Francesco Devoti, Vincenzo Sciancalepore, Xavier Costa-Perez
©SHUTTERSTOCK.COM/ZUBADA
The pandemic outbreak has profoundly changed our life, especially our social habits and communication behaviors. While this dramatic shock has heavily impacted human interaction rules, novel localization techniques are emerging to help society in complying with new policies, such as social distancing. Wireless sensing and machine learning are well suited to alleviate virus propagation in a privacy-preserving manner. However, their wide deployment requires cost-effective installation and operational solutions.
In public environments, individual localization information, such as social distancing, needs to be monitored to avoid safety threats when rules are not properly observed. To this end, the high penetration of wireless devices can be exploited to continuously analyze and learn the propagation environment, thereby passively detecting breaches and triggering alerts if required. In this article, we describe a novel passive and privacy-preserving human localization solution that relies on the direct transmission properties of millimeter-wave (mm-wave) communications to monitor social distancing and notify people in the area in case of violations, thus addressing the social distancing challenge in a privacy-preserving and cost-efficient manner. Our solution provides an overall accuracy of about 99% in the tested scenarios.
The COVID-19 pandemic has turned everyone’s life upside down, forcing national governments around the world to take measures to drastically reduce the rate of contagion. Localization information may support social distancing as well as the contact tracing of infected people, thereby playing a fundamental role in the fight against the virus’ spread. The scientific community proposed social distancing as the first effective measure against uncontrolled virus spread, able to stop the transmission chains of the virus and prevent new ones from appearing. This has resulted in strict rules to limit personal contact and maintain interpersonal distance. However, in many situations (offices, grocery stores, shops, and so on), guaranteeing social distancing may be challenging, as people cannot continuously measure their interpersonal distance, especially if placed in small rooms and indoor environments, and some people might be less careful than others in complying with social distancing.
This has potentially raised industrial interest for developing advanced solutions to issue notifications, through alert messages, about minimum distance violations. The solutions are mainly designed as infrastructure-free approaches that rely on wearable sensors based on ultrawideband technology (UWB) and smartphones provided with Bluetooth Low Energy (BLE) or near-field communication (NFC), which determine (by means of power measurements) whether the intersocial distance is below a certain threshold, thereby notifying neighbors about a potential threat. In addition, identity and contact duration information might result in privacy regulations violations [1]. The main drawback of such solutions is that the ability to detect interpersonal distance is limited to people with such specific devices (and running given applications). This assumption might not hold in many cases, as neither telecommunications operators nor national governments can force people, in general, to wear such devices and install such applications [2].
Therefore, pervasive wireless passive solutions are preferred, as they gather localization information to detect undesired contacts and trigger exposure alerts in a privacy-preserving manner. This information can be used for many other use cases, including intrusion detection, privacy-preserving health applications, location-based services, and so on. The development of advanced wireless sensing techniques is thus taking a prominent role in this space since variations in wireless propagation signals can be exploited for such novel services. Nonetheless, such applications generally require detailed and instantaneous information on the channel status, thus calling for specific radio-frequency hardware that can directly expose the information via programmable interfaces [3]. When only incomplete or partially hidden data are available, artificial intelligence (AI) might be helpful: designed neural networks (NNs) can continuously study the obtained pairwise channel information, learn unexpected variations, and proactively map unusual fluctuations onto deterministic system state changes, i.e., the Boolean state of social distancing violations in open and closed spaces.
In this work, we leverage the narrow directivity of mm-wave wireless communications to pioneer a new passive AI-based sensing system capable of instantaneously detecting human gatherings violating social distancing policies and sending corresponding alerts to warn people. The proposed system does not rely on dedicated wearable devices, applications, and active connections; thus, it overcomes the limitations of infrastructure-free solutions. At the same time, the system leverages reinforcement learning to automatically learn the human gathering detection task based on the observation of the reactions of people upon receiving the alert. Hence, our passive solution is privacy guaranteed and does not require human intervention. To validate our proposed framework, we carry out a synthetic trace simulation campaign and realistic deployment-based experiments with commercial mm-wave devices in selected environments. The proposed system can be seamlessly installed on existing Wi-Fi network deployments in public and hybrid environments, such as rail stations (as depicted in Figure 1), airport terminals, bus stops, and so forth, with limited installation and maintenance costs.
Figure 1 The passive detection of social distancing in public scenarios, using mm-wave off-the-shelf devices provided with AI-based algorithms.
Localization information-based applications, such as social contact tracing, have become a key objective to effectively keep under control unwanted human virus spreading. However, this has proved to be a daunting task, due to the simultaneous lack of the following constraints:
We detail, in the following, the main effort in the literature toward those essential challenges, mostly taken individually.
Computer vision and image processing from surveillance and dedicated cameras offer means to constantly monitor social distancing enforcement in public environments. On this line, [4] exploits the deep learning technique to efficiently carry out the object recognition process that can identify (and automatically locate) humans in video sequences and, in turn, estimate the Euclidean distance among people. This is approximately performed by counting the pixels within the snapshots where people are detected. Similarly, [5] proposes a deep learning-based framework that analyzes the data from mass video surveillance to monitor social distancing and trigger, in necessary cases, instantaneous alerts. Such solutions are proved to be very effective in terms of high reliability and extreme accuracy (c1 and c2), but they require the installation of appropriate cameras, which might not be allowed in many everyday environments (e.g., offices, factories, hospitals, and schools), and might even incur untenable costs.
In parallel, the very limited energy consumption of small (and sometimes wearable) devices enables a focus on infrastructure-free approaches capable of building low-cost and flexible ad hoc networks (c3 and c4). Attaining such desirable objectives precludes relying on a supporting (and fixed) infrastructure while pushing for widespread technologies, such as BLE, NFC, and UWB sensors. In particular, [6] provides specific means to combat virus spread by minimizing the human approaching time, e.g., with ultrasound-based proximity detection systems and by means of wearable magnetic field proximity sensors. Finally, [7] blends together the need for accurate group tracking models with the infrastructure-free requirement that helps to detect contagion-related misbehavior within whatever environment. On top of dedicated hardware, tracing applications may leverage complex wearable devices, such as smartphones, tablets, and smart watches, to query public repositories, thus notifying all direct human contacts about potential (diffusion) threats [8].
Most of the data thus collected must be carefully (and efficiently) analyzed to rapidly block uncontrolled diffusion (c1 and c4). This requires proper and complex mathematical models [9] (c2) that present scalability issues for very crowded scenarios. To cope with the complexity of the scenario, machine learning can be exploited to facilitate good approximations and to reconstruct hidden and unavailable data [10]. The preceding techniques may help cover all presented challenges and thus prevent infections only if combined together. Generally, this does not apply in realistic contexts, highlighting the need for a sustainable solution that can trade off all described features. Our pioneering proposal paves the road toward an agile and flexible framework that can be readily installed on existing wireless infrastructures, without requiring people to wear electronic devices but, at the same time, providing the comparable accuracy and reliability levels of an infrastructure-free system.
While existing solutions might achieve reliability, fast actuation, and high responsiveness, exclusively, they can be fatally impaired by human misbehaviors. Therefore, there is a compelling need for an agile and flexible solution that is able to pursue high accuracy with affordable installation costs. Hereafter, we detail our solution, which relies on conventional mm-wave communication, by shedding light on the mathematical models and implementation details.
Social distancing breaches can be ideally spotted by continuously monitoring the surrounding propagation environment to promptly detect suspicious variations. This operation can be performed in a passive way wherein a transmitter pair interacts and keeps track of the channel response. Specifically, after establishing the mm-wave link, power measurements can be regularly collected and analyzed to detect unexpected changes. To be compliant with IEEE 802.11ay standard guidelines (please refer to https://standards.ieee.org/standard/802_11ay-2021.html), power measurements are regularly performed during the beam training phase, i.e., when two mm-wave devices discover each other by selecting the transmitting beam (direction) based on best response channel quality [11].
IEEE 802.11ay (like its predecessor, IEEE 802.11ad) covers many relevant aspects to establish and sustain a communication link among mm-wave-enabled devices. To provide the required beamforming capabilities, such devices are equipped with electronically steerable antenna arrays controlled by predefined weights vectors that are included in the so-called codebook. Each wave vector in the codebook corresponds to the activation of a specific transmitting/receiving beam pattern. Those beam patterns are designed to be directional, and their choice is subject to the instantaneous channel condition: a beam adaptation process is executed to avoid (nomadic) obstacles and efficiently follow channel variations so that communication is never disrupted.
The beam pattern selection is performed by means of a complex beamforming training phase, wherein devices activate sequentially all available beam patterns as per their codebook and correspondingly collect power measurements that are used to select the best transmitting direction. This is started during the initial connection establishment (i.e., the device pairing phase) and periodically repeated to avoid connection drops [12]. In parallel, the wide spatial diversity provided by all available beam patterns allows obtaining a complete snapshot of the propagation environment, as shown in the “Experimental Evaluation” section, inspecting the surrounding area and keeping track of potential state changes.
The high-frequency feature of mm-wave communication hinders the overall signal propagation, which appears strongly affected by the propagation environment itself, including human bodies, walls, and even glass objects [13]. On the one hand, this aspect may require an additional effort for properly designing a mm-wave-based network to guarantee affordable communication quality levels. On the other hand, a passive environment monitoring system may capitalize on this issue to build a reliable and low-cost solution that uses the effect on the short wavelength signal to continuously monitor selected areas. In particular, the different displacement of people in the environment can radically change the propagation conditions experienced by mm-wave device pairs. Such changes are reflected in the outcome of the beam training procedure: we exploit such power measurements performed during standard operations to devise a complete sensing map of the propagation environment that can be smartly used to retrieve information on the environment itself, without requiring the extraction of advanced channel state information from the devices and deploying additional dedicated hardware. Hence, it will dramatically lower implementation costs.
Figure 2 provides an example of different beam training phase outcomes as a function of the displacement of the people in the monitored area. The example shows how the position of potential signal blockers (i.e., people) has an impact on the power measurements related to different beam activations. Indeed, beams with directions toward potential signal blockers will experience larger attenuation with respect to the beams pointing toward unblocked paths, which translates into huge power variations reported in the beam training outcome. We leverage this feature to build a system capable of detecting and reporting safe distance violations in areas covered by a mm-wave transmission service. The rationale behind this is related to the beam training phase periodically performed among deployed devices (i.e., without involving user equipment), which we use to understand where blockages occur, based on selected transmitting/receiving beam and consequent directions, so as to infer the mutual distance between people. Naturally, as we show in the “Experimental Evaluation” section, the more directive the selected beam patterns, the higher the granularity of the environment sensing map, and consequently, the higher the sensitivity of the system against the position of the blocks, the higher the accuracy of the social distance detection model.
Figure 2 The effect of the interpersonal distance on the beam training outcome. (a) Social distance observance. (b) Social distance violation.
Analytically modeling the effect of interpersonal distance on power measurements appears very challenging and biased due to the relevant dependence of the mm-wave channel on the propagation environment. It is necessary to rely on machine learning techniques capable of automatically approximating the link between the measured power and distance violations. This outstanding dependence on the environment turns into huge differences even in those areas that are relatively close to each other. Therefore, we analyzed a bench of classifiers specifically trained for each alert area.
Our safe distance violation detection and alert system is depicted in Figure 3. Let us consider a public area (e.g., a train station, airport, shopping mall, office, and so on) wherein people, in accordance to the virus spread prevention measures, must keep a minimum safe distance and wherein high-speed connectivity service is provided through IEEE 802.11ay-enabled access points (APs) (see Figure 1).
Figure 3 The safe distance violation detection and alert system. The system collects power measurements from the device beam training phase, maps the obtained measurements onto the device deployment to define alert areas, and performs AI-driven human gathering estimation. Training is performed via reinforcement learning. RSSI: received signal strength indicator; SNR: signal-to-noise ratio; FFNN: feedforward NN. (Top) AI-based models. (Bottom) Alert areas.
APs can provide connectivity to mobile stations (e.g., smartphones, laptops, and so forth) as well as fixed stations (e.g., wireless displays, computers, cameras, and similar devices). Our system exploits the channel variations detected by mm-wave devices. Those variations can be caused both by different displacements of blockers and movements of involved devices. Therefore, to filter out the device mobility effect, only devices that are fixedly deployed within the monitored area are considered an information source for the system. Moreover, we assume the system to be deployed in controlled environments, such as rail stations, shopping malls, and so on, wherein the movements of nomadic obstacles (e.g., trains, buses, and other vehicles) are periodically repeated in a quasi-deterministic fashion. Accordingly, their effect on the power measurements would affect many observations in the input of the classifier, which, given the high number of examples, would automatically filter out the contribution of such objects in the classification process.
According to the standard, devices in the area periodically activate the beam training procedure and perform power measurements. Note that APs and stations can detect and collect beam training frames transmitted from nearby devices even if they are executing the connection handshake process. The power measurements thus obtained are transmitted through a control plane to the safe distance violation system and collected to build a snapshot of the propagation environment state. Given the limited coverage of mm-wave APs, only a portion of the monitored area is considered relevant for the power measurement campaign related to a given pair of devices. Thus, we map all collected measurements onto the corresponding portion of the monitored area, based on device deployment and device coverage (assumed to be known), namely, the alert area. This allows the system to send targeted alerts only to the specific areas wherein safe distance violations occur. Note that we assume an optimal deployment phase being executed beforehand, ensuring good channel conditions among devices to enable passive sensing: this enables strong communication paths, i.e., line-of-sight and/or second-order reflections, among devices covering a given alert area. Moreover, the denser the device deployment, the higher the granularity of the sensing/reflecting in smaller and more accurate target areas.
Such power measurements become the input feature of our detection system. For each alert area, the corresponding power measurements are collected and arranged into a feature vector whose elements contain power measurements corresponding to each activated beam pair and are thus normalized. Feature vectors are fed into a bank of feedforward NN (FFNN) classifiers (one per alert area), which is in charge of performing safe distance violation detection; i.e., given the observation of the power measurements, it provides the probability distribution over the classes of safe distance violation and safe distance observance. We select FFNNs due to their relatively low complexity and ability for nonlinear model generalization. However, different types of classifiers can be easily plugged into the proposed framework, depending on the desired system efficiency and complexity. It is worth pointing out that the lack of an active connection with the people in the area prevents our system from sending targeted notifications. Instead, if a violation is detected, a notification is sent to the specific alert area through advising systems, e.g., voice warnings and display boards. However, in situations where a group of people is not required to spread apart, e.g., people belonging to the same household, an unwanted alert might be triggered. Nonetheless, the alert is eventually sent to security officers in charge of evaluating the situation and enforcing safety policies within the target area, if needed.
The bank of FFNN classifiers needs to be properly trained to detect safe distance violations. To this extent, we rely on a reinforcement learning approach that exploits people’s reaction to an alert to reveal the correctness of notified alerting messages, and based on this, the technique automatically learns how to truly detect safe distance violations. The rationale behind this approach is as follows. In the event of a safe distance violation detection, an alert is sent to the corresponding alert area. If the detection is correct, people in the area will automatically react to the alert by rearranging themselves to return to a safe condition (the Hawthorne effect). This reaction reflects a noticeable change of the propagation environment that can be clearly captured in the power measurement snapshots following the alert. Conversely, if the safe distance violation is incorrectly detected, the sent alert will not push people to move far away. Consequently, an imperceptible change in the snapshots following the alert will occur.
In the reinforcement learning architecture we propose, the actor network is in charge of classifying safe distance violations, while the critic is in charge of monitoring the channel variations following an alert and providing a reward to the actor. For this reason, we design the critic as a deep convolutional NN, which is fed with the series of feature vectors following the alert concatenated over time and, involving the convolutional layers, particularly suitable to recognize the space–time features characterizing the movement of people from its input. This mechanism facilitates automatically learning how to correctly detect safe distance violations independently from the environment in which the system is deployed. It is worth highlighting that the learning mechanism we propose keeps unaltered the level of privacy guaranteed by our passive solution. Moreover, although relying on an instinctive reaction, such as the Hawthorne effect, which, in some cases, could be ignored, alert messages surely reach people in charge of guarding the monitored area, thus forcing lawbreakers to return to a safe situation.
Hereafter, we provide the performance evaluation of our system, which we carry out through a simulation campaign, where we generate synthetic beam training phase traces, as well as through a real implementation of our solution as a software module installed in four commercial IEEE802.11ad-enabled devices deployed within a real office environment.
For the beam training synthetic traces generation, we emulate the system by means of an ad hoc MATLAB simulator. We consider four different experimental environments, as follows:
mm-Wave propagation and beam patterns are modeled as per [14]. Moreover, to take into account the typical nonidealities of beam patterns generated by commercial devices [12], we vary the half-power width of the beams W according to the equation ${W} = {2}{\pi}\left[{{1}{-}\left({{1}{-}\left({{1} / {N}}\right)}\right){\alpha}}\right]$, where N is equal to the number of transmitting/receiving beam configurations in the codebook and ${\alpha}$ is a scaling factor that allows us to modulate the directionality in our experiments and ranges from a maximum value of one, resulting in $\left({{{2}{\pi}} / {N}}\right) {-} {wide}$ beams corresponding to 100% directionality, to a minimum value of zero, resulting in a beam width of ${2}{\pi}$ corresponding to 0% directionality, i.e., omnidirectional beam patterns. We consider devices with 32 transmitting beams and one, three, or six receiving beams. Following the IEEE 802.11ad standard, the beam pattern alignment procedure is performed at every beacon interval (i.e., every 10 ms) per device pair in both the transmitting and receiving directions.
People’s bodies are emulated as fully absorbing cylinders with a radius of 0.25 m. We randomly drop up to six people in the simulation playground. The minimum safe distance to be kept is set to 1.5 m, as usually recommended by European health-care institutions to reduce the spread of the virus. We consider a total of 200 different people arrangements in which social distancing regulations are violated and an additional 200 people arrangements in which the minimum social distance is fulfilled. For each people displacement, we collect the power measurements of 200 beam training procedures, for a total of 80,000 channel measurement snapshots that are arranged to form input feature vectors and normalized via independent standardization. The overall dataset of snapshots is split according to a 60/20/20 ratio for the purposes of training, validation, and testing procedures, respectively. The classification process is performed by a fully connected FFNN with a single hidden layer of neurons with a rectified linear unit activation function. We train our NN with a batch size of 1,000, 30 epochs, a learning rate of 0.001, and an Adam optimizer. Figure 4 demonstrates the performance of our system in terms of the safe distance violation detection accuracy achieved in the different scenarios we consider, with different beam pattern configurations both in terms of the number of transmitting/receiving beam patterns available as well as beam directionality. We consider different NN complexities by varying the number of neurons forming the hidden layer.
Figure 4 The accuracy of the distance violation detection process in four different simulated environments, considering different numbers of beams available at the devices N and varying the beam directionality from a maximum of 100% (the half-power beamwidth, equal to 2π/N) to a minimum of 0% (omnidirectional). The environments include (a) an office, (b) a hall, (c) underground, and (d) a train station. Tx: transmitter; Rx: receiver.
From the results, it can be seen how the directional capabilities of the devices have a direct impact on the achievable performance. Indeed, the beam directionality directly affects the directional sensing effectiveness of the system, as the more directional the beams, the higher the granularity of the sensing map available to the system and, thus, the better the achievable performance (the higher the system accuracy). Additionally, the number of available beam patterns also affects the overall system performance. Indeed, increasing the number of beams increases the different points of view that the system can exploit to efficiently run the violation detection process, with a consequent performance increase. On the AI algorithm, results show that different NN complexities (i.e., the number of neurons in the hidden layer) slightly change the system performance when a high number of beams with high directionality is available, though this has a greater impact when the devices are equipped with fewer beam patterns. This is a direct consequence of the quality of the input features in relation to the devices’ directional capabilities, as previously described. The better the quality of the input features, the easier the detection task and vice versa. This is reflected in the NN complexity required to achieve high detection performance.
Finally, the results show how the different considered scenarios impact the system performance: the wider the monitored area, the lower the system performance. Recalling that we are keeping constant the number of deployed devices in the simulated alert areas, this behavior is mainly due to the density of the devices involved in the measurement process, which naturally affects the system performance. Nonetheless, our system shows outstanding performance that varies from 75% to 99%, depending on the selected scenario.
To validate our system in a real environment, we implement our solution as a software module running on commercial off-the-shelf devices, thus considering realistic beamforming patterns and propagation conditions. In particular, we deploy four Talon ad7200 devices in a real office environment. The devices are provided with 36 transmitting beams with different pointing directions and one quasi-omnidirectional receiving beam. The default device firmware does not include easy access to the beam training and received power values; therefore, we use LEDE-ad7200 custom firmware [12] on such devices that allows us to retrieve the power measurements performed by the devices during the beam training phase and use the measurements to reveal infractions of the minimum safe distance. To comply with current social distancing regulations, we emulate the presence of two people in the office, with a real person and a human phantom. Figure 5 depicts our testbed implementation, wherein devices are mounted on the ceiling according to the office map. Additionally, we consider a second deployment setup, where the two devices on the right-hand side of the office map are placed on top of the corresponding desk. We refer the reader to [15] for additional information on the testbed.
Figure 5 The executive experiment set in a real office environment with four IEEE802.11ad devices mounted on the ceiling and a human phantom. The proposed device deployment at the top-right can efficiently and accurately cover the overall office environment.
To validate our system per each deployment setup of the devices, we consider a total of 40 different dispositions of the person and human phantom in the office: 20 with safe distance violations and the remaining 20 with sufficient interpersonal distances. For each disposition, we collect the beam training measurements performed by the devices for 3 min, thus obtaining 2 h of measurements with about 80,000 beam training procedures. The measurements thus collected are divided according to a 60/20/20 ratio for the purposes of training, testing, and validation. We report in Figure 6 the performance obtained with an FFNN classifier with a single hidden layer constituted by 32 neurons. We follow the same training process as described in the “Synthetic Scenario” section.
Figure 6 The accuracy of the distance violation detection process obtained in a real office environment. (a) Two different configurations of the device deployment. (b) Different areas of the office, with the devices mounted on the ceiling.
From the obtained results, it can be noticed that while our passive detection system can be easily installed on an existing Wi-Fi network, resulting in cost-effective, reliable, and agile deployment, it is able to spot safe distance violations with an accuracy higher than 99% when the devices are mounted on the ceiling. Such performance is maintained all over the office area, with a minimum variation depending on the relative proximity of the devices and monitored persons. The overall accuracy is slightly lower when the devices are placed according to the top desk setup, wherein office furniture (e.g., screens) causes higher attenuation among deployed devices and makes channel variations caused by the presence of people more difficult to sense. Nevertheless, the overall accuracy in our settings is higher than 98%. However, in line with the simulations, the accuracy performance might slightly reduce in other realistic deployments.
The COVID-19 virus explosion turned everyone’s life upside down. Social distancing proved to be an effective measure to control the virus’ spread, but unfortunately, compliance was difficult for people, due to deeply rooted social habits. In this article, we proposed an AI-based mm-wave sensing solution that, by passively monitoring changes in a wireless environment, can infer localization information, thereby detecting social distancing breaches and triggering corrective actions in a privacy-preserving manner without requiring an active connection with user equipment.
Our advanced proposal combines cost efficiency, agility, reliability, and accuracy challenges together into a novel passive detection system that can be supported by a variety of applications. The solution has been evaluated through a simulation campaign and a real deployment with commercial mm-wave devices. Proof-of-concept results show a promising detection accuracy of social distancing above 99%. The system has been designed such that it can be seamlessly added as a software module to off-the-shelf commercial mm-wave devices.
This work was supported by the European Union (EU) Horizon 2020 RISE-6G project (grant 101017011), MINECO/NG EU (grant TSI-063000-2021-6), and Council for Electronic Revenue Communication Advancement Program.
Francesco Devoti (francesco.devoti@neclab.eu) is a senior research scientist in the 6G Network group, NEC Laboratories Europe, 69115 Heidelberg, Germany. His research interests include reflective intelligent surfaces, millimeter-wave technologies in 5G and 6G networks, and network slicing. He received his Ph.D. degree in information technology from Politecnico di Milano in 2020. He is a Member of IEEE.
Vincenzo Sciancalepore (vincenzo.sciancalepore@neclab.eu) is a senior 5G researcher at NEC Laboratories Europe, 69115 Heidelberg, Germany. His research interests include network virtualization and network slicing challenges. He received his Ph.D. degree in telecommunications engineering and telematics engineering in 2015. He is the industrial chair of the IEEE Emerging Technologies Initiative on Reconfigurable Intelligent Surfaces, an editor of IEEE Transactions on Wireless Communications, and a Senior Member of IEEE.
Xavier Costa-Perez (xavier.costa@ieee.org) is the head of Beyond 5G Networks R&D at NEC Laboratories Europe, 69115 Heidelberg, Germany; the scientific director at the i2Cat R&D Center, 08034 Barcelona, Spain; and a research professor at the Catalan Institution for Research and Advanced Studies, 08034 Barcelona, Spain. He received his Ph.D. degree in telecommunications from the Polytechnic University of Catalonia, Barcelona, Spain. He is a Senior Member of IEEE.
[1] E. Y. Chan and N. U. Saqib, “Privacy concerns can explain unwillingness to download and use contact tracing apps when COVID-19 concerns are high,” Comput. Hum. Behav., vol. 119, p. 106,718, Jun. 2021, doi: 10.1016/j.chb.2021.106718.
[2] K. Soltesz et al., “The effect of interventions on COVID-19,” Nature, vol. 588, no. 7839, pp. E26–E28, Dec. 2020, doi: 10.1038/s41586-020-3025-y.
[3] P. S. Ranaweera, M. Liyanage, and A. D. Jurcut, “Novel MEC based approaches for smart hospitals to combat COVID-19 pandemic,” IEEE Consum. Electron. Mag., vol. 10, no. 2, pp. 80–91, Mar. 2021, doi: 10.1109/MCE.2020.3031261.
[4] I. Ahmed, M. Ahmad, J. J. Rodrigues, G. Jeon, and S. Din, “A deep learning-based social distance monitoring framework for COVID-19,” Sustain. Cities Soc., vol. 65, p. 102,571, Feb. 2021. [Online] . Available: http://www.sciencedirect.com/science/article/pii/S2210670720307897, doi: 10.1016/j.scs.2020.102571.
[5] M. Shorfuzzaman, M. S. Hossain, and M. F. Alhamid, “Towards the sustainable development of smart cities through mass video surveillance: A response to the COVID-19 pandemic,” Sustain. Cities Soc., vol. 64, p. 102,582, Jan. 2021. [Online] . Available: http://www.sciencedirect.com/science/article/pii/S2210670720308003, doi: 10.1016/j.scs.2020.102582.
[6] C. T. Nguyen et al., “A comprehensive survey of enabling and emerging technologies for social distancing—Part II: Emerging technologies and open issues,” IEEE Access, vol. 8, pp. 154,209–154,236, 2020, doi: 10.1109/ACCESS.2020.3018124.
[7] A. Biri, N. Jackson, L. Thiele, P. Pannuto, and P. Dutta, “SociTrack: Infrastructure-free interaction tracking through mobile sensor networks,” in Proc. 26th Annu. Int. Conf. Mobile Comput. Netw., ser. MobiCom’20, 2020, pp. 1–14, doi: 10.1145/3372224.3419190.
[8] D. Leith and S. Farrell, “Contact tracing app privacy: What data is shared by Europe’s GAEN contact tracing apps,” in Proc. IEEE INFOCOM 2021 – IEEE Conf. Comput. Commun., doi: 10.1109/INFOCOM42981.2021.9488728.
[9] Y. C. Chen, P. E. Lu, C. S. Chang, and T. H. Liu, “A time-dependent SIR model for COVID-19 with undetectable infected persons,” IEEE Trans. Netw. Sci. Eng., vol. 7, no. 4, pp. 3279–3294, Oct./Dec. 2020, doi: 10.1109/TNSE.2020.3024723.
[10] D. Vekaria, A. Kumari, S. Tanwar, and N. Kumar, “ε boost: An AI-based data analytics scheme for COVID-19 prediction and economy boosting,” IEEE Internet Things J., vol. 8, no. 21, pp. 15,977–15,989, Nov. 2021, doi: 10.1109/JIOT.2020.3047539.
[11] S. Aggarwal, M. Ghoshal, P. Banerjee, D. Koutsonikolas, and J. Widmer, “802.11ad in smartphones: Energy efficiency, spatial reuse, and impact on applications,” in Proc. IEEE INFOCOM 2021 – IEEE Conf. Comput. Commun., pp. 1–10, doi: 10.1109/INFOCOM42981.2021.9488763.
[12] D. Steinmetzer, D. Wegemer, M. Schulz, J. Widmer, and M. Hollick, “Compressive millimeter-wave sector selection in off-the-shelf IEEE 802.11ad devices,” in Proc. 13th Int. Conf. Emerg. Netw. EXperiments Technol., 2017, pp. 414–425, doi: 10.1145/3143361.3143384.
[13] C. Slezak, V. Semkin, S. Andreev, Y. Koucheryavy, and S. Rangan, “Empirical effects of dynamic human-body blockage in 60 GHz communications,” IEEE Commun. Mag., vol. 56, no. 12, pp. 60–66, Dec. 2018, doi: 10.1109/MCOM.2018.1800232.
[14] F. Devoti, I. Filippini, and A. Capone, “Mm-wave initial access: A context information overview,” in Proc. 2018 IEEE 19th Int. Symp. ‘A World Wireless, Mobile Multimedia Netw.’ (WoWMoM), pp. 1–9, doi: 10.1109/WoWMoM.2018.8449790.
[15] F. Devoti, V. Sciancalepore, I. Filippini, and X. Costa-Perez, “PASID: Exploiting indoor mmwave deployments for passive intrusion detection,” in Proc. IEEE INFOCOM 2020 – IEEE Conf. Comput. Commun., pp. 1479–1488, doi: 10.1109/INFOCOM41043.2020.9155401.
Digital Object Identifier 10.1109/MVT.2022.3202025