provide a comprehensive survey on how emerging technologies,e.g., wireless and networking, artificial intelligence (AI) can enable, encourage, and even enforce socialdistancing practice. In Part I, an extensive background of social distancing is provided, and enablingwireless technologies are thoroughly surveyed. In this Part II, emerging technologies such as machinelearning, computer vision, thermal, ultrasound, etc., are introduced. These technologies open many newsolutions and directions to deal with problems in social distancing, e.g., symptom prediction, detection andmonitoring quarantined people, and contact tracing. Finally, we discuss open issues and challenges (e.g.,privacy-preserving, scheduling, and incentive mechanisms) in implementing social distancing in practice.As an example, instead of reacting with ad-hoc responses to COVID-19-like pandemics in the future, smartinfrastructures (e.g., next-generation wireless systems like 6G, smart home/building, smart city, intelligenttransportation systems) should incorporate a pandemic mode in their standard architectures/designs.

INDEX TERMS Social distancing, pandemic, COVID-19, wireless, networking, positioning systems, AI,machine learning, data analytics, localization, privacy-preserving, scheduling, incentive mechanism.

I. INTRODUCTIONIn the presence of contagious diseases such as the currentCOVID-19 pandemic, social distancing is an effective non-pharmaceutical approach to limit the disease transmission.

The associate editor coordinating the review of this manuscript and

approving it for publication was Derek Abbott .

By reducing the frequency and closeness of human physicalcontacts, social distancing can lower the probability of thedisease transmission from an infected person to a healthy one,thereby significantly limiting the disease’s spread and sever-ity. During the ongoing COVID-19 pandemic, many govern-ments have implemented various social distancing measuressuch as travel restrictions, border control, public places

VOLUME 8, 2020 This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/ 154209

https://orcid.org/0000-0001-6461-4560https://orcid.org/0000-0001-8862-2004https://orcid.org/0000-0002-7696-0521https://orcid.org/0000-0002-9302-6474https://orcid.org/0000-0002-6488-404Xhttps://orcid.org/0000-0001-7306-9349https://orcid.org/0000-0003-2659-8648https://orcid.org/0000-0002-9528-0863https://orcid.org/0000-0002-8374-443Xhttps://orcid.org/0000-0002-4268-9286https://orcid.org/0000-0001-5122-0001https://orcid.org/0000-0003-2298-6774https://orcid.org/0000-0002-0945-2674

C. T. Nguyen et al.: Comprehensive Survey of Enabling and Emerging Technologies for Social Distancing—Part II

closures, and quarantines. Nevertheless, the implementationof such aggressive and large-scale measures is facing signifi-cant challenges such as negative economic impacts, personalrights violation, difficulties in changing people’s behavior,and the difficulties arisen when there are many people stayingat home. In such context, emerging technologies such asArtificial Intelligence (AI) can play a key role in addressingthose challenges.

In this two-part paper, we present a comprehensive surveyof enabling and emerging technologies for social distanc-ing. In Part I [1], we provide a comprehensive backgroundon social distancing and how wireless technologies can beleveraged to enable, encourage, and enforce proper socialdistancing implementation. In this Part II, we discuss variousemerging technologies, e.g., AI, thermal, computer vision,ultrasound, and visible light, which have been introducedrecently in order to address many new issues related to socialdistancing, e.g., contact tracing, quarantined people detectionand monitoring, and symptom prediction. For each technol-ogy, we have provided an overview, examined the state-of-the-art, and discussed how it can be utilized in differentsocial distancing scenarios as illustrated in Fig. 1. Finally,some important open issues and challenges (e.g., privacy-preserving, scheduling, and incentive mechanisms) of imple-menting technologies for social distancing will be discussed.Furthermore, potential solutions together with future researchdirections are also highlighted and addressed.

As illustrated in Fig. 2, the rest of this paper is orga-nized as follows. We first discuss emerging technologies forsocial distancing in Section II. After that, we discuss openissues and future research directions of technology-enabledsocial distancing in Section III, and conclusions are given inSection IV.

II. EMERGING TECHNOLOGIES FOR SOCIALDISTANCINGIn addition to the wireless technologies, emerging technolo-gies such as AI, computer vision, ultrasound, inertial sen-sors, visible lights, and thermal also can all contribute tofacilitating social distancing. In this section, we categorizethose technologies into sensing intelligence and machineintelligence technologies, provide a brief overview of eachtechnology, and discuss how they can be applied for differentsocial distancing scenarios.

A. SENSING INTELLIGENCE1) ULTRASOUNDThe ultrasound or ultrasonic positioning system (UPS) is usu-ally used in the indoor environment with the accuracy of cen-timeters [6]. The system includes ultrasonic beacons (UBs)as tags or nodes attached to users and transceivers. Beaconunits broadcast periodically ultrasonic pulses and radio fre-quency (RF) messages simultaneously with their unique IDnumbers. Based on these pulses and messages, the receiver’sposition can be determined by position calculation methods

such as trilateration or triangulation [7]. In comparison withother RF-based ranging methods, the UPS does not require aline of sight between the transmitter and the receiver, and italso does not interfere with electromagnetic waves. However,since the propagation of the ultrasound wave is limited, mostUPS applications for social distancing are only limited withinthe indoor environment.

a: KEEPING DISTANCEFor this purpose, UPS can be used to position and notifypeople. One of the first well knownUPS systems is Active Bat(AB) [8] based on the time-of-flight of the ultrasonic pulse.Typically, an AB system consists of an ultrasonic receivermatrix located on the ceiling or wall, a transmitter attached toeach target, and a centralized computation system to calculatethe objects’ positions. As presented in [8], by using a receivermatrix with 16 sensors, the AB system can achieve very highpositioning accuracy, i.e., less than 14 centimeters. However,a limitation of this system is its high complexity, especially ifa large number of ultrasonic sensors are deployed.

Another limitation of the AB system is the privacy riskfor users since the location of users under the AB system iscalculated at the central server. To address that, the Cricket(CK) system is proposed in [9] wherein the position calcu-lation is executed at the receivers. Specifically, a receiver inthe CK system passively receives RF and ultrasound signalsfrom UBs located on the wall or ceiling, and then the receivercalculates its position by itself based on UBs’ ID and coordi-nates. Since the receivers do not transmit any signals, the pri-vacy of users will not be compromised. Fig. 3 demonstratesthe two systems in the keeping distance application.

b: REAL-TIME MONITORINGIn the context of social distancing, UPS can be an effectivesolution for real-time monitoring scenarios, especially gaug-ing the number of people in public buildings. In particular,the main characteristic that makes UPS different from otherpositioning technologies is confinement, i.e., the ultrasoundsignal is confined within the same room as the UBs [7].Among the other positioning technologies, only infrared tech-nology shares the same characteristic. Nevertheless, infraredsignals are prone to interference from sunlight and other ther-mal sources, and they also suffer from line-of-sight loss [7].As a result, ultrasound is the most efficient technology forbinary positioning [7], i.e., determine if the object is in thesame room as the UBs or not. Thus, UPS can be particularlyuseful in the social distancing scenarios where the exact posi-tions of people are not as necessary as the number of peopleinside a room (e.g., to limit the number of people). Thistechnology is more efficient because it needs a few referencenodes (e.g., UBs) to determine the binary positions of people,which can significantly reduce implementation costs.

c: AUTOMATIONUltrasound can also be applied in the social distancing sce-narios that utilize medical robots or UAV. Mobile robots,

154210 VOLUME 8, 2020


FIGURE 1. Application of technologies to different social distancing scenarios. Some technologies, e.g., AI and Thermal, can be applied to manyscenarios, whereas technologies such as Visible Light and Ultrasound are applicable to fewer scenarios. Scenarios from the same group have the samecolor. The arrows that show the links from one technology to different scenarios have the same color.

FIGURE 2. The organization of this (Part II) paper.

especially medical robots, can play a key role in reducingthe physical contact rates between the healthcare staff (e.g.,doctors and nurses) and the patients inside a hospital, therebymaintaining a suitable social distancing level. In such sce-

narios, UPS can help to improve the navigation of medicalrobots. In [11], a navigation system based on Wi-Fiand ultrasound is proposed for indoor robot navigation.To deal with the uncertainties which are very common in

VOLUME 8, 2020 154211


FIGURE 3. Ultrasound application for keeping distance using a) Cricket system [9], and b) Active Bat system [8]. The maindifference between the two systems is that the user’s position is calculated by the user device in Cricket and by a centralserver in Active Bat.

crowded places like hospitals, the system employs a PartiallyObservable Markov Decision Process, and a novel algorithmis also introduced to minimize the calibration efforts.

In the social distancing context, besides outdoor applica-tions, UAVs can also be employed to reduce the necessity ofhuman physical presence. For example, UAVs can be usedto deliver goods inside a building or to manage warehouseinventory. However, most of the existing studies focus onUAV navigation for the outdoor environment, which oftenrelies on GNSS for UAV positioning. Since GNSS’s accuracyis low for the indoor environment, these methods cannotbe applied directly for UAV navigation inside a building.To address that limitation, a navigation system is proposedin [10], which utilizes ultrasound, inertial sensors, GNSS,and cameras to provide precise (less than 10 cm) indoornavigation for multiple UAVs.Summary: Ultrasound can be applied in several social

distancing scenarios. In the keeping distance scenarios, UPSsystems such as AB and CK can be applied directly to localizeand notify people to keep a safe distance. Moreover, dueto its confinement characteristic, ultrasound is one of themost efficient technology for binary positioning, which isparticularly useful for monitoring and gauging the numberof people inside the same room. In the automation scenarios,ultrasound can facilitate UAVs and medical robots naviga-tions, especially for the indoor environment.

2) INERTIAL SENSORSIn the context of social distancing, inertial-sensors-basedsystems can be applied in distance keeping and automationscenarios as illustrated in Fig. 4. For example, positioning

FIGURE 4. Inertial-sensors-based systems for several social distancingscenarios. In the keeping distance scenario, the built-in inertial sensors ofsmartphones can be utilized for user positioning. Based on this,the smartphone can warn the user when there are other users or crowdsin close proximity. Inertial sensors can also help to localize and navigateUAVs and medical robots.

applications utilizing built-in inertial sensors can bedeveloped for smartphones which can alert the users whenthey get close to each other or a crowd. Moreover, inertialsensors can be integrated into robots and vehicle positioningsystems, which can facilitate autonomous delivery servicesand medical robot navigation. All of these scenarios cancontribute to reducing the physical contact rate betweenpeople.

Inertial sensors consist of two special types of sensors,namely gyroscopes and accelerometers, attached to an objectto measure its rotation and acceleration. Based on the mea-sured rotation and acceleration data, the orientation andposition displacements of the object can be determined [95].Because inertial sensors do not require any external referencesystem to function, they have been one of the most commonsensors for dead reckoning navigation systems, i.e., cal-culation of the current position is based on a previously

154212 VOLUME 8, 2020


determined position. Such navigation systems can provideaccurate positioning within a short time frame. However,since the current position is determined based on the previ-ously calculated positions, the errors accumulate over time,i.e., integration drift. Therefore, Inertial-Navigation-System(INS) is often used in combination with other positioning sys-tems, e.g., GPS, to periodically reset the base position [95].

a: KEEPING DISTANCETraditionally, INS has been widely used for aviation, marine,and land vehicle navigation. Recently, the ever-increasingpresence of smartphones has enabled many INS applicationsfor pedestrian positioning and navigation, which can supportsocial distancing scenarios. Moreover, INS is one of the fewtechnologies that can enable accurate pedestrian positioningfor the outdoor environment, especially when combined withother outdoor positioning technologies such as GPS. In [96],a smartphone-based positioning system is proposed. The sys-tem makes use of a smartphone’s built-in sensors, includinggyroscopes, accelerometers, andmagnetometers (sensors thatmeasure magnetism), to calculate the smartphone’s position.In particular, magnetometers are combined with gyroscopesto improve accuracy of rotation measurements. This is doneby correlating their measurements via a novel algorithmwhich uses four different thresholds to determine the weightsof the gyroscope and magnetometers measurements in thecorrelation function. In [97], a wearable body sensors sys-tem using inertial sensors is proposed to measure the lowerlimb motion. The proposed system consists of three sensorsattached to different parts of the human lower limb tomeasureits orientation, velocity, and position. Using this informationand the initial position of a person, the person’s location canbe tracked.

In [46], a novel indoor positioning system is developedusing Wi-Fi and INS technologies. In this system, INS isutilized for the area where Wi-Fi coverage is limited, whileWi-Fi positioning is used to compensate INS’s integrationdrift. Another positioning system using inertial sensors andWi-Fi is presented in [98], where Wi-Fi fingerprinting tech-nique is used to improve the accuracy of the dead reckoningnavigation. Because of the integration drift, a dead reckoningnavigation system needs to frequently update its positionby referencing to an external node. In the proposed system,a Wi-Fi fingerprinting map is set up in advance and the deadreckoning system can use the map to update its position.Moreover, in [46], the authors propose using Kalman filter tocombine the measurement data from Wi-Fi and INS, whichcan reduce the error to 1.53 meters.

Besides Wi-Fi, INS can be used in combination with otherpositioning technologies. In [99] and [100], INS has beencombined with the UWB technology for pedestrian position-ing and tracking. Generally, INS helps to reduce UWB’s highimplementation cost and complexity, while INS’s integrationdrift can be compensated. Particularly, INS is employed tocompensate for the UWB’s low dynamic range and pronenessto external radio disturbances in [99]. To enable the combi-

nation, an information fusion technique using the extendedKalman filter is proposed to fuse the measurement datacoming from both the INS and UWB sensors. The resultshows that the hybrid system can achieve better performancethan both the individual systems. In [100], the informationfusion problem between the INS and UWB is optimized tominimize the uncertainties in the measurements. As a result,the positioning accuracy can be significantly improved.

b: AUTOMATIONBesides pedestrian positioning, INS can also be appliedfor social distancing scenarios involving autonomous vehi-cles, e.g., medical robots and drone delivery. Generally, INShas been commonly used for medical robot applications,including surgeon assists, patient motion assists, and deliveryrobots. In this section, we will only focus on the medicaland delivery robot applications for social distancing pur-poses. In [101], a novel INS system is developed specificallyfor mobile robot navigation. In addition, an error model isproposed to increase the accuracy of the involved inertialmeasurements. A Kalman filter is also proposed to preciselyestimate the velocity and orientation of the robot in the pres-ence of noises. A novel data fusion algorithm, leveraging anadaptive Kalman filter is presented in [102] for indoor robotpositioning based on an INS/UWB hybrid system.

Unlike INS for mobile robots that are mostly developedfor the indoor environment, INS for UAV focuses on outdoorapplications. Note that UAV navigation must also consider itsaltitude, which adds more complexity. The authors of [103]leverage inertial sensors and cameras to determine the UAV’sposition, velocity, and altitude. Particularly, the camerasattached to the UAV capture the images of the surroundingenvironment and send them to a control station. This stationwill then process the images to determine the UAV’s pose inregards to the surroundings. The pose’s data is then combinedwith the inertial sensors data via a Kalman filter to deter-mine the UAV’s position and velocity. Similarly, a systemcombining inertial and vision sensors is developed in [104]for UAV positioning and navigation. The system utilizestwo observers which have inertial and vision sensors. Thefirst observer calculates the orientation based on gyroscopeand vision sensors, and the second observer determines theposition and velocity based on data from the accelerometersand vision sensors. The experimental results show that thevision sensors measurements can be used to compensate forthe inertial sensors errors, thereby achieving a high accuracyeven with low-cost inertial sensors.Summary: The omnipresence of smartphones with built-

in inertial sensors has opened many opportunities for devel-oping positioning systems based on INS. For the distancekeeping scenarios, INS positioning systems, especially forpedestrians, can play a vital role as they are readily available.In the other scenarios such as medical robot navigation andUAV delivery, INS-based techniques can help to increase theefficiency (more accurate path, and lower traveling time) ofthe existing navigation systems.

VOLUME 8, 2020 154213


3) VISIBLE LIGHTThe recent development in the light-emitting diodes (LEDs)technology has enabled the use of existing light infras-tructures for communication and localization purposes dueto attractive features of visible lights such as reliability,robustness, and security [12]–[14]. Visible light communi-cation (VLC) systems usually comprise two major compo-nents, i.e., LED lights corresponding to transmitters to sendnecessary information (e.g., user data and positioning infor-mation) via visible lights and photodetectors (e.g., photodi-odes) and imaging sensors (e.g., camera) playing the role ofreceivers [2]. Due to the ubiquitous presence of LED lights,VLC can be leveraged in many social distancing scenarios asdiscussed below.

a: REAL-TIME MONITORINGCommunication systems using visible light (e.g., LED-basedcommunications) can provide precise navigation and local-ization solutions in indoor environments. Utilizing this tech-nology, some applications can be implemented to supportsocial distancing such as tracking individuals who are beingquarantined, detecting and monitoring crowds in publicplaces as shown in Fig. 5(a).Due to many advantages such as low cost and ease of

implementation, the VLC receiver using photodiodes canbe employed as a ‘‘tag’’ that is integrated into mobile tar-gets such as trolleys/shopping carts, autonomous robots, etc.People attached with these tags can perform self-positioningbased on the triangulation method so that they can avoidcrowded areas. Furthermore, the tags’ locations can be col-lected by the authorities to monitor people in public areas.Based on this location data, further actions can be carried outsuch as warning people by varying the color temperature ofthe lights in the crowded areas. It is worth noting that thissolution will not reveal any personal information of users(e.g., customers) because it only requires communicationsbetween VLC-based tags and light fixtures. However, mostVLC systems only provide half-duplex communications dueto the fact that LED lights operate in the role of transmit-ters. Therefore, they should be combined with other wirelesstechnologies like Bluetooth [15], [16], and Infrared [17] toenable an uplink communication with the server for locationinformation exchange. Moreover, to improve the accuracyof positioning people in indoor environments using photo-diodes, some advanced techniques can be used such as datafusion of AOA and RSS methods proposed in [18] and theAOA method using a multi-LED element lighting fixtureintroduced in [19]. Onemain disadvantage of the photodiode-based VLC systems is the need for hardware (i.e., the pho-todiode receiver) mounted on smart trolleys/shopping cartsto receive light signals. Consequently, the system might failto detect the locations of people who do not carry them.Nevertheless, pureLiFi company has recently invented a tinyoptical front end which can be integrated into smartphonesto take benefits of the photodiode receiver in high accuracyVLC-based localization services [20].

The rapid development of smartphones has enabled VLC-based applications on handheld devices such as indoorlocalization and navigation applications (e.g., smart retailsystems [15], [16], [21]). These systems use front-facingcameras of mobile phones to receive visible light signalscontained positioning information (e.g., the LED light’s IDor location) from visible light beacons [22]. The capturedphotos collected regularly by the front-facing camera aresent to a cloud/fog server for image processing to alleviatethe computation on the phone. Then, the beacon’s ID andcoordinates can be extracted and sent back to the phone.After that, the AoA algorithm is implemented to estimatethe location and orientation of the phone. An attractive usecase of the camera-based VLC systems [15], [16], [21] isto assist users to quickly find specific products in shoppingmalls or supermarkets. Thus, we can adopt this functionto implement tracking and monitoring crowds in publicplaces as well as assisting people in avoiding crowds ina proactive manner. It is worth noting that this solution ismore convenient than using photodiodes since it uses front-facing cameras of smartphones as the VLC receivers, thuseveryone using smartphones can be tracked. However, dueto continuous photo shooting, these positioning applicationsare very energy-consuming, which is a major drawback ofcamera-based VLC systems when they are used for trackingpeople.

b: AUTOMATIONIn public places, there is always a need for assistance inspecific circumstances (e.g., information or physical supportsfor customers, older and disabled people). For instance, sup-porting staff in supermarkets can assist customers in findingproducts or help elder/disabled people to carry their goods.Similar assistance scenarios can be seen in hospitals, banks,and libraries. This results in an increase in close physi-cal contacts between customers and assistants. Therefore,autonomous assistance systems using VLC technology canbe employed to minimize the physical contacts as shown inFig. 5(a) and (b).Besides the navigation purpose, the smart retail sys-

tems [15], [16] can also provide information assistance ser-vices for shoppers. For example, the product description, saleinformation, or other necessary information can be displayedon the screen when the phone is under a certain LED light.Another example is information assistance services in muse-ums [23], [24]. This can help to reduce the number of closephysical contacts in these places.

Similar to the information assistance systems for reduc-ing close physical contacts, autonomous robots using theVLC technology for communication and localization can alsobe deployed to assist people in certain circumstances, forexample, elderly-assistant robots, walking-assistant robots,shopping-assistant robots, etc., [25], [26]. Moreover, visiblelight signals do not cause any interference to RF signals,and thus they can be effectively deployed in diverse indoorenvironments such as hospitals, schools, and workplaces.

154214 VOLUME 8, 2020


FIGURE 5. Visible light communications supporting social distancing in several scenarios. In indoor environments, visible light sensors can be utilized forreal-time monitoring, information assistance system, and navigation application. For outdoor environments, visible light sensors can support trafficcontrol.

c: TRAFFIC CONTROLIn the context of social distancing, high demand traffic cancause a large concentration of people in a certain area (e.g.,city center). By adopting smart traffic light systems in [27],[28], we can deploy an intelligent traffic controlling systemusing the VLC technology to control large traffic flows asillustrated in Fig. 5(c). That can help to reduce vehicle densityin public areas. The VLC technology provides a communi-cation method between vehicles and the light infrastructure(e.g., traffic lights, street lights). First, vehicles can sendtheir information (e.g., their IDs) to the light infrastructureby using its headlights as transmitters, thus the system candetect and monitor the traffic flow. However, in this case,it is required that the light infrastructure must be equippedwith VLC receivers (e.g., traffic cameras or photodiodes).Second, based on the awareness of the traffic, the system cancontrol the vehicles by sending instructions to guide them.In this case, the system uses traffic lights, or street lights astransmitters to send information and the vehicles use dashcameras to receive the information. For example, the systemwill notify them about hot zones that have a high density ofvehicles and do not allow them to enter, so that they can avoidthese zones.Summary: The availability of smart retail systems is proof

of the superior performance and convenience of VLC tech-nology compared to other RF technologies in high precisionindoor localization and navigation. By leveraging such com-mercial approaches, we can deploy the cost-effective crowd

monitoring system on a large scale, not only in shoppingmalls or hypermarkets but also in other public places, suchas airports, train stations, and hospitals, based on the existingilluminating infrastructures. Building/facilities managers canimmediately alert or notify the users if they are in the middleof a crowd (e.g., varying the color temperature of the lights inthe high-density zones). People can also take the initiative inplanning their move to the desired locations without encoun-tering the crowds. On the other hand, assistance systems helpto reduce the number of staff/volunteers, nurses inside publicbuildings; or limit the close contacts between them and cus-tomers, patients. Moreover, the combination with other RFtechnologies such as Bluetooth and Infrared also ensures thelocation-based services are not interrupted when the smart-phone is not being actively used by the user (e.g., the phoneis in the pocket). Last but not least, the VLC technology canbe a potential communication method between the intelli-gent traffic controlling system and vehicles in the outdoorenvironment. However, the main disadvantage of the VLCtechnology is that interference from ambient and sun lightshave significant impacts on the visible light communicationchannels [12], [14]. It results in poor performance of the RSS-based positioning approaches and outdoor communications.

4) THERMALThermal based positioning systems can be classified into twomain categories which are infrared positioning (IRP) systemsand thermal imaging camera (THC). Typical IRP systems

VOLUME 8, 2020 154215


such as [29], [31], [32] are low-cost, short-range (up to10 meters) systems that use infrared (IR) signals to determinethe position of targets via AOAor TOAmeasurementmethod.On the other hand, the THC, which constructs images fromthe object’s heat emission, can operate at a larger range (upto a few kilometers) [35]. Because of this difference, IRP andTHC can be applied in different social distancing scenariosas discussed below.

a: KEEPING DISTANCEIn keeping distance scenarios, IRP systems such as ActiveBadge [31], Firefly [33], and OPTOTRAK [32] can be uti-lized. In the Active Badge, badges that periodically emitunique IR signals are attached to the targets. Based on the dis-tances from the fixed infrared sensors to the badges, the tar-get’s position can be calculated. As a result, this applicationcan be useful to determine the distance between two people aswell as to identify crowds in indoor environments. The mainadvantages of this solution are low cost and easy implemen-tation. However, it requires users to wear tag devices to tracktheir locations.

To achieve a higher positioning accuracy, the Firefly [33]and OPTOTRAK [32] systems can be implemented. Thesesystems contain infrared camera arrays and infrared trans-mitter called markers. Due to the difference in setups (onetarget is attached with one tag in Firefly and multiple tags inOPTOTRAK), the Firefly system can accurately determinethe target’s 3D position, whereas the OPTOTRAK systemcan capture the target’s movement. The main disadvantage ofthese systems is that they are prone to interference from otherradiation sources such as sunlight and light bulbs. Combinedwith their short-range, IRP is mostly applicable in smallrooms with poor-light conditions.

b: PHYSICAL CONTACT MONITORINGSince the Firefly and OPTOTRAK systems can accuratelycapture movements, they can be useful for contact tracingscenarios in social distancing. For example, markers can beattached to the target’s body parts which are usually used inphysical contacts, e.g., hands for handshakes and body forhugs. The movement of these body parts can then be capturedby the IR camera as illustrated in Fig. 6, and the recordeddata can be analyzed later to determine if there are closecontacts between the target and other people. Based on thisinformation, the contacts that the target made can be tracedlater if necessary.

c: REAL-TIME MONITORINGFor traffic monitoring in social distancing contexts, both IRPand THC can be utilized, especially in poor-light conditions.The authors in [34] propose a robust vehicle detector based onthe IRP under the condition to quantify traffic level and flow.The collected data can be sent to assist the authorities in socialdistancing monitoring. However, since IRP has a short range,THC systems such as [37] can be a better choice in a largerarea with high vehicle density.

FIGURE 6. Physical contact monitoring by infrared system [33].IR cameras can be utilized to detect and monitor physical contact amongpeople. If there is close contact between any two people, the event canbe recorded for future usage, e.g., contact tracing.

Due to its very high observation range (a few kilome-ters) [36]–[38], THC is particularly effective for real-timemonitoring scenarios, such as public building monitoring,detecting closure violation, and non-essential travel detec-tion, which does not require high positioning accuracy. THCsystems such as those proposed in [29], [30] are efficient inthese scenarios since they are light-weight and can cover awide area with medium accuracy.

Another application of thermal technology is to detectsusceptible groups. Since the THCs measure heat emittedfrom people or other objects, they can be used for checkingpeople’s temperature quickly from a far distance [39], [40].Further, the THC system has the ability to detect slight tem-perature differences with a resolution of 0.01 degrees [41].Thus, it can be a good means to check health conditionsand sickness trends of patients. Moreover, the system can bedeployed in shopping centers to measure customers’ temper-ature remotely. This can help to detect infection symptomsearly and also prevent the disease spread.Summary: Thermal based positioning systems are helpful

in some social distancing scenarios, especially in poor-light conditions. For short-range communication applica-tions, the IRP is cost-effective and can be used for positioningand tracing purposes. Whereas, some light-weight THC sys-tems can be leveraged for real-time monitoring over long dis-tances due to their high range. However, the high cost of THCshould be considered when implementing THCs in practice.

Table 1 provides a summary of the surveyed sensing intelli-gence technologies. Generally, each technology has a specialcharacteristic that makes it a very effective solution for aspecific scenario. For example, ultrasound signals are con-fined by walls, which enables low-cost ultrasonic positioningsystem to efficiently monitor people in a small room. Fur-thermore, since inertial sensors are built-in in most smart-phones, they can be quickly utilized for keeping distance insmartphone applications. In addition, visible light technologycan be leveraged for building information assistance sys-tems which help to reduce human presence. Finally, thermal

154216 VOLUME 8, 2020


FIGURE 7. Thermal cameras used in susceptible group detection and traffic monitoring. Thermal cameras can be used tocheck body temperature, thereby detecting susceptible groups and people with symptoms. For traffic monitoring, bothinfrared positioning systems and thermal cameras can be utilized, especially in poor-light conditions, e.g., at night.

camera is the only technology that can detect people overa large distance (a few kilometers) without the need forattached devices, which makes it an ideal solution to detectviolation of quarantines or closures.

B. MACHINE INTELLIGENCE1) COMPUTER VISIONComputer vision technology trains computers to interpretand understand visual data such as digital images or videos.Thanks to recent breakthroughs in AI (e.g., in pattern recog-nition and deep learning), computer vision has enabledcomputers to accurately identify and classify objects [48].Such capabilities can play an important role in enabling,encouraging, and enforcing social distancing. For exam-ple, computer vision can turn surveillance cameras into‘‘smart’’ cameras which can not only monitor people butalso can detect, recognize, and identify whether peoplecomply with social distancing requirements or not. In thissection, we discuss several social distancing scenarios wherecomputer vision technology can be leveraged, includingpublic place monitoring, and high-risk people (quaran-tined people and people with symptoms) monitoring anddetection.

a: PUBLIC PLACE MONITORINGDespite government restrictions and recommendations aboutsocial gathering, some people still do not comply withthem, which can cause the virus infection to the community.In such context, human detection features in object detec-tion [49], a major sub-field of computer vision, can helpto detect crowds in public areas through real-time imagesfrom surveillance cameras. An example scenario is describedin Fig. 8(a). If the number of people in an area does notmeet the social distancing requirement (e.g., gathering above10 people), the authorities can be notified to take appropriateactions.

There are two main approaches to detect humansfrom images in object detection namely region-based and

unified-based techniques. The former detects humans fromimages in two stages including the region proposal andthe processing according to the regions [50]. Based on thisapproach, several frameworks including Fast-RCNN [56] andFaster-RCNN [57] are developed in combination with Con-volution Neural Network (CNN) [54]. In [58], the authorsimprove the Faster-RCNN by proposing the Mask Regionswith the CNN features (Mask RCNN) method which masksthe bounding box to detect the object with high accuracywhile adding a minor overhead to the Faster-RCNN. MaskRCNN outperforms previous methods by simplifying thetraining process and improving the accuracy in detectinghumans in the images for calculating the density of peoplein a particular area.

Although the above region-based approach has high recog-nition accuracy [58], it has high complexity, which isunsuitable for devices with limited computational capacity.To address this, the unified approach is more appropriate toimplement, which can reduce the computational complexityby detecting humans from images with only one step. Thisapproach maps the pixels from the image to the boundingbox grid and class probabilities to detect humans or objectsin real-time. Following this direction, the You Only LookOnce (YOLO) method proposed in [59] can detect/predictobjects (even small ones) in real-time with high accuracy.In addition, in [60], the authors propose the Single ShotMultibox Detector (SSD) framework which uses a convo-lution network on the image to calculate a feature map andthen predict the bounding box. Through experimental results,they demonstrate that this method can detect objects fasterand more accurately than those of both YOLO and Faster-RCNN. For public place monitoring, both YOLO [59] andSSD [60] can be used to detect fast and accurately humansfrom real-time images or videos of surveillance cameras.After identifying people, we can use a real-time automaticcounter to count and identify whether the number of gather-ing people is complying with social distancing requirementsor not.

VOLUME 8, 2020 154217


TABLE 1. Summary of sensing intelligence technologies applications to social distancing.

FIGURE 8. Computer vision technologies for social distancing: (a) human detection to identify the number of people in the public place [51], (b) facerecognition to identify (b1) the full face of isolated person, (b2) person with mask or person behind the mask [52], and (c) pose estimation to detect onewith coughing symptom [53].

b: DETECTING AND MONITORING QUARANTINED PEOPLETo prevent the spread of the virus from an infected personto others, the infected person or people who had physicalcontact with them must be isolated at the restricted areas orat home. For example, people who come back from highlyinfected countries/regions of COVID-19 are often requestedto be quarantined or self-isolate for 14 days. Due to the lack offacilities, most countries require these people to self-isolate athome. In this case, the face recognition capability of computervision can help to enforce this requirement by analyzing theimages or videos from cameras to identify these people (i.e.,to check whether they breach the self-isolation requirementsor not). If these people are detected in public, the authoritiescan be notified to take appropriate actions.

Unlike object detection, the dataset including the fullface images of the isolated people needs to be built. Theface recognition system firstly learns from this dataset andthen analyzes the images from public surveillance camerasto identify their appearances as in Fig. 8(b1). The authorsin [61] propose a framework named DeepFace using DeepNeutral Network (DNN) which can detect with an accuracyof 97.35% and 91.4% in Labeled Faces in the Wild (LFW)and YouTube Faces (YTF) dataset, respectively. To improve

the accuracy in detecting humans from surveillance cameras,some advanced techniques can be implemented such as [62],[63] and [64].

To prevent the spread of infectious diseases such asCOVID-19, people are often required to wear masks in publicplaces, which necessitates approaches to recognize or identifypeople with or without masks as illustrated in Fig. 8(b2).For example, the cameras in front of a public building canrecognize and send warning messages (e.g., a beep sound)to remind the person who does not wear a mask when he/sheintends to get into the building. This idea is introduced in [70]by using CNN to detect people who do not wear the masks.However, this work is just at the first step, which still requiresmuch more efforts to demonstrate the effectiveness as well asimprove the accuracy.

c: SYMPTOMS DETECTION AND MONITORINGAfter a few days of being infected with the virus, the infectedperson may have some symptoms such as coughing or sneez-ing. To minimize spreading the virus to others, it would bevery helpful if we can detect these symptoms from people inpublic and inform them or the authorities. The idea here issimilar to that of using thermal imaging cameras at airports

154218 VOLUME 8, 2020


or train stations. Specifically, detecting human behaviors,motion, and pose in computer vision can play a pivotalrole [65]. Pose estimation captures a person with differentparts (as illustrated in Fig. 8(c)) then detects human behaviorsby studying the parts’ movements and their correlation. Forexample, a coughing person in Fig. 8(c) usually moves hishand near his head, and his head would have a vibration.

Recognition of human behaviors from surveillance cam-eras is a challenging problem because the same behaviorsmay have different implications, depending on the relation-ship with the context and other movements [66]. The recentadvances in AI/ML are instrumental in correlating differentmovements/parts to interpret the associated behavior. In [67]the authors propose to use CNN [54] to enhance the accuracyof the model of the interaction between different body parts.In addition, the authors in [68] introduce several methodsto detect body parts of multiple people in 2D images, andthe authors in [69] propose methods to estimate 3D posesfrom matching of 2D pose estimation with a 3D pose library.These works can be further developed for future studies todetect people with symptoms of the disease such as coughingor sneezing in real-time. To improve the accuracy of thesymptom detection in social distancing, computer vision-based behavior detection methods can be combined withother technologies, e.g., thermal imaging.

d: INFECTED MOVEMENT DATATo prevent the spread of the virus, tracing the path of aninfected person plays an important role in finding out thepeople who were in the same place as the infected person. Forthis purpose, computer vision technology can not only detectinfected people by facial recognition but also contribute tothe positioning process. In [43], the movement of people isdetermined by analyzing the key point of transition framescaptured from smartphone cameras. This method can drawthe trajectory of movements and the location with an accuracyaround two meters. In [44], the authors propose to combinethe human detection techniques of computer vision with dig-ital map information to improve the accuracy. In this study,the user path from cameras is mapped to the digital mapwhich has the GPS coordinates. This method can achieve avery high accuracy within two meters. In another approach,the authors in [45] propose to use both smartphones’ camerasand inertial-sensor-based systems to accurately localize tar-gets (with only 6.9 cm error). This approach uses the fusionof keypoints and squared planar markers to enhance theaccuracy of cameras to compensate for the errors of inertialsensors.

e: KEEPING DISTANCEComputer vision can also be very helpful to support peoplein keeping distance to/from the crowds. In [42], the authorsdevelop an on-device machine-learning-based system lever-aging radar sensors and cameras of a smartphone. Whenthe radar sensor detects the surrounding moving objects,the smartphone camera can be utilized to capture its

surrounding environment. Taking into account the recordeddata, the smartphone can train the data using machine learn-ing algorithms to determine the existence of nearby peopleand its distance from those people with respect to the socialdistancing requirements. We can also use a smartphone toestimate the distance between the mobile user and otherpeople using radar sensors and cameras along with machinelearning algorithms.Summary:Computer vision can be utilized in several social

distancing scenarios, especially the ones that require peoplemonitoring and detection. Particularly, computer vision is theonly method that can differentiate between people and iden-tify complex features such asmasks and symptoms. To furtherimprove the effectiveness of computer vision in the socialdistancing context, future research should focus on increasingthe accuracy and reducing the complexity of computer visionmethods, so that they can be integrated into existing systemssuch as surveillance cameras.

2) ARTIFICIAL INTELLIGENCEOver the last 10 years, we have witnessed numerous appli-cations of AI in many aspects of our lives such as health-care, automotive, economics, and computer networks [106].The outstanding feature of AI technologies is the ability toautomatically ‘‘learn’’ useful information from the obtaineddata. This leads to more intelligent automation, operatingcost reduction as well as the great compatibility to adaptto changing environments. For that, AI (and its underlyingmachine learning algorithms) can also play a key role in socialdistancing, especially in modern lives, with many practicalapplications, as discussed below.

a: DISTANCE TO/FROM CROWDS ANDCONTACT TRACINGApplications of machine learning to users’ location dataallow us to effectively monitor the distance between peopleand trace the close contacts of infected people. In [150],the authors analyze the accuracy of a user’s location pre-diction based on his/her friends’ location datasets. In thiscase, a temporal-spatial Bayesianmodel is developed to selectinfluential friends considering their influence levels to theuser. Thus, the service provider can predict the exact locationof a mobile user by using the temporal-spatial Bayesianmodel. Then, when the user is too close to other mobileusers/people at crowded public places or his/her friends whenthey go in a group as illustrated in Fig. 9(a), his/her smart-phone can alert to keep a safe distance. In addition, usingthe list of influential friends based on their ranks, the ser-vice provider can utilize it for the contact tracing purposewhen the mobile user or one of his/her influential friendsin the list gets infected. Moreover, the local-experts-findingscheme proposed in [151] can be utilized to find the localsocial media users of a certain area. Based on this, informa-tion such as current crowds locations can be extracted moreefficiently.

VOLUME 8, 2020 154219


FIGURE 9. Application of artificial intelligence to several social distancing scenarios. By predicting people location, AI can be utilized to warnpeople to avoid potential crowded places in (a) distance/to from crowds scenario, alert people about infected locations in (b) infected movementprediction scenario, detect quarantine violation in (c) quarantined/At-Risk people location prediction. AI can also be leveraged to warn people orpropose alternate routes in (d) people/traffic density prediction. Using data from audio and image sensors, AI can predict potential infected placesbefore the real disease occurs in (e) sickness trend prediction.

b: INFECTED MOVEMENT PREDICTIONAnother application of machine learning is to predict infectedpeople movement from one location to another one and hencecan potentially predict the geographic movement of the dis-ease. The prediction is particularly crucial as infected peoplemay travel to various places and can accidentally infect othersbefore know that they carry the disease. In [152], the authorsintroduce a smartphone-based location recognition and pre-diction model to detect the current location and predict thedestination of mobile users. In particular, the location recog-nition is implemented using the combination of k-nearestneighbor and decision tree learning algorithms, and the des-tination prediction is realized using hidden Markov mod-els. Given the history of infected people movement, we canadopt the above model to recognize and predict the potentialgeographic movement of the disease. Using the information,people can be advised to stay away from the possible infectedlocations through alerts from their smartphones as illustratedin Fig. 9(b).

c: QUARANTINED/At-RISK PEOPLE LOCATIONPREDICTIONThe current location prediction of quarantined people, e.g.,infected people, and at-risk people, e.g., sick and elder peo-ple, is very important to monitor whether they currentlystay at the self-quarantined and self-protection areas, e.g.,their homes, or not. To this end, a machine-learning-basedlocation prediction approach can help to detect the currentposition of those people in a certain area. In [153], the authorsapply the auto-encoder neural networks and one-class sup-port vector machines to verify whether a user is within aspecific area or not. Considering various channel models,i.e., path-loss, shadowing, and fading, the proposed solu-tions can achieve Neyman-Pearson optimal performance byobserving the probability ofmiss-detections and false-alarms.The authors in [154] propose a novel localization systemleveraging the federated learning to allow mobile users tocollaboratively provide accurate location services withoutrevealing mobile users’ private location. As such, the authors

154220 VOLUME 8, 2020


utilize deep neural networks with the Gaussian process toaccurately predict the desired location of the mobile users.As a result, we can apply the proposed solutions to detect ifinfected people or at-risk people currently move away fromtheir homes as illustrated in Fig. 9(c). Moreover, we canutilize the proposed solutions to determine the movementfrequency of the self-isolated people outside the protectionfacility. Using the movement frequency history, the author-ities can enforce them to stay at the protection facility forfurther infection prevention.

d: PEOPLE/TRAFFIC DENSITY PREDICTIONPredicting the density of people or the number of people inpublic places allows us to efficiently schedule or guide peopleto stay away or refrain from coming to soon-to-be over-crowded places. For example, when the predicted number ofpeople in a certain place almost reaches a predefined thresh-old (e.g., according to the social distancing requirement),the service provider can broadcast a local notification toincoming people via cellular networks, aiming at encouragingthem to move to another area. In [155], the authors adoptadvanced machine-learning-based approaches for edge net-works to predict the number of mobile users within base sta-tions’ coverages. Particularly, the framework first groups thebase stations into clusters according to their network data anddeployment locations. Then, using various machine learningalgorithms, e.g., the Bayesian ridge regressor, the Gaussianprocess regressor, and the random forest regressor, we canpredict the number of mobile users within their network cov-erages. From the preceding architecture, one can utilizeWi-Fihotspots and cluster them based on their locations. By doingso, we can predict the number of people within each cluster’scoverage. Using the same architecture, we can extend theapplication to predict the traffic level on the roads. Specifi-cally, upon predicting the number of vehicular users on theroads, we guide the drivers to choose particular routes tosatisfy the social distancing requirements, e.g., suggest alter-native routes to avoid crowded areas. In [156], the authorsintroduce a UAV-enabled intelligent transportation system topredict road traffic conditions using the combination of con-volutional and recurrent neural networks. In particular, sensorcameras on the UAVs are utilized to capture the current roadtraffic. By using this information, the UAVs can then predictthe road traffic conditions using the aforementioned deeplearning methods. Thus, from the traffic prediction, the UAVscan work as mobile road-side units to orchestrate road trafficfor over-crowding avoidance by informing the upcoming roadtraffic conditions to vehicular users via cellular networksaccordingly (Fig. 9(d)).

e: SICKNESS TREND PREDICTIONMachine-learning-based location prediction method is alsoof importance to predict the sickness trend in specific areas.This sickness trend prediction can be used to inform peo-ple to stay safe from possible infected places. For exam-ple, the work in [157] designs a contactless surveillance

framework, i.e., FluSense, to predict the influenza-like dis-ease 7-14 days before the real disease occurs in the hos-pital waiting areas. In particular, a set of real-time sensorsincluding a microphone array to detect normal speech/coughsounds and a thermal camera to detect crowd density areembedded into an edge computing platform. Consideringmillions of non-speech audio samples and hundred thousandsof thermal images for audio and image recognition models,the proposed framework can accurately predict the number ofdaily influenza-like patients with Pearson correlation coef-ficient of 0.95. The prediction model from this work canbe correlated/combined with the localized medical/healthinformation (e.g., from local hospitals/clinics) to furtherimprove the prediction accuracy as shown in Fig. 9(e).We canthen inform the local mobile users about the sickness trendprediction to avoid the potential areas where many influenza-like patients exist.

f: SYMPTOM DETECTION AND MONITORINGCoughing is one of the most common and detectable symp-toms of influenza pandemics. In the presence of a pandemic,the early detection of such symptoms can play a key role inlimiting the disease spread from the infected to the suscep-tible population. For example, if a coughing person can bedetected and identified in public places, that person and thepeople in close proximity can be tested for the disease.

In several studies, such as [158]–[161], AI technologies areleveraged to identify the cough patterns in audio recordingscollected from microphones or acoustic sensors. In [158],audio signals are analyzed using recurrent and convolutionalneural networks to detect coughs with high accuracy (upto 92%). Similarly, a hidden Markov model is proposedin [159] to detect cough from continuous audio recordings.In addition to audio signals, signals from motion sensors arealso analyzed in [160] by a novel classification algorithm.However, a common limitation of these approaches is thatthey require the sensors to be attached to the person, which isnot always possible in social distancing scenarios. To addressthis problem, a cough detection system is proposed in [161].This system utilizes a wireless acoustic sensors network con-nected to a central server for both cough detection and local-ization. In particular, when a sound is detected, the sensorsfirst localize the sound source by the AOA technique. Then,the sensors send the measured sound signals to the centralserver for cough identification using a novel classificationalgorithm. In the social distancing context, this system canbe applied directly to monitor and detect coughing peoplein public places. Nevertheless, a limitation of this system isthat the localization and measurement errors increase signif-icantly when the sound source is too far from the sensors.Besides coughs, other physiological metrics such as cardio-vascular activity, body temperature, and respiration can alsobe meaningful indicators. Especially, based on those metrics,an infected case can be potentially detected before clinicalsymptoms, e.g., fever, occur [162]. However, early detectionalgorithms need to be developed specifically for COVID-19.

VOLUME 8, 2020 154221


Summary: Various AI technologies can be leveraged tofacilitate social distancing implementations, especially in thescenarios that require modeling and prediction. In particular,AI technology can help to predict people’s locations, trafficdensity, and sickness trends. Moreover, AI-based classifica-tions algorithms can be utilized to detect symptoms such ascoughs in public areas.

III. OPEN ISSUES AND FUTURE RESEARCH DIRECTIONSIn this section, we discuss the open issues of social distanc-ing implementation such as security and privacy concerns,social distancing encouragement, work-from-home, and theincreased demands in healthcare appointments, home health-care services, and online services. To addressed these issues,potential solutions are also presented.

A. SECURITY AND PRIVACY-PRESERVING IN SOCIALDISTANCINGMost aforementioned social distancing scenarios (see Table 1for more details) call for people’s private information, to adifferent extent, ranging from their face/appearance to loca-tion, travel records, or health condition/data. These data,if not protected properly, attract cyber attackers and canturn users into victims of financial, criminal frauds, andprivacy violation [125]. Users’ data like health conditionscan also adversely impact people’s employment opportunitiesor insurance policy. Given that, to enable technology-basedsocial distancing, it is critical to develop privacy-preservingand cybersecurity solutions to ensure that users’ private dataare properly used and protected.

The general principle of users’ privacy-preserving isto keep each individual user’s sensitive information pri-vate when the available data are being publicly accessed.To do so, data privacy-preserving mechanisms includingdata anonymization, randomization, and aggregation canbe utilized [116]. For example, Apple, Google, and Face-book have developed people mobility trend reports whilepreserving users’ privacy during the COVID-19 outbreak.In particular, Apple utilizes random and rotating identi-fiers to preserve mobile users’ movements privacy [119].Meanwhile, Google aggregates and uses anonymized datasetsfrom mobile users who turn on their location history set-tings in their Android smartphones. In this case, a differ-ential privacy approach is applied by adding random noiseto the location dataset with the aim to mask individualidentification of a mobile user [117]. Similarly, Facebookutilizes aggregated and anonymized user mobility datasetsand maps to determine the mobility trend in certain areasincluding the social connectedness intensity among nearbylocations [118]. In addition to the Apple’s, Google’s, andFacebook’s latest privacy-preserving implementation, in thefollowing, wewill thoroughly discuss how the latest advancesin security and privacy-preserving techniques can help tofacilitate social distancing without compromising users’interest/privacy.

1) LOCATION INFORMATION PROTECTIONTo protect the exact location/trajectory information of par-ticipating mobile users in social distancing, some advancedlocation-based privacy protection methods can be adopted.Specifically, we can anonymize/randomize/obfuscate/perturbthe exact location of each mobile user to avoid maliciousattacks from the attackers using the following mechanisms.For example, the authors in [126] develop a privacy-preserving location-based framework to anonymizespatio-temporal trajectory datasets utilizing machine-learning-based anonymization (MLA). In this case, the frame-work applies the K -means machine learning algorithm tocluster the trajectories from real-world GPS datasets andensure theK -anonymity for high-sensitive datasets. Using theK -anonymity [127], [128], the framework can collect loca-tion information from K mobile users within a cloakingregion, i.e., the region where the mobile users’ exact locationsare hidden [129], [130]. In [131], the use of K -anonymityis extended into a continuous network location privacyanonymity, i.e., KDT -anonymity, which not only considersthe average anonymity size K , but also takes the average dis-tance deviationD and the anonymity duration T into account.Leveraging those three metrics, the mobile users under real-istic vehicle mobility conditions can control the changes ofanonymity and distance deviation magnitudes over time.

The authors of [132] propose a mutually obfuscating pathsmethod which allows the vehicles to securely update accuratereal-time location to a location-based service server in thevehicular network. In this case, the vehicles first hide theirIP addresses due to the default network address translationoperated by mobile Internet service providers. Then, theygenerate fake path segments that separate from the vehicles’actual paths to prevent the location-based service serverfrom tracking the vehicles. Exploiting dedicated short-rangecommunications (DSRC) among vehicles and road naviga-tion information from the GPS, the vehicles can mutuallygenerate made-up location updates with each other whenthey communicate with the location-based service server(to obtain spatio-temporal-related information). In [133],vehicles which use location-based services can dynamicallyupdate virtual locations in real-time with respect to the rela-tive locations of current nearby vehicles. This aims to providedeceptive information about the driving routes to attackers,thereby enhancing location privacy protection.

In addition to the anonymization and obfuscating meth-ods, randomization and perturbation are the methods thatcan be employed to protect user’s location privacy in socialdistancing scenarios. In [134], a location privacy-preservingmethod leveraging spatio-temporal events of mobile usersin continuous location-based services, e.g., office visita-tion, is investigated. Specifically, an �-differential privacyis designed to protect spatio-temporal events against attack-ers by adding random noise to the event data [137]–[139].In [140], the authors present a location privacy protectionmechanism using data perturbation for smart health systemsin hospitals. In particular, instead of reporting the patient’s

154222 VOLUME 8, 2020


FIGURE 10. Location-based privacy preserving for social distancing scenarios. In (a) location information protection, the exact location of a vehicle canbe obfuscated to protect people’s privacy. To protect (b) personal identity, a user can exchange its identity with nearby trusted users in each location, andthus that user cannot be identified by the attackers. For (c) health-related information protection, the health information can be anonymized.

real locations directly, a processing unit attached to a patient’sbody can adaptively produce perturbed locations, i.e., the rel-ative change between different locations of the patient. In thiscase, the system considers the patient’s travel directions andcomputes the distance between the patient’s current locationsand the patient’s sensitive locations (i.e., patient’s predefinedlocations which he/she does not want to reveal to anyone, e.g.,patient’s treating room). Using this dynamic location pertur-bation, the need for a trusted third party to store real locationscan be removed. Leveraging the aforementioned methods,we can also prevent the service provider from accessingmobile users’ and vehicles’ exact locations/trajectories/pathswhen they implement social distancing for crowd/trafficdensity and movement detection. Specifically, a platoonof mobile users/vehicles in a certain area can collaboratetogether to mix their real locations/trajectories/paths anony-mously (Fig. 10(a)). In this way, the service provider willonly obtain the aggregated location/trajectory/path informa-tion of the platoon instead of each individual’s exact loca-tion/trajectory/path for its location privacy.

2) PERSONAL IDENTITY PROTECTIONIn addition to protecting mobile users’ location-relatedinformation, preserving their personal identities is of impor-tance to improve users’ acceptance of the latest technolo-gies to social distancing. Specifically, we can exchange oranonymize personal identities among trusted mobile users toavoid the attackers identifying the actual identity of each indi-vidual user. In [141], the authors develop a pseudo-identityexchanging protocol to swap/exchange identity informationamong mobile users when they are at the same sensitivelocations, e.g., hospital and residential areas. In particular,when a mobile user receives another trusted user’s identityand private key, themobile user will verify if the encryption ofanother user’s identity hash function and public key is equalto the encryption of the received private key. If that condition

holds, the mobile user will change his/her identity with thatuser’s identity and vice versa.

Another method to protect personal identity in social dis-tancing scenarios is individual information privacy protec-tion through indirect- or proxy-request as proposed in [142].In particular, instead of directly submitting a request tothe server, a mobile user can have his/her social friendsthrough the available social network resources, i.e., trustedsocial media, to distribute his/her request anonymously tothe server. The request result can be returned to his/hersocial friends and then forwarded to the requested mobileuser, thereby preserving the requested mobile user’s iden-tity. In fact, there may exist some malicious friends whoexpose the identity of the mobile user. Therefore, the authorsin [143] investigate a user-defined privacy-sharing frameworkon social networks to choose his/her particular friends whoare trusted to obtain the mobile user’s identity information.In this case, the mobile user only shares his/her identityinformation with the particular friends whose pseudonymsmatch the mobile user’s identity through the authorizedaccess control. Using the same approaches from the aboveworks, we can use local wireless connections, e.g., Bluetoothand Wi-Fi Direct, to anonymously exchange actual locationinformation in a mobile user group, i.e., between a mobileuser and his/her trusted nearbymobile users, in an ad hocway.As shown in Fig. 10(b), when the service provider requiresto collect location-related information for the current crowddensity detection, a representative mobile user from the groupcan send the group’s anonymous location information to theservice provider, aiming at preserving the personal identity ofeach mobile user in the group.

Moreover, Apple and Google have recently introduced akey schedule for contact tracing to ensure the privacy ofusers [3]. Specifically, there are three types of key: (i) tracingkey, (ii) daily tracing key, and (iii) rolling proximity identifier.The tracing key is a 32-byte string that is generated by using

VOLUME 8, 2020 154223


a cryptographic random number generator when the app isenabled on the device. The tracing key is securely storedon the device. The daily tracing key is generated for every24-hour window by using the SHA-256 hash function withthe tracing key. The rolling proximity identifier is a privacy-preserving identifier which is sent in Bluetooth advertise-ments. This identifier is generated by using the SHA-256 hashfunction with the daily tracing key. Each time the BluetoothMAC address is changed, the app can derive a new identifier.When a positive case is diagnosed, its daily tracing keys areuploaded to a server. This server then distributes them to theclients who use the app. Based on this information, each ofthe clients will be able to derive the sequence of the rollingproximity identifiers that were broadcasted from the user whotested positive. In this way, the privacy of the users can beprotected because, without the daily tracing key, one cannotobtain the user’s rolling proximity identifier. In addition,the server operator also cannot track the user’s location orwhich users have been in proximity.

Similarly, several solutions have been proposed in [4], [5].The key idea of these solutions is generating a unique iden-tifier and broadcasting it to nearby devices. In particular,PACT [4] regularly (every few seconds) emits a data string,called chirps, generated by cryptographic techniques basedon the current time and the current seed of the user to ensurethe privacy. Similarly, in [5], the identifier EphID (calledephemeral ID) is created as follows:

EphID = PRG(PRF(SKt , broadcast key)

), (1)

where PRF is a pseudo-random function (e.g., SHA-256),broadcast key is a fixed and public string, and PRG is a streamcipher (e.g., AES in counter mode). SKt is the secret key ofeach user during day t which is computed as follows:

SKt = H (SKt−1), (2)

where H is a cryptographic hash function. Upon receivingthe identifier, other nearby devices will keep it as a log. If auser is diagnosed with the disease, other users who may haveencountered the infected person will receive a warning of apotential contact.

With outstanding performance in data integrity, decen-tralization, and privacy-preserving, blockchain technologycan be an effective solution to preserve privacy to enabletechnology-based social distancing scenarios. A blockchainis a distributed database shared among users in a decentral-ized network. This decentralized nature of blockchain ensuresits immutability property, i.e., the data stored within cannotbe altered without the consensus of the majority of networkusers [147]. Another advantage of blockchain technologyis that the users’ anonymity is ensured due to the public-private keys pair mechanism [148]. As a result, blockchaintechnology can effectively address the personal identity issuein social distancing scenarios where people have to share theirmovement and location information but not their exact iden-tities. For example, in the infected movement data scenario,we only need to know the movement path of a person, and

whether or not that person is infected. In this case, the personanonymity can be ensured with the public-private keys pairmechanism, since there is no way to link the public key tothat person’s true identity.

3) HEALTH-RELATED INFORMATION PROTECTIONTo monitor the sickness trend in a certain place, e.g., the hos-pital, for the social distancing purpose (i.e., to inform theupcomingmobile users not to enter a high-risk area/building),the health-related condition information of visiting mobileusers has to be shared to provide reliable learning dataset.To protect this highly sensitive information, the authorsin [144] propose a differential privacy-based protectionapproach to preserve the electrocardiogram big data by uti-lizing body sensor networks. In particular, non-static noisesare applied to produce sufficient interference along withthe electrocardiogram data, thereby preventing the maliciousattackers to point out the real electrocardiogram data.

To provide secure health-related information access forauthenticated users, a dynamic privacy-preserving approachleveraging the biometric authentication process is introducedin [145]. Specifically, when a user wants to access themedicalserver containing his/her health condition, a secure biometricidentification at the server for the user’s validity is employedwhere the exact value of his/her biometric template remainsunknown to the server. In this way, the personal identity of theauthenticated user can be preserved. To further enhance theanonymity of his/her medical information, the random num-ber that is used to protect the biometric template is updatedafter every successful login. Then, the authors in [146] pro-pose a secure anonymous authentication model for wirelessbody area networks (WBANs). Specifically, this frameworkenables both patients and authorized medical professionals tosecurely and anonymously examine their legitimacies priorto exchanging biomedical information in the WBAN sys-tems. Motivated by the above works, we can utilize mobiledevices, secure service provider, and the aforementionedprivacy-preserving approaches to anonymously collect peo-ple’s health condition information for illness monitoring inthe hospital/medical center (Fig. 10(c)). In this way, the socialdistancing through monitoring the sickness trend can beimplemented efficiently while preserving the sensitive infor-mation of the people in the illness areas.

B. REAL-TIME SCHEDULING AND OPTIMIZATIONIn the context of social distancing, real-time scheduling andoptimization techniques can play a key role in preventingan excessive number of people at a given place (e.g., super-markets, hospitals) while maintaining a reasonable Quality-of-Service level. Fig. 11 illustrates several social distancingscenarios where scheduling and optimization techniques canbe applied. In particular, proper scheduling can help reducethe number of necessary employees at the workplace and thenumber of patients coming to the hospital, thereby minimiz-ing the physical contacts among people. Moreover, trafficscheduling can help to reduce the peak number of vehicles

154224 VOLUME 8, 2020


FIGURE 11. Scheduling and optimization for several social distancingscenarios including reducing the simultaneous presence of employees(workforce scheduling), patients (healthcare appointment and homehealthcare scheduling), and traffic (traffic control). Moreover, networkresources can be optimized to meet surging demands on online serviceswhile more people are working remotely from home.

and pedestrians, and network resource optimization [71](e.g., network/resource slicing) can meet surging demandson online services while more people are working remotelyfrom home.

1) WORKFORCE SCHEDULINGWorkforce scheduling can help to limit the number of peo-ple at the workplaces while ensuring the necessary workis done. While working from home is encouraged in socialdistancing, some essential work requires people to be presentat the workplace for important tasks (e.g., health, trans-portation and manufacturing). Moreover, different types oftasks impose various constraints such as due date (time con-straints), dependence among tasks (precedence constraints),skill requirements (skill constraints), and limited resourcesusage (resource constraints) which further complicate thescheduling problem. For such scenarios, workforce schedul-ing techniques can be utilized to optimally align and reducethe number of required employees to practice social distanc-ing. In [72], a novel three-phase algorithm is proposed forworkforce scheduling to optimize the operational cost andservice level simultaneously. Another Genetic-Algorithm-based hybrid approach is presented in [73], which optimizesthe schedules of the workforce according to multiple objec-tives including urgency, skill considerations, and workload

balance. Similarly, in [74], a Mixed-Integer-Programming-based approach is developed to minimize the operational costwith consideration of skill constraints. It is worth noting thatthe main objective of these approaches is to minimize cost,which is not the highest priority in the context of socialdistancing. In [75], [76], and [77], several methods are pro-posed to optimize the workforce schedules with considerationof rotating shifts, which indirectly reduce the number ofemployees to a certain extent. Nevertheless, the main objec-tive of these approaches is reducing costs. Therefore, devel-oping techniques to reduce the physical contacts or distanceamong employees at the workplace is critical for workforcescheduling in social distancing scenarios.

2) MEDICAL/HEALTH APPOINTMENT SCHEDULINGBesides workforce planning, scheduling techniques can alsohelp to optimize healthcare services, especially healthcareappointments and home healthcare services, thereby decreas-ing unnecessary traffic and the number of patients com-ing to hospitals. Several approaches have been proposedto effectively schedule appointments. In particular, a localsearch algorithm is proposed in [78] to minimize patientwaiting times, doctor idle times, and tardiness (lateness).Moreover, a two-stage bounding approach and a heuristic arepresented in [79] and [81], respectively. However, a commonlimitation of these techniques is that they do not take intoaccount the uncertainties in the duration of the appointmentsand the possibility that the patient will not come to thescheduled appointment. To address that, the uncertainty inthe processing times (e.g., of surgeries) is considered by aconic optimization approach in [80]. Similarly, a multistagestochastic linear program is developed in [82] to minimizepatient waiting times and overtime, which takes into accountthe unpredictable appointment duration and unplanned can-cellations. Although there are many effective approaches tooptimize appointment scheduling, the open issue is to developtechniques that specifically minimize or control the numberof patients simultaneously coming to the hospitals tomaintaina suitable level of social distancing, similar to that of theworkforce scheduling scenario.

3) HOME HEALTHCARE SCHEDULINGSimilar to appointment scheduling, home healthcare ser-vices (HHS) can help to reduce the pressure on hospi-tals and traffic in the social distancing context. In [83],a multi-heuristics approach is proposed for HHS schedul-ing to minimize the total traveling times of HHS staff.An extended problem is presented in [84], where the objectivealso includes minimizing the tardiness and additional skillsand time constraints are considered. For this problem, localsearch-based heuristics are proposed in the paper. Anotherlocal search-based heuristic is proposed in [85] for HHSscheduling with the objective to minimize traveling times andoptimize Quality-of-Service while considering workload andtime constraints. In [86], a Genetic-Algorithm-based hybridapproach is proposed for HHS scheduling with uncertainty

VOLUME 8, 2020 154225


in patient’s demands to minimize transportation costs. Alsoaddressing uncertainties, a branch-and-price algorithm isproposed in [87] to minimize the traveling costs and delay ofservices while considering stochastic service times. Unlikein workforce planning and appointment scheduling, HHSscheduling techniques can be more effectively applied tosocial distancing scenarios because they can minimize thetraveling distances while ensuring Quality-of-Service.

4) TRAFFIC CONTROLScheduling techniques have also been applied for traffic con-trol. In social distancing scenarios, scheduling techniquescan help to regulate the traffic level, especially the num-ber of pedestrians. In [88], a novel scheduling algorithmis developed for traffic control, considering both vehiclesand pedestrians, to minimize the delays. Similarly, a macro-scopic model and a scheduling algorithm are proposed fortraffic control, which jointly minimize both the pedestriansand vehicle delays in [89]. Another scheduling approach isproposed in [90] that considers both pedestrians and vehi-cles. Different from the previously mentioned approaches,this approach only focuses on minimizing pedestrian delay.Although there is a vast literature on traffic scheduling tech-niques, the social distancing implications have not been takeninto account. For example, to maintain social distancing,a more meaningful objective would be to reduce/constrainthe peak number of pedestrians on the street at thesame time.

5) ONLINE SERVICES OPTIMIZATIONWhen social distancing measures are implemented, morepeople will be staying at home e.g., working from home.Physical meetings/gatherings will move to virtual platforms,e.g., webinars. That results in much higher Internet traffic andcorresponding virtual service demands (e.g., video streaming,broadcasting, and contents delivery). Therefore, optimizingonline services delivery is a challenging issue in the socialdistancing context. Fortunately, online services optimizationis a well-studied topic with a substantial body of supportingliterature.

For example, in [91], a novel algorithm is proposed to opti-mize the contents delivery process in a CDN semi-federationsystem. In particular, the algorithm optimally allocates thecontent provider’s demand to multiple Content Delivery Net-works (CDNs) in the federation. The results show that thelatency can be reduced by 20% during peak hours. Anothertechnique to reduce the delay and network congestion is edge-caching, which brings the contents closer to the networkusers (e.g., [92]). In [93], the performance of two edge-caching strategies, i.e., coded and uncoded caching, are ana-lyzed. Moreover, two optimization algorithms are developedto minimize the content delivery times for the two cachingstrategies.

Besides the contents delivery, the demands on videostreaming traffic are also much higher during social

distancing implementation because there are many peoplewho work from home. In that context, emerging networkingtechnologies can be an effective solution. For example,an architecture utilizing HTTP adaptive streaming [94]and software-defined networking technology is proposed toenable video streaming over HTTP. Moreover, a novel algo-rithm is developed to optimally allocate users into groups,thereby reducing communication overhead and leveragingnetwork resources. The results show that the proposed frame-work can increase video stability, Quality-of-Service, andresource utilization.

Scheduling and optimization are well-studied topics witha vast literature available, which can be utilized for differentsocial distancing scenarios such as workforce, healthcareappointment, home healthcare, and traffic scheduling, andoptimization of online services delivery. Nevertheless, exceptfor the home healthcare service scenario, the existing tech-niques’ objectives do not align with the objectives of socialdistancing. Moreover, scheduling algorithms are often devel-oped such that they are only efficient for specific problems.Therefore, developing novel optimization/scheduling algo-rithms in operations research and adopting social distancingas a new performance metric or design parameter is verymuch desirable. Furthermore, the optimization of Internet-based services such as content delivery can help to encouragepeople to stay at home during social distancing periods byensuring the service levels.

C. INCENTIVE MECHANISM TO ENCOURAGE SOCIALDISTANCINGDue to the people’s self-interested/selfish nature character-istics in their daily life [163] (especially during the pan-demic outbreak), incentive mechanisms can be very helpfulin encouraging people to accept or share relevant informa-tion to enable new social distancing methods. These mech-anisms have been thoroughly discussed in crowdsourcingas implemented in [135], [168]–[171]. Therein, the serviceproviders can provide incentives to a large number of peopleto attract their contributions in data collection for crowd-sourcing processes. For example, the contract theory-basedincentivemechanism for crowdsourcing is discussed in [168],[169]. In particular, this approach is considered an effi-cient mechanism to leverage common agreements betweenthe participating entities, e.g., a service provider and itsmobile users, in a certain area under complete and incom-plete information from the participants [164]. The use of agame theory-based incentive mechanism to encourage a setof mobile users to form a crowdsourcing community networkis investigated in [135], [170]. Then, in [171], the authorsutilize an auction theory-based approach incentive mech-anism to stimulate mobile users’ participation in crowd-sourcing tasks such as traffic monitoring. In the following,we also highlight the existing incentive mechanisms and howthey can be further adopted to encourage social distancingapplications.

154226 VOLUME 8, 2020


1) DISTANCE BETWEEN ANY TWO PEOPLE AND DISTANCETO/FROM CROWDSTo motivate people to keep safe distances from themselves toothers, contract theory-based incentive models via D2D com-munications, e.g., Bluetooth, Wi-Fi Direct, can be employed.In [165], the authors propose a contract theory-based mech-anism to provide a higher reward for D2D-capable mobileusers if they send the information to a requesting mobileuser with a higher transmission data rate. Taking into accountthe number of potential nearby mobile users in proximity,the authors in [166] introduce the same mechanism suchthat a mobile user will receive a higher payment if theycan share the information with more nearby users. Likewise,the same approach considering a higher reward for a mobileuser who has shorter distances in sharing its informationto nearby D2D pairs is presented in [167]. Inspired by the

Documents

A Comprehensive Survey of Enabling and Emerging Technologies … · 2020. 12. 15. · C. T. Nguyen et al.: Comprehensive Survey of Enabling and Emerging Technologies for Social Distancing