An Inclusive Survey of Contactless Wireless Sensing: A Technology Used for Remotely Monitoring Vital Signs Has the Potential to Combating COVID-19

With the Coronavirus pandemic showing no signs of abating, companies and governments around the world are spending millions of dollars to develop contactless sensor technologies that minimize the need for physical interactions between the patient and healthcare providers. As a result, healthcare research studies are rapidly progressing towards discovering innovative contactless technologies, especially for infants and elderly people who are suffering from chronic diseases that require continuous, real-time control, and monitoring. The fusion between sensing technology and wireless communication has emerged as a strong research candidate choice because wearing sensor devices is not desirable by patients as they cause anxiety and discomfort. Furthermore, physical contact exacerbates the spread of contagious diseases which may lead to catastrophic consequences. For this reason, research has gone towards sensor-less or contactless technology, through sending wireless signals, then analyzing and processing the reflected signals using special techniques such as frequency modulated continuous wave (FMCW) or channel state information (CSI). Therefore, it becomes easy to monitor and measure the subject’s vital signs remotely without physical contact or asking them to wear sensor devices. In this paper, we overview and explore state-of-the-art research in the field of contactless sensor technology in medicine, where we explain, summarize, and classify a plethora of contactless sensor technologies and techniques with the highest impact on contactless healthcare. Moreover, we overview the enabling hardware technologies as well as discuss the main challenges faced by these systems. INDEX TERMS Channel state information, Frequency modulated continuous wave, Radio frequency signals, Contactless, Wireless, Vital signs, Survey, Overview, wireless Technologies, Wireless sensing.


I. INTRODUCTION
R ecent statistics have shown that the population aged 65+ is projected to increase from 6.9% to 12% and in particular from 15.5% to 24.3% in Europe [1]. This population range is usually exposed to suffering from chronic diseases like Parkinson's disease, diabetes, and cardiovascular disease (CVD) [2]. In the USA, for example, 80% of the elderly population have at least one chronic illness and 50% have at least two. This list presents the major cause of death around the world, where 30% of global deaths are caused by CVD each year [1]. Furthermore, this age group has the highest mortality rate in the ongoing Coronavirus (COVID- 19) pandemic. Therefore, continuous monitoring is needed to continuously observe their vitals. Nevertheless, this care will certainly increase the cost of healthcare [2], which is not available for all patients, especially in many underdeveloped countries.
Recent research has shown that the solution to this challenge is to integrate wireless communication and sensing technology, through the deployment of sensor devices on or around the human body to supervise their vital signs, forming what is called wireless body area networks (WBANs) [3]. However, implanted sensors on the patient's body make them uncomfortable and require frequent expert intervention. Moreover, physical contact between the patient and the healthcare provider may lead to the spread of dangerous infections such as the Coronavirus [4].
Thus, a robust alternative preference to the contact body area network related solutions has recently emerged under the name wireless sensing, which uses contactless (devicefree or sensor-less) technology for supervising human vital signs. This choice ensures patient comfort, eliminates physical contact between the patient and the healthcare provider, and overcomes other challenges associated with the use of WBANs such as limited resources, fault tolerance [5], security [6], and others.
Several research studies [7] [8] [9] [10] showed that it is possible to supervise and measure a user's vital signs by sending RF signals, processing, and analyzing the reflection from the patient. In this paper, we give an overview of the most recent advances in wireless sensing techniques used in many healthcare applications. More specifically, a classification framework is proposed for categorizing the existing wireless sensing techniques related to contactless technologies. Also, a detailed comparison between these techniques is provided to help facilitate choosing the right technique for a given case or scenario. Moreover, we overview the main challenges faced by these systems as well as enabling technologies.
The remainder of this paper is organized as follows (as can be observed from Fig.1): Section II presents and explains the essential functions related to wireless sensing technologies. Section III overviews the contactless technologies proposed and used in the literature to supervise human vital signs. Section IV discusses the hardware technology available in literature that facilitates contactless sensing research. Section V talks about the challenges faced by each discussed technology. In section VI we discuss lessons learned and research direction, then we conclude the paper in section VII. II

II. CONTACTLESS SENSORS: FUNCTIONALITY
Contactless sensor monitoring is a new technology that is used to supervise vital signs (motion, emotion, sleep, heart rate, respiration rate), without obliging the patient to wear any sensor device. In most technologies the contactless sensor system broadcasts the wireless signal, then analyzes and processes the reflected signals from the objects. In a typical contactless sensor setup shown in Fig. 2, there are four reflectors (human, wall, bed, and couch), the system begins by suppressing reflected signal from permanently-placed objects (wall, bed, and couch), and focus only on the reflected signal concerning the person under study. Other contactless technologies have the same setup as shown in Fig. 2, where the sensing device is installed away from the subject and is designed to recognise the target from all other objects in the environment. In the following subsection, we will discuss how each of the mentioned technologies operates. We put more emphasis on the FMCW technique since according to our research it is the most commonly used technique and has shown great potential in its applications in future human vital signs monitoring. Moreover, many contactless technologies such as radio frequency (RF) based sensing are based on FMCW technique.

A. FMCW RADAR
FMCW is used in radar systems for sensing, localizing, and tracking an object in front of it by measuring the range, velocity, and angle of arrival of the reflected wave. In healthcare, this technique sends a special sinusoid signal to a subject then the system analyzes the reflected signal based on its frequency and amplitude to extract information about the observed person [11] [12]. Here, we present the principle of this technique by answering the following four questions: First: How do you configure the radar to estimate the range of an object? As shown in Fig. 3 (A), the distance d from the object to the FMCW-radar must be known by the radar. The approximation of intermediate frequency ω if is typically used to find the distance d.
where c m is the propagation speed of the electromagnetic (EM) wave in a given medium and .
ω is the quotient of the chirp's slop frequency [13].
Second: What if there are multiple objects? Fig. 3 (B) depicts the scenario where multiple objects are observed by the FMCW-radar. Therefore, the system must be correctly configured to recognize this scenario.
Third: What is the minimum distance between two objects? Fig. 3 (C) shows two objects very close to each other, the system must be correctly configured to distinguish between one object and two objects which are very close to each other.
Fourth: What determines the furthest distance a radar can see? It is critical that the maximum observable distance by the radar system is determined [14] as shown by Fig. 3 (D).
Moreover, we further discuss important features of FMCW-radar in subsequent subsections.

1) Chirp
It is a signal that exists at the heart of the FMCW-radar. It is considered a sinusoid wave whose frequency increases with time [15], see Fig. 4. As observed from the figure, the chirp starts as a sine wave with the frequency of say, FC and gradually increases its frequency until it reaches the frequency of say FC+B, where B is the bandwidth of the chirp, thus the frequency is modulated and this is what is meant by frequency modulated continuous wave (FMCW).

2) Characteristics of a chirp
The frequency vs. time plot is a convenient way to represent a chirp, as shown in Fig. 5, since the frequency of the Chirp increases linearly with time, it takes the form S = At, where S is the instantaneous frequency of the chirp and A is a constant. In other words a straight line with a given slope. From the figure, the following characteristics of a chirp can be extracted: 1) A start frequency (FC), Bandwidth(B), and duration (Tc). 2) The Slope (S) of the chirp defines the rate at which the chirp ramps up. In this example, the chirp is sweeping a bandwidth of 4GHz in 40µs, which corresponds to a slope of 100MHz/µs.

3) FMCW radar operation
After explaining what a chirp is, now we move on to explain how an FMCW-radar works. For this, we consider the simple FMCW-radar shown in Fig. 6 (A) with : • a single Tx antenna, • a single Rx antenna, and • a mixer. A mixer is a three(3) port device with two(2) inputs and one(1) output, as shown by Fig. 6 (B), where for two input sinusoid signals x 1 and x 2 , there is an output signal x out , its frequency equals the difference of the two incoming signals' frequencies, and its amplitude also equals the difference of the two amplitudes. For example, if x 1 and x 2 are as shown below, then x out is: The FMCW-radar operates according to the following steps:    7 shows the TxChirp and RxChirp signals (the transmitted and reflected signals from the object), observe that signal from Rx is a delayed version of signal from Tx, and that IF signal is a straight line with a constant f value. Notable properties of the Fourier transform are as follows: • FT converts a time-domain signal into the frequency domain.
• A sinusoid in the time-domain produces a single peak in the frequency domain (see Fig. 8).

FIGURE 8: Fourier transform
• Within the observation window T below, the orange tone completes 2 cycles, while the blue tone completes 2.5 cycles. The difference of 0.5 cycles is not sufficient to resolve the tones in the frequency spectrum, see Fig. 9.
• Doubling the observation window results in a difference of 1 cycle, the tones are resolved in the frequency spectrum, see Fig. 10. • As a result, we can say that a more extended period of observation deduces better resolution. Since the signals used in an FMCW-radar are sinusoidal, the FT is typically used to estimate parameters such as range, angle, and velocity. Depending on the number of parameters that need to be detected, the FT dimensions are adjusted accordingly [16]. For instance, to detect range and velocity a two-dimensional (2D) FT is required while a fourdimensional (4D) FT is needed to estimate range, velocity, azimuth, and elevation. Hence, the complexity exponentially increases as the dimensions of FT increases.

6) Multiple objects in front of the radar
Applying the same principle, multiple objects in front of the radar yield multiple reflected chirps at the Rx antenna, see Fig. 11. A frequency spectrum of the IF signal will reveal multiple tones, the frequency of each being proportional to the range of each object from the radar, see Fig. 12.

7) Range Resolution in a radar
To understand how to recognize two objects which are very close to each other, we discuss range resolution in a radar. 1) Range Resolution refers to the ability to resolve two closely spaced objects, see Fig. 13 and Fig. 14. 2) The two objects can be resolved by increasing the length of the IF signal, see Fig. 15 and Fig. 16. 3) Note that this also proportionally increases the bandwidth. Thus intuitively, the greater the Bandwidth, the better the resolution. 4) The range resolution (d res ) depends only on the Bandwidth swept by the chirp: d res =c/2B (c:speed of light, B:Bandwidth). The resultant IF signal is passed through a low pass filter to be digitized by an ADC, then sent to a suitable processor such as DSP that begins by doing the FT to estimate the range, the velocity, and angle of arrival of the object, see Fig. 17. The bandwidth of interest of the IF signal depends on the desired maximum distance: f IF −max =S2d max /c, where S is the slope of the chirp.
An ADC sampling rate of Fs limits the maximum range of 1) The synthesizer generates chirp, 2) The generated chirp is transmitted. Rx receives delayed versions of this chirp, 3) The IF signal consists of multiple tones, the frequency (f) of each tone being proportional to the distance (d) of the corresponding object, 4) The IF signal is digitized. The ADC must support an IF bandwidth of (S2d max /c), 5) An FFT is performed on the ADC data. The location of peaks in the frequency spectrum directly corresponds to the range of objects.

B. CSI-BASED
Apart from FMCW radar, CSI-based sensing has attracted tremendous attention in detection and interpretation of human activities in healthcare applications such as assisted living and remote monitoring. For instance, due to its pervasive and unobtrusive nature in indoor environments, WiFibased technologies have evolved to a promising residential contactless monitoring technique. In this contactless technique, advanced signal processing algorithms are used to precisely extract the CSI of the WiFi signal which includes phase shifts, frequency variations and signal levels [17] [18]. Fig.19 depicts the general structure of CSI-based sensing, where the first step is to capture the signal, then processing the captured signal to extract CSI, and the last step is using machine learning techniques to extract the human behaviour. activity recognition performance or to fill the gaps in time and location coverage for continuous and seamless monitoring.

III. WIFI CSI SIGNAL PROCESSING AND BEHAVIOR RECOGNITION METHODS
Both signal processing and machine learning techniques are used in conjunction with WiFI CSI measurements for behaviour recognition and activity monitoring. In this section we outline various approaches which have been employed to extract WiFi CSI, including those which have been used to identify frequency (Doppler) shifts caused from physical body movements in the vicinity of a WiFi AP. We also discuss methods used to classify CSI/Doppler signatures inherent to a given motion, and the potential improvements afforded by making use of temporal patterns within the frequency data. Figure 1 illustrates the main processing and classification steps in scenarios where people move through WiFi fields.
A. Extracting the WiFi CSI 1) Commercial Off-the-Shelf Devices: The majority of commercial WiFi-enabled devices are able to parse WiFi signal data and output information about the state of the channel, the most common being the received signal strength indicator (RSSI). However factors such as the orientation of scatters, multipath and shadowing act as major sources of error in RSS measurements. Techniques such as 'Fingerprinting' Particularly for Doppler shift and frequency component, there are two main approaches. One approach is to apply STFT or DWT on CSI for extracting frequency component. However, this approach has a limitation on discriminating moving target reflections and stationary reflections. Another approach which is based on passive radar principle applies CAF processing on sampled signals from reference and surveillance channels that contains stationary source signal and moving target reflection respectively. The passive radar approach shows good performance on cancelling the impact from stationary reflection. In Section IV we present a high resolution passive WiFi Doppler   In healthcare, WSNs are a set of sensors nodes randomly positioned and co-operate with each other to monitor, process, and send collected patients' data and their environment to the relevant person [19]. The sensors employ routing protocol for low power and loss network (RPL) for routing packets in a WSN [20]. Emergency services and daily monitoring services in a hospital can be very difficult to manage. Hence, systems which can automate patient monitoring have the ability to enhance patient care. Existing systems include CodeBlue [21], which is a wireless sensor network system which can be deployed at the hospital or at home. The device integrates various medical sensors such as EKG Electrocardiograph, pulse oximeter and EMG. Moreover, MEDiSN [22] is another system that aims to automate physiological monitoring in different settings. The system comprises of several physiological sensors monitoring the patient and have a temporary memory where information is stored and scripted before transmission. Also, a wireless patient monitoring system developed by Washington University is closely related to MEDiSN [23]. Fig.20 depicts WSN nodes deployed in a hospital and home setting.

D. CAMERA-BASED
This system uses cameras to capture the images of subjects (for example their faces) and uses intelligent algorithms to decipher information (for example, pain, confusion, fear, etc). In other cases, thermal camera are used, which have the ability to register the temperature of a subject. In medicine, these special cameras are deployed in quarantine facilities to continuously monitor people and help recognise people with infectious diseases by monitoring their temperature [24] [25].
Moreover, in advanced camera-based vital signs monitoring RGB-thermal image sensors are used to monitor not 1 3 the de facto technology for wireless networking and radiobased wireless networks are exclusively used in hospitals to transmit wirelessly information. Radio communications is a well-developed technology, flexible, inexpensive and widely standardized. Two major approaches are used, namely HoF. Since there are multiple radio technologies operating in the same environment, the challenge is to mitigate against interference between wireless systems, exposure to electromagnetic radiation by patients and hospital staff, spectrum congestion and others.  only body temperature, but also heart rate and respiration rate. The RGB camera is able to measure blood volume pulse (BVP) through the variations of light absorption on the human face [24]. Fig.21 shows a thermal image of a subject being observed. In this section, we will discuss the applications of contactless sensors in human's vital signs monitoring. We will study the latest implementation of this emerging technology using FMCW, CSI, WSN, RF, and camera based.

A. FMCW-BASED
All the approaches presented in this subsection use the FMCW technique to process and analyze the signals that are reflected off the objects, introductory details about this technique was presented in the preliminaries section.
1) Authors in [11] developed a human vital sign monitoring system based on 77 Ghz mm-wave FMCW radar. The system, shown in Fig. 22, also utilizes compressive sensing based on orthogonal matching pursuit (CS-OMP) algorithm and rigrsure adaptive soft threshold noise reduction based on discrete wavelet transform (RA-DWT) algorithm to separate and reconstruct breathing and heartbeat signals. According to the authors, results indicate that the proposed algorithms are able to filter out noise and improve accuracy, where the accuray of respiratory rate and heartbeat rate was about 93%.
Sensors 2020, 20,2999 the radar field of view. The distance information is obtained by range FFT, and then the ran map (RTM) is constructed according to the range distribution. After the human target is dete the DC offset is corrected, and the arctangent demodulation phase is unwrapped by using the e DACM algorithm. The heartbeat signal is enhanced by using the differential phase to furthe the accurate phase change information in Step 2.

Target Detection
Before phase extraction, we must first perform target detection to determine the locatio human being. We assume that the tested person is sitting in front of the radar and stay stabl the human body movement usually covers the heartbeat and breathing signals, it is difficult t vital sign of human body in the presence of body movements. The authors in [30] apply such a problem with a Doppler radar. Although the work in [30] cannot be directly applie FMCW radar, it provides some research direction for the vital sign detection under body mov with FMCW radar, and we leave it for future work). These tasks will be done by the range F RTM construction. A single-frame beat signal obtained after the A/D is a two-dimensional which composes of fast sampling and slow sampling. The vertical axis corresponds to the sl sampling points constructed by the chirp frequency N, and the horizontal axis is the numbe time sampling points M samples . In order to suppress side lobe leakage, a Hamming window i and the range FFT vector is obtained by FFT at the fast time sampling points, so as to ob distance distribution from the radar field of view. In [25], the vertical axis uses a single chir time, and N = 1. However, we experimentally found that using multiple chirps achieves a FIGURE 22: Human breathing monitoring using 77 GhHz mm-wave FMCW radar [11].
2) Authors in [26], developed a short range-Doppler radar (1m) that may be used in many applications, like driving, health (through the monitoring of the heart rate), and security screening. They designed it according to the following scheme, see Fig. 23. It has two helical antennas, each with a 40-degree beam-width, the upper antenna is used to illuminate the heart, and the lower antenna is used for motion compensation, this system uses the ISM band at 2.4GHz and 2.5 GHz. The front view and the circuit of the proposed radar are shown by Fig. 24, where the antennas are constructed by winding copper tape on a low-loss acrylic tube. 3) A CMOS Doppler radar sensor was developed in [27], to estimate motion from heart rate and respiration rate, up to range 1m and 1.5m respectively, authors integrated this Doppler radar transceiver in printed circuits on a simple CMOS chip which has dimensions 5cm x 10cm, where patch antennas ASP PT 2988 are used for both transmission and reception, with a 60°by 80°beam-width and 8-dB gain. The authors made two contributions, in the first, they gave the theoretical derivation as a prediction tool, and second, they achieved an experimentation validation by using PC with Matlab, as digital signal processing(DSP) for SNR. 4) Insomnia monitoring at home is essential for both diagnosis and treatment, authors in [28] proposed the EZ-Sleep system for this purpose. Moreover, the monitoring is done without any contact on the human body. Authors begin by giving the general parameters relating to insomnia and sleep, as shown in Fig. 25. In the figure, SL is the sleep latency, which means the time from coming to bed to the beginning of sleep, TST is the total time of sleep, TIB is the total time spent in the bed, and WASO (wake after sleep onset), is the total time of weakness after sleep. The EZ-Sleep standalone is shown in Fig. 26. It consist of four processes that cooperate in gathering the subject's details to realize the application: First, capturing the user's location in the home using Frequency Modulated Carrier Waves (FMCW), and an array antenna to separate other reflected RF signal and focus only in the location, this permit to monitor the location and breathing for future use. Second, inferring bed area, it defines where the user stays stationary (chair, bed, desk, couch) then classify it as bed or no-bed, by using   [28] machine learning and image processing tools. Third, detecting bed entry and exit by using Hidden Markov Model (HMM). Fourth, a neural network model is used to classify sleep versus awake. 5) Authors in [29] proposed a system to monitor and supervise breathing and heart rate signs through wireless signal and without any body contact. They also studied and presented the major challenge that must be addressed for complete efficiency in this system, that is, the inhale and the exhale effect on the reflected signal time as shown in Fig. 27. The proposed system is FIGURE 27: Inhale and exhale motion [29] called Vital-Radio, and it operates according to the following three steps: First, vital-radio transmits a wireless signal, then separates the reflected signals from different objects into buckets according to their TOF's, as shown in Fig. 28 using FMCW technique. After the separation of the reflected signals into buckets, Vital-Radio eliminates the signal of the static objects like furniture and walls. Second, Vital-Radio analyzes each bucket to identify those giving breathing and heart rate. Moreover in the third step, vital-Radio extracts breathing and heart rate by applying signal processing through FT. Vital-Radio differs from previous systems due to its ability to monitor several peoples simultaneously. 6) Authors in [30], began by presenting the importance of the study of behaviour in the home, for either academic goals like social research or to be used by health-care givers. They then proposed a wireless system called FIGURE 28: Separating reflectors into different buckets [29].
Marko, with the ability to collect information about the user's behaviour at home, without asking them to write diaries, answer questions or wear sensors. As an example of the user behaviour at home, consider when a person reaches out to the fridge, indicating the desire to eat. The proposed Marko system can be used to treat multiple users in the same home. For example, for a couple who live together Marko can answer the following questions: Does the couple sleep in the same room/bed? Who wakes up First? Who prepares dinner? Does the couple eat together? The proposed system accomplishes its task through the following steps: Marko system transmits an RF signal and processes its reflections to extract two types of data, RF frames and short user expected direction of motion. It uses FMCW technique as a tool for the first goal, then uses a simple technique to compute user locations. It then connects locations of the consecutive frames to construct short trajectories called tracklets. After this step, Marko system starts to eliminate extraneous data concerning the static objects, then uses a conventional neural network classifier to identify users by using RF frames and tracklets holding to the same time interval as shown in Fig. 29. 7) A passive fall detection system (Aryokee) is proposed in [31]. It is a multi-antennas wireless system composed of three essential components as observed in Fig.  30. In the figure, fall detector, stand-up detector, and state machine take as input a short time window of RF signal transmitted by the FCMW technique. The first two component operate in parallel, each one has two cascading CNN's and a none-maximum suppression (NMS). The state machine takes as input the results of the two previous components to infer the exact fall time and fall duration. 8) Authors in [32] combined convolutional and recurrent neural networks to detect sleep stages using RF reflection. This prediction model consists of three components: feature encoder, sleep-stage predictor, and source discriminator. As shown in Fig. 31, the predictor cooperates with the encoder to predict one of the following states (Awake, Light sleep, Deep Sleep, and REM (Rapid Eye Movement)), and prevent the discriminator from decoding the source label.
In the figure, x is a 30-second RF spectrogram and y is the predicted sleep stage. 9) Authors in [33] developed WiGait, which is a home sensor shown in Fig. 32 used to monitor gait velocity and stride length according to the following steps: First, the device takes a set of locations provided by the wireless localization system, then identifies periods when the person is stationary or moving in the place to remove them. Second the system processes the remainder to focus only on the walking periods. Third WiGait estimates the gait velocity and stride length using stable phase based on time series. FIGURE 32: WiGait sensor system [33] 10) Author in [34] utilized FMCW-based radar system to monitor the respiratory rate and the heartbeat of a patient in a challenging indoor environment in order to simulate ambient assisted living (AAL) conditions. The authors considered four different scenarios where the patient's chest, left, right, and back side were facing the antennas. Moreover, the authors compare the performance of the proposed system with a pho-toplethysmograph and a respiratory belt for all the four cases and conclude that the proposed system measures both respiratory rate and the heartbeat with great precision. The designed system is a SISO FMCW-based radar working with 5.8GHz industrial, scientific, and medical (ISM) band shown in Fig. 33. 11) Authors in [35] developed a wireless healthcare life and vital sign detection system using FMCW radar. Unlike conventional systems, the authors claim that the proposed system is able to detect motionless patients by analysing inherent characteristics of human respiration motion such as spatial correlation and periodicity. Moreover, multiple algorithms are employed to remove the effects of the environment and distance. Results indicate that the proposed system is able to accurately detect motionless patients. Fig. 34 shows the flow chart of life detection. 12) Authors in [36] use Emerald, which is a contact free wireless sensor utilizing machine learning for deep analysis of movement disorders. The authors use the device to measure the progression of facioscapulohumeral muscular dystrophy (FSHD). According to the authors FSHD affects around 16,000-38,000 USA citizen, moreover, its effects slowly progress over the years making it extremely difficult to track using hospital facilities. Twelve FSHD patients were continuously observed for three months using two of this device for each patient (One in the living room and one in the bedroom). The devices collected about 40,000 movement data which were processed to extract gait and time to exit the bed. The condition of the patient was raked from mild to severe and allocated a clinical severity score (CSS) from one to four. According to the authors, results indicate that there is a relationship between time to exit bed and CSS. 13) Authors in [37] also used Emerald to monitor postoperative recovery process for patients recovering from endometriosis surgery. The study was conducted on three women where the device was placed on each of their homes to study the patients movement and sleep for five weeks before surgery and five weeks after surgery. The patient was also required to conduct a daily pain assessment survey where they rate the level of pain they are experiencing. According to authors, results indicate that the longer it took the patient to fall asleep the higher the pain rating was the next day. 14) Authors in [38] used infrared images to monitor the head position of patients suffering from obstructive sleep apnea (OSA). The authors use three machine learning algorithm to process the images. Data from fifty patients with different levels of OSA severity was collected. According to the authors, the best performing algorithm Darknet19, correctly identified lateral and supine head position with an accuracy of 92% and 94% respectively. 15) Moreover, authors in [39] deployed the Emerald device which uses radio waves to detect the agitation and pacing pattern of patients suffering from Alzheimer's disease (AD). The authors claim that the device is able to detect changes in movement patterns that are able to reveal underlying behavioral symptoms and environmental triggers. The device was installed for seventy days in the house of a single patient suffering from AD. According to the authors results indicate that there is a relationship between pacing of the patient and environmental triggers. 16) Authors in [40] exploited the RF signal to view the human skeleton without wearing any smart device. The essential elements in this process are RF signal sent by a radio device, the antenna array is used to define the angle direction of the arriving signal. FMCW is used to measure RF reflections and static reflection elimination and focus only on the reflection of the human body. These elements are integrated into the developed RFcapture device. The big challenge facing the proposed system is: there is a lot of RF signal reflected away from the device, and only a subset reflected towards it as seen in Fig. 35. Authors tried to address this challenge by exploited body motion to trace the body of the person. 17) Authors in [41] proposed BodyCompass (see Fig. 36), an RF-based sleep poster monitoring device to help patients avoid bedsores after surgery, reduce apnea event, monitor the progression of Parkinson's disease, and even alert epilepsy patients to a potential lifethreatening sleeping position. The device works by analyzing RF signals that bounce off the patient using machine learning algorithms. According to the authors, the device was used to evaluate data from twentysix patients over a period of two hundred nights. The observed accuracy of the system was over 80%. 18) Authors in [42] proposed the use of an RF-based home monitoring device to servile patients with Parkinson's decease (PD) focusing on gait, home activity, and time in bed. The study was conducted on seven patients over a period of eight weeks where the RF-based device was installed in their rooms. According to the authors, promising results were obtained to help better understand PD, moreover, the device can be used to treat other chronic diseases. 19) Authors in [43] claimed that RF signals may be used for inferring emotions (happy, sad, confused, etc.), then they demonstrated that the reflected RF's have information that can be used to extract breathing and average heart rate without any contact, they also presented the challenges facing this technique. First, the reflected RF signal is modulated by breathing and heartbeats. Second, the heartbeat in the RF signal lacks the sharp pick that characterizes the ECG signal. As a result, the boundaries of identification become hard, to address these challenges, the authors designed and developed the EQ-Radio system, which consists of three compo-  The transmitting chain of the radar is composed of a voltage controlled oscillator (VCO), with an output power of 5 dBm, a power amplifier (PA), a power divider, and the transmitting antenna. An attenuator was added between the VCO and the amplifier to reduce the input power to this component and to make it work in the best conditions. For the VCO, a triangularly-shaped feeding signal was preferred to a ramp since it did not present strong discontinuities, which could adversely affect the measured signals with the introduction of unwanted harmonic components. The feeding signal V feed has a fundamental frequency f T = 300 Hz and a duty cycle of 10%. The receiving chain includs, instead, the receiving antenna, a low noise amplifier (LNA), a mixer, and an ADC. To reduce the costs and the dimensions of the system, discrete components were preferred. In particular, the chosen components were HMC587LC4B [38] for the VCO, HMC392ALC4 [39] for both the PA and the LNA, ZN2PD2-63-S+ for the power divider [40], and HMC557A [41] for the mixer. The transmission and the reception of the signal were controlled and synchronized by the NI DAQ USB 6361 [42]. Considering that the triangular signal produced by the NI DAQ USB 6361 stayed constant in time, we could assume that also the chirps generated by the VCO were essentially identical.

Antenna Design
The antennas were specifically designed for the proposed application [43]. The receiving and transmitting antennas are two identical series-fed arrays composed of six patches. The shape of the patch (see Figure 3c) was suitably modified to increase the bandwidth of this antenna topology, which is typically narrow. Since for FMCW radars, the range resolution is inversely proportional to the bandwidth [44], it was crucial to design an antenna with a band equal to or larger than the one available in the ISM frequency range used. The bandwidth increase was achieved through a dual band structure, given by the superposition of two tapered patches. These patches were obtained by spline interpolation within the red and the green dashed polygons in Figure 3c. The two resonating frequencies depend on the lengths L 1 and L 2 in Figure 3c. To reduce the side lobe level (SLL), an alternative technique to the amplitude tapering was applied. The curvature degree of the upper edge of the patch, controlled by the indent parameter i in Figure 3c, was used to modulate the amount of power transferred from one patch to the following ones. In particular, enlarging the indent increases the power radiated by the patch and reduces the transmitted one. Feeding the antenna from one of the two central patches (see Figure 3) and choosing the appropriate value of i, a superficial current distribution ensuring a higher radiation from the central patches than from the most external ones could be guaranteed, thus allowing a fine regulation of the SLL. Furthermore, it is worth noting that, contrary to what happens for the amplitude modulation methods, the presented SLL control does not require a variation of the patch shape along the array. With the proposed design, the antenna has a fractional bandwidth of 5.92%, which is more than the double the 2.6% required and wider than the n be (4) . The n the GHz e by good body hase appear among the corresponding neighboring range bins [10].
Compared with other stationary objects, the human target has periodical breathing, exhalation and inhalation, causing the body undulation. For stationary objects, the amplitude A and phase φ of the IF signal remain unchanged to some extent : For living human targets, they change over time due to body undulation, resulting in fluctuations of the amplitude and phase. Moreover, the fluctuation is much larger than that of stationary objects and noise. Therefore life detection is converted into body undulation detection. And two parameters are utilized: std fftVal and std Phase, standard variances of amplitude in IF frequency spectrum and phase. As presented in last section, the phase is very sensitive to small distance change and it is used as a parameter to describe the body undulation in this section. In addition, due to that the SNR (Signal Noise Ratio) and signal strength of radar signal drop along the distance, the correlation between range bins will be disturbed and changed. Based on the aforementioned    phasizes pixels with a signicant contribution from the subject's breathing while still retaining the FIGURE 36: Signal reflections in a patient's room [41] nents as shown in Fig. 37. First, the radio transmits FIGURE 37: EQ-Radio Architecture [43].
a low power signal, then estimates its reflection time using the FMCW technique previously described, this is used to differentiate between reflected signal off different objects. Each one in a specific bucket, EQ-Radio achieves a pre-processing to mitigate noise and to improve the signal-to-interference-and-noise ratio (SINR). Second, the heartbeat extraction algorithm takes the reflected signal after the pre-processing phase previously achieved as input and returns a series of segments as output. This output is used by the third component which is the emotion classification, to determine an emotion according to the received physiological signals using an SVM classifier. Through experimentation, authors showed that EQ-Radio performs better compared to ECG-based and image-based systems. 20) Authors in [44] proposed Witrack system, for tracking users in two scenarios, these are line of sight and nonline of sight (through a wall). The system is composed of four antennas in T form. One for transmission and three for reception as can it be observed from Fig.  38. The device transmits an RF signal then computes and estimates the TOF using FMCW technique, it then focuses on the human's body reflections by suppressing those of the static objects after de-noising. Witrack FIGURE 38: Antenna 'T' and FMCW signal generator [44] estimates the traveled distance from the transmission antenna to the human and back to the reception antenna, in other words, three estimated distances are used to identify the user's location. In the end, Witrack uses elevation through the Z-axis to detect fall with the following two conditions. First, the elevation change must be significantly observed. Second, the final value of this elevation must be near to the ground level. 21) Authors in [45] proposed a wireless system called WiTrack2.0, as an improvement to WiTrack [44], it has 5 antennas for transmission and 5 antennas for reception as shown in Fig. 39. Unlike Witrack, it can localize multiple users in a multi-path indoor environment by transmitting an RF signal and then processing its reflection. It has two essential components, the first component is the FMCW to deal with the multi-path effect, and the second component is the successive silhouette cancellation algorithm to overcome the nearfar problem. 1) Authors in [17] proposed a WiFi CSI-based human activity detection for assisted-living patients. The system works by analysing the Doppler shifts in the WiFi CSI caused by human activity. Three case studies are performed in the experiment, through-wall detection of vital signs, daily activity recognition in residential home, and activity monitoring in a residential environment, sequentially inference as case 1, case 2, and case 3 respectively. Fig. 40 shows an example of 24-hour monitoring in a house environment using the proposed system.

IEEE Communications Magazine • May 2018
136 applied to the Doppler data, and the predictions were made "batch by batch." Examples of such activities are shown at the bottom of Fig. 5. It is accepted [15] that a sequence of activities is correlated in the sense that certain activities are more likely to follow others, and likewise, some are impossible to occur one after another (e.g., a person cannot transition from sitting to running without standing first). Such a sequence of activities can be represented by a hidden Markov model (HMM) or related models such as Gaussian random processes and conditional random field (CRF). These sequential methods incorporate future and past observations (here, Doppler information) to improve predictions for the current estimate. It is widely reported in other areas where HMMs are used (e.g., speech recognition) that such strategies can lead to improved performance compared to techniques that rely only on current observations. In the context of residential healthcare, sequential prediction of activities can plausibly help in prediction of future activities or warn of increased risk of a given event.

CONCLUSIONS
In this article we explain how WiFi CSI can address the challenges of behavior recognition and activity monitoring in residential healthcare. State-of-the-art signal processing techniques make it possible to extract accurate Doppler data, allowing us to characterize activities and behaviors of monitoring human subjects. We have presented three case studies to illustrate the capabilities of a passive WiFi Doppler sensing system within a healthcare setting, including, monitoring of vital signs, fall detection, and living pattern monitoring. The results confirm that WiFi CSI-based sensing technologies show good potential for healthcare applications. We have also identified four key challenges that must be overcome to facilitate a transition of the techniques into real-world assisted living applications.
•WiFi Signal Processing: Many existing techniques such as SpotFi [14] and Matrack [3] rely on high-data-rate OFDM WiFi transmissions to extract CSI data. However, only a few methods such as [9] can be applied to lower-bit-rate or WiFi beacon signals, albeit with performance deterioration. High-resolution algorithms for processing low-data-rate signals are therefore critical.
•Time Alignment: The classification performance has been shown to depend on accurate time alignment during the gesture cycle. Thus, determining the time points of of start and end of a gesture or behavior is crucial to performance.
•Multiple Users: Previous studies such as [4,[8][9][10] have only considered sensing individuals, and extrapolate to multiple users by proposing additional devices. However, additional users significantly add to the sensing complexity in terms of shielding and multipath, sensor deployment, cost, and so on, and is a challenging next step.
•Sensor Fusion has shown promising results in WiFi CSI-based behavior recognition. It is generally accepted that a variety of sensors will be deployed in future smart homes. Fusing WiFi CSIbased recognition data with other sensors like cameras, accelerometers, or electricity, humidity, and water meters will provide more accurate and seamless recognition and modeling of human behavior.   2) Authors in [46] proposed a non-invasive method for monitoring the respiration rate using received signal strength measurements (RSS). The goal of the proposed work is to address two critical problems encountered when observing breathing using a single transmitter-receiver pair. The first problem is the difficulty in observing the breathing signal using RSS because inhaling and exhaling causes finite change in propagation channel. The authors addressed this problem by increasing the signal-to-noise ratio (SNR). The second problem addressed is the distinction of other movements of the person from breathing motion, these movements dominate most of the observed RSS response. To counter this problem, authors suggest to identify time instances of the movements as monitoring continues, then the device is disabled at these instances in the long run. The following scheme presents the different components of the breathing monitoring system as observed in Fig. 41. First, the noised RSS measurement is pre-filtered to increase the SNR then downsampling is performed. Second, the mean removal is applied to the sampled signal, then passed through a motion interference detector and finally through a breathing estimator. 3) Motion detection become an essential component in most modern systems, therefore, authors in [47] proposed WiDetect system. Unlike most existing motion detectors which require particular installation, calibration, and have limited coverage, WiDetect is a highly accurate, calibration-free, and low-complexity wireless motion detector. The proposed system exploits the statistical properties of electromagnetic (EM) waves to find a relationship between the autocorrelation function FIGURE 41: Breathing monitoring system components [46].
of the physical layer CSI and movement in the surrounding environment. Authors consider buildings and rooms with scatters that diffuse EM as reverberation cavities. This is because the building or room contain internal multipath signals between a transmitter (Tx) and a receiver (Rx) as shown in Fig. 42.  [48] used reflected Wi-Fi signals to recognize humans through their gait patterns in an indoor environment utilizing the architecture depicted in Fig.  43. They used two transmitter antennas (such as a Wi-Fi router) and a receiver antenna (such as a laptop). The proposed system operates according to the following steps:

4) Authors in
The system starts by collecting the CSI measurements, then extracts the principal component from CSI mea-FIGURE 43: Application scenario and data collection environment [48].
surements using the Principal Component Analysis (PCA) technique. Next, it transforms the resulting PCA into a spectrum using short time fourier transform (STFT), then enhances this spectrum by applying frequency domain denoising algorithms to reduce the noise. After that, a machine learning technique (SVM classifier) is applied to classify the extracted feature into two classes (the person understudy and others). 5) Authors in [49] designed WIFID system, which is a device-free system used to identify humans indoors.
The system exploits the ability of PHY layer CSI to obtain the frequency assortment of a widband channel. The layout of the experiment is shown in Fig. 44. A pair of Wi-Fi transmitter and receiver is deployed. The transmitter has a single antenna while the receiver has three antennas. Moreover, the transmitter continuously broadcast a signal every 3 milliseconds. After receiving this signals the system runs the following three modules: First, WIFID analyzes CSI to focus on the user segment and to extract a novel feature of sub-carrier amplitude frequency (SAF). Second, WIFID applies PCA on the resulting SAF, and third, WIFID uses an SVM classifier for classification. 6) A non-contact system to detect paraparesis is proposed by authors in [50]. The proposed system has SIMO architecture and operates on a pipeline mode according to the following steps as observed in Fig. 45. First, the system collects data through wireless sensing. Second, the collected data is passed to the pre-processing stage where the system uses a Humpl filter to detect outliers values, then applies Wavelets for de-noising signal and finally uses data calibration to increase sample data. Third, a CNN classifier is applied to classify data and detect paraparesis. 7) Authors in [51] used CSI from WiFi 802.11n in a home environment as shown in Fig. 46, to detect human activity such as sitting, standing, walking, and running. Moreover, the human activity recognition system is able to detect the number of people in a room as well as recognising human fall, which is very instrumental in the case of assisted living conditions. The proposed system is developed using two algorithms: support vector machine (SVM) used for classification and long short-term memory (LSTM) recurrent neural network. where a person does not need help after a fall and the fall will be considered as a false alarm and labeled as such, i.e. as FALSE ALARM. Captured packets are logged into the file that is later used to extract CSI values. The interval between logged in packets between samples is 5 s. To separate one sample from the next one the timestamp of packets is used.  8) Author in [52] proposed a reconfigurable intelligent surface (RSI) based radio frequency system for human poster recognition. As shown in Fig. 47, the system comprises of a single transmitter, a receiver and an RIS. The RSI can reflect and modify the incident signal from the transmitter. During sensing, the transmitter continuously generate a single-tone signal of a given frequency which is reflected by the RIS and the human body before reception by the receiver. According to the authors, the proposed system is able to customize the environment to provide the needed propagation properties and a plethora of transmission channels. 9) Authors in [53] proposed a monitoring system to supervise events in an indoor environment. The proposed work exploits the information embedded in the channel(CSI) and uses commercial WIFI devices for real time monitoring. The system is made up of a feature extracting algorithm utilized during training to recognize the most distinctive sequence of CSI. The system also uses principle component analysis (PCA) to reduce noise and to remove correlation between subcarriers and links. Authors also note some challenges associated with real-time monitoring. The challenges include unknown start and end time of an event, change in occurrence instances, and precise detection with low latency. Authors propose a modified k-nearestneighbor (kNN) classifier to address the mentioned constraints. Another problem facing real-time monitoring is durable robustness due to the eventual and unpredictable environmental change. Authors integrated in their model an unsupervised retraining algorithm that guarantees high accuracy during an environmental change. The proposed system is depicted in Fig. 48. 10) Authors in [54] proposed MultiSense human respiration sensing device. According to the authors, the WiFi-based device accurately measures similar breathing pattern of multiple people in a room even when they are very close to each other. The device utilizes multiple antennas to receive a blended reflected signal from monitored individuals. Authors claim that the device is able to accurately measure respiration even in a room of four people with the mean absolute respiration rate error of 0.73 bpm (breaths per minute). 11) Authors in [55] developed a continues real-time heartbeat monitoring device utilizing 24 GHz continuouswave Doppler radar as well as artificial neural networks (ANNs). The device was designed to have very low latency of less than one second. Moreover, the device was tested on twelve volunteers and the results were evaluated against results from an electrocardiogram (ECG). Authors claim that results indicate that the device is viable to detect accurate heartbeat from patients. Fig.49 shows the proposed model. 12) Authors in [56] proposed WiSee, a whole-home gesture recognition system that requires neither sensors on humans nor a camera system. The system can recognize gestures behind obstacles because it uses Wifisignal that propagate not only by line-of-sight but also through the wall. WiSee achieves the recognition task by following the steps below: First, the system extracts Doppler shifts that appear This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final on Selected Areas in Communications

A. RIS Model
The RIS is an artificial thin film of electromagnetic and reconfigurable materials, which is composed of a large number of uniformly distributed and electrically controllable RIS elements. We denote the number of RIS elements by N and the set of them by N . As shown in Fig. 1, the RIS elements are arranged in a two-dimensional array.
Each RIS element is made of multiple metal patches connected by electrically controllable components, e.g., PIN diodes [14], which are assembled on a dielectric surface. Each PIN diode can be switched to either an ON or OFF state based on the applied bias voltages. The state of an RIS element is determined by the states of the PIN diodes on the RIS element. Each state of the RIS element shows its own electrical property, leading to a unique reflection coefficient for the incident RF signals. Suppose that an RIS element and its state s. Th element is assume An example of is depicted in Fig  complex number, and \r(✓ I , ✓ R , s) shift between the r which can be obt RIS element. As a specific reflectio and proposed sys designed for a dif However, since large, it is costly independently. To divide the RIS ele same number of R same group are lo Fig. 1. Specificall L groups, and th denoted N l , whic l 6 = l 0 ) and S L l=1 each group by N G We control the group, that is, th the same state, a controlled indepen by s l . We refer s = (s 1 , ..., s L ), changing the con waveforms of the By using the beam generate various d recognition.
Remark: In th between different the reflection coe on its own state group to be in elements at differe situation more co experimental resul It is shown that t FIGURE 47: Reconfigurable intelligent surface (RSI) based radio frequency system model for human posture recognition. [52] FIGURE 48: Indoor event monitoring system diagram. [53] Sensors 2020, 20, 2351 8 of 16 where j refers to the current layer, N is the number of inputs to the current layer, w is the weights and b is the corresponding biases. The output of each neuron was passed through the hyperbolic tangent function, with the exception of the output neuron which used a sigmoid function. The weights were calculated through a numerical optimization of the mean squared error loss function in Equation (6): where M is the number of data points, b i is the binarized target signal and a i is the network output. This was an iterative procedure that was set to run for a maximum of 1000 epochs or to stop early if the solution became su ciently close to the minimum, that is, if the gradient became smaller than 10 7 .
The training would also stop if the error on the portion of the data set aside for validation (30%) failed to decrease for 6 consecutive epochs.
To remove fast noisy changes, the sequence of outputs of the ANN calculated for each sample was smoothed. This stage of the detection algorithm was implemented as a moving average filter with a width of 10 consecutive ANN outputs.
The next stage of the algorithm was the peak detection subroutine which marked local maximums of the continuous probabilities output, imposing established constraints on the minimal distance between the consecutive peaks based on the known physiological range in rest 40-120 beats/min [38,39] and their prominence (detection amplitude). When the duration of the detected inter-pulse interval (IPI) was twice as large as the previously detected IPIs (within the established constraints), a beat was interpolated as having occurred at the point in time that was the arithmetic mean between the occurrences of the current and previously detected heartbeats. The detection amplitude was defined empirically for each ANN topology on a small test sample using the error of the number of detected heartbeats as a metric. The same detection amplitude was then used for all the subjects.    [55] in the wireless signal, caused by the movement of the object. Second, WiSee maps this Doppler shifts to the gestures. WiSee can classify the nine wholebody gestures shown in Fig. 50. Moreover, WiSee system also takes into consideration multiple humans scenarios using MIMO.

C. WSN-BASED
A wireless sensor network (WSN) is a set of interconnected sensors deployed in a given area to supervise a particular phenomena. The sensors mostly use multi-hop communication technique to send their data. In this section, we present three main approaches of the contactless vital-signs supervising based on a set of WSN devices. [57] showed that 4% of the male and 2% of the female adult population is affected by sleep apnoea. Therefore, the authors developed the SleepMinder, which is a novel contactless technology to supervise sleep and breathing at home. The sensor operates in a license-free band of 5.8 GHz, and its emitting average power is less than 1mV and it is able to measure movement and breathing across the distance of between 0.3 and 1.5m. Also, it can process the case of two peoples in the bed, by combing sophisticated sensor design and signal processing to focus only on the respiration of the nearest person. The study aims to validate Sleep Minder as a sensor contactless technology with the ability to accurately estimate the Apnoea-Hypopnoea index (AHI). The model of the proposed Sleep Minder system is shown in Fig. 51. In the proposed model, algorithms used in [58] is integrated and used to analyse recordings of 129 subjects with suspected sleep-disordered breathing (SDB).

1) Authors in
Numerous observations were made such as sleep state and motion flags which were used in further analysis. Moreover, a new algorithm was used to detect SDB using phase demodulation and amplitude as well as correlation-based signal processing methods. The AHI is estimated by dividing the total number of the detected events over the Total Sleep Time (TST) as seen in Fig. 52 FIGURE 52: Signal processing steps [57].
2) Authors in [59] developed a WSN-based network to monitor patients' psychological parameters and track patients inside a hospital. According to the authors, the experiment was done in collaboration with Henri Mondor University Hospital Center in France. Fig. 53 depicts the setup of the experiment where the application used in the proposed system simply collects temperature and luminosity values captured by sensor.
hose the MDA100CB sensor and data acquisition board h has a precision thermistor and light sensor/photocell. ptures thereby light and temperature. A base station s the aggregation of sensor network data onto a PC or computer platform. Any MICAz mote can function as a station when it is connected to a standard PC interface ateway board [16]. Our platform includes the MIB520 ay board which provides a serial/USB interface for both ramming and data communications. e Imote2 is an advanced wireless sensor node platform. ntains the Marvell PXA271 CPU. The Imote2 uses the 420 IEEE 802.15.4 radio transceiver from Texas Instrus. The CC2420 supports a 250kb/s data rate with 16 nels in the 2.4GHz band. The Imote2 platform integrates GHz surface mount antenna which provides a nominal e of about 30 meters [17]. e Cricket platform, is an indoor location system which ides two forms of location information : space identifiers position coordinates. The most common way to use et is to deploy actively transmitting beacons on walls r ceilings, and attach listeners to host devices (handhelds, ps, etc.) whose location needs to be obtained. Because is a difference between RF and ultrasonic travels, listener se the time difference of arrival between the start of the essage from a beacon and the corresponding ultrasonic to infer its distance from the latter. Cricket's radio runs at quency of 433 Mhz, with the default transmit power level antennas providing a range of about 30 meters indoors there are no obstacles. The maximum ultrasound range .5 meters when the listener and the beacon are facing other and there are no obstacles between them [18].
ests and analysis e purpose of this section is to describe the different steps ploying the Crossbow WSN platform and to present test ts and their analysis. In a first time, we present results of ementing MICAz platform (in the first three subsections). in second time, we present those of Imote2 and Cricket e last subsection). e set up a Crossbow platform formed by 17 MICAz s (16 motes + base station). These motes were deployed in the laboratory buildings. We took into account the tial obstacles (wall, glass, etc.) and set up a network various radio link types. There are direct links and other h get over obstacles. The application provided has simply purpose to collect temperature and luminosity values red by sensors. Deployed network's map is illustrated in nodes which are close to the base station (15, 16) have a higher percentage of transmitted packets that the ones that are far away. However, mote «4» which is also close to the base has a lower percentage. This proves that there is a second factor that impacts this percentage. Figure 4 illustrates the number of parent changes for each node. Nodes «5», «7» and «8» have zero change whereas nodes «4» and «3» have several changes. This can be explained by the effect of pedestrian traffic on communication links, and the asymmetric nature of link quality between nodes. Nodes which are deployed in offices with low pedestrian movements have a lower number of parent changes than those in eventful offices. Thus, we can conclude that distance between mote and the base station, signal range and pedestrian mobility are three factors that impact the WSN performances. Deployment 3) Authors in [60] deployed a set of sensor devices as shown by the testbed network in Fig. 54, these wireless devices (black points) collaborate with each other to localize a breathing person and estimate his breathing rate in-home using RSS. The authors demonstrated two situations. First, the case when the person is stationary (sitting, lying down, standing, or sleeping.), in this case, a prior calibration is needed. The Second situation is when a person is in motion (moving), in this case, we do not need calibration. They presented two methods to estimate the breathing rate. The primary method takes the maximum frequency at the sum, it is based on the power spectral density (PSD) calculated over links. Moreover, authors mentioned the challenge facing the RSS-based breathing rate estimation. Interference due to links sharing with other objects in the environment is one of them. To address this problem, authors use time index to identify at which time the sudden RSS change occurs. This technique is called a break point. As the second method for the localization, authors exploit the previous two techniques and the amplitude of the signal component at each link to construct the map and identify the chest position. For experimentation, they deployed TI CC2531 dongle nodes.

D. CAMERA-BASED
Observing the human face and body without any contact and measuring physiological signs using a camera is a technology FIGURE 54: Sensor devices test-bed network [60].
that has gained research momentum in the last decade. Without the need of contact, remote cameras can be used to measure vital signs such as heart rate irregularity, breathing rate, blood oxygenation saturation, pulse transit time and body fever. According to [61] the most popular type of external sensors are cameras. In this section, some applications of camera-based sensing are highlighted.
1) A contactless vital signs measurement paradigm using RGB-Thermal image sensors is proposed in [24]. The RGB camera sensors measures blood volume pulse (BVP) using light absorption deviations on the human face. Moreover, infrared thermography (IRT) was used to measure changes in temperature near the nose or mouth during breathing. The authors conclude that the results obtained by the proposed contactless system for measurements such as heart rate, respiration and temperature are highly similar to the results obtained using contact systems such as a respiration belt and a thermometer. Fig. 55 shows the proposed contactless system. 2) Authors in [63] developed a screening system utilising thermography and other wireless sensors to observe multiple vital signs from a subject. Moreover, the authors employ machine learning and apply six different classification algorithms and compare their performance. According to authors, after performing a test where 92 people were screened using the proposed method, results indicate that there is 50% more accuracy increment compared to using thermography alone. Fig.56 shows the proposed system model. Other previous works using infrared cameras include mass blind fever inspection [64][65] [66], infrared thermometers [67], and RGB-thermal screening with facial tracking [68]. 3) Authors in [69] proposed a see-through-wall imaging radar system that is used by law enforcement agencies to observe objects behind a wall. The system is an Ultra-wideband high-resolution short pulse imaging radar paradigm operating at 10GHz. According to the authors, the model investigates two design properties, first is the the electromagnetic wave propagation property through wall, and second is the pulse fidelity. 4) Authors in [70,71] discussed the principal behind non-contact measuring of respiration and heartbeat using infrared/RGB facial-image. Moreover, potential applications such as the detection and observation of a person with infectious diseases were also discussed. According to the authors, an RGB camera was used to observe the heartbeat of the patient from the face due to the degree of skin exposure and ease of detection. The camera measures the reflected light from the arteries to get the heartbeat signal. Moreover, an infrared camera was used to measure respiration activity by observing the nasal-region of the subject. An infrared camera is used because during inhalation the cool air from the environment lowers the temperature around the nasal cavity and during exhalation the hot air from the lungs raises the temperature around the nasal cavity. Fig. 57 shows the principal utilized to measure respiratory and heart rate using infrared and RGB imaging as well as an infectious decease screening system. 5) The current omnipresent property of smart phones with optical sensors have provided an opportunity for lowcost vital signs remote monitoring. Authors in [72] proposed a novel system for monitoring heart rate, perfusion index and oxygen saturation of a person using a camera system from a smart phone. According to the authors, they employed principal component analysis (PCA) technique and the results obtained demonstrated that the proposed system performs better than conventional systems. Fig. 58 shows the proposed system model, where, HR, PI, and SpO2 represents heart rate, perfusion Index, and saturation of peripheral oxygen respectively .

IV. NEW HARDWARE TECH FOR WIRELESS SENSING
In this section, we will discuss new advancements in hardware platforms that facilitate research on wireless sensing. Numerous hardware platforms that facilitate research on wireless sensing have emerged due to its resent popularity among researchers and the high potential of contactless sens-can automatically detect infected individuals based on their vital signs, which are measured in real time. This system was tested on patients with an influenza-like illness in a clinical setting to evaluate the performance of this vital-sign-based screening approach using IRT alone.
This was a cross-sectional investigation that was undertaken at the Takasaka Clinic in Fukushima in Japan. The study involved 16 outpatients (11 male and five female) who visited the Takasaka Clinic with an influenza-like illness that included fever, headache, and sore throat, between January 22, 2015, and February 25, 2015.    1. Schematic diagram of the multi-modal infection screening system. The system acquires heart rate, respiration rate and facial temperature readings and feeds them into a classifier, which classifies the measurement as potentially infected or healthy.
from the measurement enters only through a kernel function k(x, x ). The kernel function is a function that depends on two input variables and returns a scalar output. More precisely, a new observation x is classified by evaluating the function: and making a decision based on the sign of y. The variables an and b are parameters of the SVM, which are learned from the training data xn by solving a quadratic programming (QP) problem. The standard algorithm for solving the QP problem arising in SVM is called sequential minimal optimization [12].
In the typical case, most of the an found after solving the QP problem are equal to zero. This means that the corresponding training sample xn does not contribute to the classification process, giving rise to the sparseness of SVM. A detailed description of SVM can be found in standard textbooks like [11].
The SVM has two hyper-parameters, which have to be fixed before optimization: the kernel width γ, which determines properties of the kernel function k and the so called box constraint C, which controls the trade-off between minimizing training errors and limiting model complexity. We choose the values of the C and γ by minimizing the leave-one-out (LOO) error via grid search. This approach, however, is susceptible to overfitting with respect to C and γ and we cannot cite the minimum LOO error as the test performance of SVM.
Therefore, we evaluate the performance of SVM using a nested cross validation scheme: Using N − 1 samples we calculate the LOO error for each (C, γ)-pair on a 2-D grid and chose the pair with the lowest LOO error. Then, we train the SVM with the chosen parameters on all N − 1 samples and test on the held out sample. This procedure is repeated N times, each time with a different sample being held out. c) k-Nearest Neighbours: kNN is considered to be a nonparametric method and despite its simplicity, it often achieves good performance in practice [11], [13]. One advantage is that kNN does not require a training phase. Instead, a new measurement x is assigned the label that holds the majority among the k training data samples which are closest to x. However, this means that for classification, the entire training set has to be stored and searched, which can be slow for large and high-dimensional training sets.
Distance between samples is often measured via Euclidean distance, but more sophisticated measures like Mahalanobis distances can also be used. In addition, preprocessing steps like e.g. neighbourhood components analysis (NCA) [14], which learns a custom Mahalanobis distance from the training data, can be applied to improve classification results. However, methods like NCA require a training phase.
Similar to SVM, there is a hyper-parameter, the number of neighbours k, which has to be optimized with cross validation. We use the same kind of nested cross validation procedure that was employed in the SVM case to optimize k and avoid overfitting. d) Logistic Regression: LR is a discriminative approach, where the logistic sigmoid function σ(a) = (1 + exp(−a)) −1 is applied to a linear function of the measurement. The output is interpreted as posterior class probability: The vector w contains parameters, which can be learned from the training data using standard optimization methods [11]. FIGURE 56: Contactless vital sign measurement system [63].
ing technology to be applied in many different areas [73] [74]. According to [75], Wi-Fi sensing technology progress has been impeded by three main obstacles. The first is the unknown baseband design and its influence on CSI. The second is the lack of access to low-level control hardware and the third is the lack of a flexible and versatile software for hardware control. The authors address each of this problem in their work and propose PicoScenes, which is a Wi-Fi sensing technology that authors claim can greatly facilitate the research on Wi-Fi sensing by giving direct access to features of QCA9300 and IWL5300.
Authors in [76] concur that health monitoring using wireless sensors is a popular topic but with open problems, hence, numerous solutions entailing wearable, wireless, open-source, and noninvasive techniques have been proposed. Moreover, authors claim that most of the available platforms for developing the wireless sensing technology are limited and lack flexibility. Therefore, authors propose a new open hardware architecture used to design sensor nodes utilized in healthcare. In addition, authors develop a simulation tool which facilitates the connections with the hardware and simplifies complex systems.
Authors in [77] and [78] acknowledge the high demand for wireless sensor networks in areas such as medicine, military, and structural and environmental monitoring. However, according to the authors, most of the existing platforms for developing these sensor networks are not power efficient. Therefore, authors review some of the most recent and ultra low power processors available in literature. Examples of the hardware systems discussed include general purpose commodity-based systems, smart dust-early event driven, sub-threshold systems, asynchronous-SNAP, charm-network stack acceleration, and Harvard eventdriven architecture.

V. CHALLENGES
In this section, we analyse the challenges faced by the above mentioned contactless sensing technologies.

A. CHALLENGES FACING FMCW-BASED SENSING
FMCW-based sensing technique discussed in details in the previous sections, is a method where a continues frequency modulated wave and its reflections' characteristics are stud-FIGURE 57: Infectious decease screening system using infrared and RGB imaging [70].
be encoded in smart tag for a person for quick assessment ry condition.

YSTEM ARCHITECTURE
stem would collect video image of the fingertip using the of the smart phone. The video image will be analyzed the software inside the camera for identifying heart rate, ion index and oxygen saturation. The perfusion index and gen saturation is calculated using personalized model. The parameters would be different for different persons. Figure s the system architecture. . Each pixel in the frame has blue. Most of the work [4] [1 green band. We first show the novel approach by applying pri identifies the component that variance. We found the princip or the component correspondin close to the green band. This im of green band for analysis of mathematical model for findin saturation of peripheral oxyg developed a method for findin peak distance for maximum acc sampling frequency. For per peripheral oxygen we showed a best to find those parameters ac

Data Collection
We used the camera of a Galax Jelly Bean operating system. image of the index finger for mobile device on. The video im in 3gp format at a frame rate of Matlab 2012 for all the analysis functions only work for avi for using a free converter named Pa

Method
The RGB color space of the vid different time series signals wit intensity (between 0 and 255); one frame and the horizontal frames. The transformed signal f (30 seconds of video) is shown Wim. et al. [13] and Jonathon the heart rate. Each pulse is de and diastole phase of the heart heart pumps out the oxygenated the blood volume; during the d the blood volume. We used component analysis (PCA) to most representative of the signa a very well-known method use is used to reduce the dimension FIGURE 58: Proposed smart phone camera vital signs measurement system model [72].
ied. The changes of the wave after reflection convey data that can be decoded to sense information about a user such as motion, heartbeat rate, sleep, and other emotions without using body contact sensors. Some challenges encountered when using this technique are as follows: 1) Bandwidth: The range resolution of FMCW is constrained by bandwidth. Higher resolutions means higher bandwidth [88]. 2) Power consumption: FMCW radar systems are able to provide fine range and Doppler information, nevertheless, the power consumption of these system is high [12]. 3) Orientation: Since the FMCW-based system sends a continues signal in a given direction, the orientation of the subject is a factor where for example if the chest is being observed for respiratory monitoring some critical signals may be missed if the patient's thorax is not facing the antenna because the signal is weak [34]. 4) Accuracy: Accuracy is one of the most important attribute for a vital signs monitoring system in an AAL environment for instance. In the FMCW-based radar set-up, the FFT is used to measure the range or property of the detected subject. Nevertheless, FFT is not always accurate and accuracy-improving algorithms need to be utilized to enhance the precision of the system [89]. Some signal processing algorithms used are windowing, zero-padding, chirp-z transform, and frequency estimators. 5) Noise and Cost: Authors in [26] mention that the main problem that persisted in their design of the Doppler radar vital sign monitoring system is finding ways to minimize the background noise during motion while maintaining low cost. 6) Interference: Authors in [28] note that interference from things such as moving fans and neighbors is one of the biggest challenge faced by the design of EZ-Sleep used to detect insomnia. This is because the RF signal provided to the classifier can easily reflect the motion from the interference.

B. CHALLENGES FACING CSI-BASED SENSING
In this section, we look at the different challenges faced by CSI-based systems discussed in section III B. CSI contains tremendous data about the environment, this data can be extracted to create new information. Nevertheless, there are numerous challenges faced by CSI-based systems such as: 1) Motion: Authors in [46] proposed a contactless respiration monitoring system using RSS. Moreover, the authors discuss further research issues such as the limits of distinguishing between subject's body motion and actual breathing motion. Body motion such as arm or body movement occupies most of the sensing frequency, therefore, the main research question posed by [46] is to investigate weather some of the frequency channels remain unaffected despite the body motion interference. Moreover, in a Wi-Fi CSI-based gesture recognition system the main problem was that the gesture movements produced minute Doppler shift changes which was very hard to detect [47,90]. 2) Subject's characteristics: In breathing monitoring using RSS, the characteristics of the patient such as age, gender, and size should be taken into consideration. For instance, the chest movement of a child is considerably faster than that of an adult, hence, it is expected that monitoring a infant's chest movement is more challenging. This factors must be taken into consideration when designing CSI-based respiratory monitoring system [46]. 3) Unreliability: According to [90], RSS provides only a rough estimate about the information of a wireless channel and does not provide detailed information about the effect of multipath. Moreover, RSS measurements are single values per packet which represent SINR over the channel. 4) Cost: In large crowd sensing application, high training cost is the main limiting factor in CSI-based technique, moreover, it is very difficult to get the actual ground number when a large group of people is involved [90]. 5) Range: Wi-Fi-based sensing techniques have limited range of detection compared to other techniques such as ulta-wide band (UWB), hence, this could limit their application [48,90].

C. CHALLENGES FACING WSN-BASED SENSING
The proposed technique used by [57] uses ultra low-power radio-frequency transceiver to detect the motion and breathing of the patient during sleep. Some challenges faced by this scheme are as follows: 1) Security: It is critical that the information of a patient is protected. The sensed signal from the patient should be secure and with limited access. Moreover, the the sensed signal from one patient should not be mistaken for another [91,92]. 2) Reliability: The system must be very reliable. Sensing and wireless communications errors must be reduced because an undetected signal could lead to fatal consequences. 3) Intelligence: Intelligent systems can also be integrated to the WSN paradigm to equip the sensors with the ability to continuously monitor the patient or perform an event driven surveillance. 4) Transmission power: The transmit power used by the sensors should be very low to prevent interference and radiation [93,92], which could be harmful to a person. However, the power should also be enough to provide power for sensing. Potential solutions for the power problem in WSN have been proposed in [91]. 5) Mobility: The communication path used for one data transfer is outdated in case of mobile WSN nodes. This means extra processing power needed to compute new routs as well as extended delays. 6) Cost: Expenditure is incurred by the routing algorithm in terms of bandwidth, delay, jitter, among other performance metrics. Based on the requirements of the system, a trade-off is needed between cost, QoS and resources. Moreover, sustaining a route table is also costly [94]. In addition, WSN have limited power, computation, and memory. 7) Delay: For real-time monitoring systems, delay is one of the most critical factor and hence, must be minimized. Delay can be caused by network characteristics such as node mobility, interrupted connection, and rerouting [94]. 8) Bandwidth: In WSN bandwidth is a limited resource and systems must carefully select routes to utilize the available bandwidth while maintaining system performance [94].

D. CHALLENGES FACING CAMERA-BASED SENSING
Even though the use of cameras for persons monitoring such as fall detection is well researched, there are still some technological challenges to overcome, such as: 1) Sensitivity: In the case of the use of thermal cameras for instance, infrared thermography (IRT) which was a popular mechanism used to detect patients carrying infectious diseases suffers from low sensitivity and specificity because monitoring high temperature alone is not sufficient to detecting an infected patient [24]. 2) Occlusion: If there are many obstacles such as furniture in a room, there might be a need to add more cameras [95]. 3) Acceptance: Many subjects are very reluctant to accept this form of sensing in their homes or around them. They user usually feels like they will loose their privacy [95].

VI. LESSONS LEARNED AND RESEARCH DIRECTION
In this section we summarise the lessons learned from this survey as well as the future expected direction of research in this field. After presenting the contactless technology to supervise human vital signs with the different techniques and systems classified as FMCW-radar, channel state information, wireless sensor network, and camera based, we summarize and compare these techniques according to supervised vital signs, used technique, and the used band as shown in Table (1), as well as according to the machine learning technique used for classification as shown in Table (2).

VII. CONCLUSION
Vital signs monitoring is a necessary, useful, and lifesaving method used in healthcare to monitor and follow the subject's well-being and trigger the required intervention, especially in high-risk scenarios. contactless sensor technology is a new technology applied in medicine that ensures that the monitoring of a patient is done in a remote manner with no patient-healthcare provider physical contact and without asking the patient to wear any body-contact sensors. In this survey, we studied the current approaches and techniques used to monitor human vital signs using contactless (devicefree or sensorless) wireless technology. Our analysis yields the following results: contactless sensing is a viable sensing technology in the field of medicine with the ability to save human life and provide valuable data while ensuring comfort to the subject since no direct contact is required. In addition, utilizing contactless sensor technology will reduce direct and physical interactions between the patient and the care giver, hence lowering the chances of spreading contagious diseases. Moreover, existing techniques are modeled around FMCWradar (which is a widely used technique as the basis of other sensing methods), channel state information, wireless sensor networks, and camera. Also, we discussed some of the main challenges faced by each of the contactless sensing techniques as well as enabling hardware technologies.