Ieeepro techno solutions 2013 ieee embedded project - analysis of an indoor biomedical radar

IEEE TRANSACTIONS ON MICROWAVE THEORY AND TECHNIQUES, VOL. 61, NO. 5, MAY 2013 2061

Analysis of an Indoor Biomedical Radar-BasedSystem for Health Monitoring

Marco Mercuri, Student Member, IEEE, Ping Jack Soh, Student Member, IEEE, Gokarna Pandey,Peter Karsmakers, Guy A. E. Vandenbosch, Fellow, IEEE, Paul Leroux, Senior Member, IEEE, and

Dominique Schreurs, Fellow, IEEE

Abstract—Innovative technology approaches have been increas-ingly investigated for the last two decades aiming at human-beinglong-termmonitoring. However, current solutions suffer from crit-ical limitations. In this paper, a complete system for contactlesshealth-monitoring in home environment is presented. For the firsttime, radar, wireless communications, and data processing tech-niques are combined, enabling contactless fall detection and tag-less localization. Practical limitations are considered and properlydealt with. Experimental tests, conducted with human volunteersin a realistic room setting, demonstrate an adequate detection ofthe target’s absolute distance and a success rate of 94.3% in distin-guishing fall events from normal movements. The volunteers werefree to move about the whole room with no constraints in theirmovements.

Index Terms—Fall detection, health monitoring, radar remotesensing, tagless localization, Zigbee communication.

I. INTRODUCTION

T HE senior citizen population of older than 60 years hasbeen steadily increasing worldwide. This situation has re-

sulted in a growing need for novel assistive technologies thatenable routine long-term home monitoring [1]. Conventionalapproaches are based on devices attached to the patient’s body,involving pressing a button, e.g., worn as a necklace, in emer-gency situations [2], [3]. However, persons in such situationmay already be unconscious or no longer sufficiently reflexiveto do so. The ideal solution is therefore a contactless health-monitoring system, avoiding the need for actions by the elderlyperson.

Manuscript received October 15, 2012; revised February 04, 2013; acceptedFebruary 07, 2013. Date of publication March 07, 2013; date of current versionMay 02, 2013. This work was supported by FWO-Flanders and KU LeuvenGOA Project.M. Mercuri, G. Pandey, G. A. E. Vandenbosch, and D. Schreurs are

with the Department of Electrical Engineering, Katholieke UniversiteitLeuven, 3001 Leuven, Belgium (e-mail: [email protected],[email protected]; [email protected]; [email protected]).P. J. Soh is with the Department of Electrical Engineering, Katholieke Uni-

versiteit Leuven, 3001 Leuven, Belgium, and also with the School of Com-puter and Communication Engineering, Universiti Malaysia Perlis, 02600 Arau,Perlis, Malaysia (e-mail: [email protected]).P. Karsmakers and P. Leroux are with the Department of Electrical

Engineering, Katholieke Universiteit Leuven, Leuven, Belgium, and alsowith Thomas More University, 2440 Geel, Belgium (e-mail: [email protected]; [email protected]).Color versions of one or more of the figures in this paper are available online

at http://ieeexplore.ieee.org.Digital Object Identifier 10.1109/TMTT.2013.2247619

In the last two decades, attention has been focused mainly oncontactless vital-sign monitoring [4]. These academic develop-ments are based on radar techniques implemented as a singledevice sensor, e.g., continuous-wave (CW) Doppler radar [5]or ultra-wideband impulse-radio (UWB-IR) radar [6]. The ca-pabilities of such devices are limited to heartbeat and respira-tion-rate monitoring under “ideal conditions,” e.g., constrainedto a single motionless person (i.e., seated or lying down). Inrecent years, remote fall detection has become a primary in-terest in connection to health monitoring in home environments[7]–[9]. Instances of such investigations are systems based onvideo cameras [10], floor vibration, and acoustic sensors [11],[12]. Besides privacy concerns, the use of video camera sys-tems is still troubled by issues related to low light environments,field of view, and image processing, resulting in a success rateof 90% using two cameras [10]. Successes in floor vibration andacoustic sensors are similarly limited due to environmental in-terference and background noise. Moreover, they are also lesseffective in detecting cases of “soft” human falls, defined as afall after the individual collides with an object (e.g., table, chair,or carpet) [11], [12].Due to the disadvantages of existing fall-detection technolo-

gies, there is a need for further solutions. An alternative ap-proach based on radar techniques has been demonstrated by theauthors [13], [14]. The approach is able to detect fall eventsand to localize a person without the need for a radio-frequencyidentification (RFID) tag on the person. However, to satisfy thespectrum mask requirements and to have a practical device, tra-ditional radar architectures are not suitable solutions for this ap-plication, and the waveform to be transmitted needs to be prop-erly designed [15].In this work, a full system is proposed enabling indoor,

noninvasive fall detection and tagless localization. It combinesradar, wireless communications, and data-processing tech-niques. Moreover, it has been designed to satisfy the Europeanand Federal Communications Commission (FCC) UWB spec-trum masks, and it can also be potentially connected to medicalmonitoring personnel to provide a prompt alert in the event ofemergencies.In Section II, the monitoring principle and underlying

theory are presented. The sensor architecture is explainedin Section III, and the experimental results are discussed inSection IV.

II. OPERATIONAL PRINCIPLE

Fig. 1 shows a simplified block diagram of the proposedhealth-monitoring system. It consists of a sensor, combining

0018-9480/$31.00 © 2013 IEEE

2062 IEEE TRANSACTIONS ON MICROWAVE THEORY AND TECHNIQUES, VOL. 61, NO. 5, MAY 2013

Fig. 1. Simplified block diagram of the health-monitoring system.

Fig. 2. Designed radar waveform.

both radar and wireless communications features, and a basestation for data processing. A radar waveform is generatedand sent to the target, and then its reflected echo, containingspeed and absolute distance information, is collected by thereceiver. The resulting baseband signals are digitized andtransmitted wirelessly to a base station that consists of a Zigbeemodule, a laptop, and a microcontroller. The latter collects andtransfers the data received of the Zigbee module to the laptopto determine remotely the target’s absolute distance and todistinguish a fall event from normal movements (e.g., walkingor sitting down). The data processing is not performed by thesensor in order to avoid complex processor on board, reducingcosts, size, and energy consumption. Moreover, this representsa flexible solution if multiple sensors will be used in the future.In fact, the base station should combine and process multipleinformation simultaneously.The radar waveform is based on a hybrid approach presented

by the authors in [15]. It consists of a single tone, at5.8 GHz in the ISM band, alternated with a stepped frequencyCW (SFCW) waveform working in the UWB band, as shown inFig. 2. Each tone lasts 1 s and is used to continuously detect thespeed of a person using the Doppler concept. The SFCW wave-form is used to detect the target’s absolute distance. It consistsof coherent CW pulses (called burst), the frequenciesof which are increased from pulse to pulse by a fixed increment

25 MHz. Each pulse is 50 s long, resulting in aburst duration of 2 ms, while its total band is 1 GHzpositioned between 6 and 7 GHz, enabling a smallest resolutionof 15 cm. The full waveform is 1.002 s. In comparison to [15],we introduce some changes in Section II-B to tackle some prac-tical issues.

Here, the waveform spectral analysis is first discussed. Next,these considerations are used to justify the choice of the wave-form parameters and to explain the operations of compensa-tion and calibration in data processing. Moreover, the techniqueto distinguish fall events from normal movements is also de-scribed.

A. Spectral Analysis

Neglecting the initial amplitudes, the transmitted waveformcan be expressed as

(1)

for and

(2)

for , with . If the waveform isreflected by a target at a distance , the received signal willthen be represented as

(3)

and

(4)

The output of the IQ mixer (Fig. 1) can be modeled as theproduct of the received signal with a copy of the transmittedsignal followed by a low-pass filter. For a quadrature sampling,it is given as

(5)

with

(6)

for the single tone and

(7)

for the SFCW signal, where

(8)

is the residual phase noise, which is negligible, while is thecontribution of the phase shift at the target surface (approx-imately 180 ) plus additional phase difference between themixer and the antenna. Only one sample for pulse width hasbeen considered. In the case of a moving target, the rangecan be written as

(9)

for and as

(10)

for . and are defined as the ranges tothe target at the particular time and , respectively.

MERCURI et al.: ANALYSIS OF AN INDOOR BIOMEDICAL RADAR-BASED SYSTEM FOR HEALTH MONITORING 2063

Combining (9) and (10) with (6) and (7), the phase of the base-band signal becomes

(11)

and

(12)

The last two equations are essential in the waveform designand for data processing, which consequently determines thehardware requirement and system complexity. Equation (11)represents that the signal is proportional to the target’s speed,and it also dictates its sampling rate. On the other hand, (12) isused to determine the target’s absolute distance. In particular,the first two terms represent the case of a stationary targetsince they are not influenced by speed . The first termrepresents a constant phase shift, which is not of any practicalsignificance, while the second term is a multiplication betweenthe rate of frequency change and the signal round-triptime . The latter term provides the distance information.The remaining terms are undesirable. The third term repre-sents the Doppler frequency shift due to target motion, whichadds to the frequency shift of the second term, resulting in ashift of target distance from its true value. The fourth termexists due to the interaction of the frequency-varying stepwaveform with the target’s motion, resulting in a spread ofthe target peak. Compensating these effects is difficult due tothe unknown instantaneous target velocity. The final term isfrequency-dependent, which adds a fixed frequency shift to theactual target distance.

B. Waveform Design

The proper choice of the parameters , , and allows tosatisfy the spectrum masks while simultaneously having suffi-cient transmit power to track a person in a typical room setting.The waveform has been designed to monitor continuously the

speed of the target and to determine its absolute distance every1 s. Assuming a maximum target movement/falling speed

of 2 m/s, the I/Q baseband signals containing speed informa-tion should be acquired at least with a sample rate of about 6.5ms. This value is defined considering the Nyquist theorem and(5) and (11), where is 5.8 GHz. Experimental evaluationson human volunteers have failed to demonstrate significant im-provements even when the sample time is decreased below 7.2ms. This indicates that themaximum speed produced by the sub-jects’ movements is about 1.8 m/s, lower than the 2-m/s initialassumption. However, a sample rate of 4 ms has been chosenproviding a sufficient margin to detect movements of humanswith different weights.To avoid losses in speed samples and to simultaneously de-

termine the target’s corresponding distance, the CW pulsesare sent, every 1 s, in between speed samples. This involves amaximum burst interval of 4 ms. In order to have an unam-biguous range of 5 m, the increment from the second term

of (12) should be fixed to 30MHz. This theoretical value cannotbe used in practice since the third and fifth terms shift the targetposition outside the unambiguous range. The third, and, conse-quently, the fourth term, can be neglected by properly choosingthe burst duration such that the target is considered staticduring this interval, i.e., speed . The burst interval of 2ms is sufficiently short to consider the static target assumptiontrue. Moreover, this ensures also that the downconverted signalsof the SFCW waveform consist of I/Q direct current (dc) levels.This means that only two samples can be acquired per pulsewidth fixing the sample time of SFCW waveform to 50 s.On the contrary, the fifth term, which is not negligible, is com-

pensated through a calibration procedure, resulting, for the pre-sented sensor, in approximately 1 m fixed shift from the correcttarget distance. These considerations resulted in to be fixedat 25 MHz to enable a 6-m coverage: 5 m for the desired targetrange and another 1-m allowance to cater for the fixed shift.

C. Data Processing

The digitized I/Q baseband signals are processed remotely inthe base station using MATLAB. The samples are first related asin (5) and then split according to category to be processed sepa-rately, namely whether the samples serve for speed monitoringor for absolute distance detection.A movement classification based on a least-square support

vector machine (LS-SVM) approach [16] is applied to analyzethe speed samples in order to distinguish falls from normalmovements. The key relies on the different changes in speedexperienced during a fall or a normal movement. During a fall,in fact, the speed continuously increases until the sudden mo-ment when the fall stops abruptly. During a normal movement,the Doppler signal experiences a controlled movement. Moreprecisely, while a person is sitting or lying down, the speedfirst gradually increases and then decreases to a smooth stop,whereas, during a walk, instead, the speed is quite constantover time. The acquired speed signals, or activities, are usedto build a data set. Before learning a model, the raw radar dataare preprocessed. Each activity is grouped in a segment of 2 s,which is considered sufficient to cover the details of the activ-ities. Given such segments, the data is then transformed usingfast Fourier transform (FFT) from which only the magnitudespectrum is retained. Two alternatives have been considered,namely: 1) to compute the FFT directly on the complete seg-ment and 2) to use the short time FFT (STFT). In the lattercase, the segment is first chopped into 50% overlapping frameswhich are each multiplied with a Hamming window after whichthe FFT is computed on each of these frames. As opposed tothe FFT, the STFT can represent time-dependent structures andresults in higher performance in case of signals that experiencea gradual change in velocity. In the LS-SVM framework, achoice concerning a kernel function must be made. A kernelfunction can be seen as a similarity function which defines howsignals are compared. For the first option, the LS-SVM is usedin combination with a linear and radial basis function (RBF)kernel on the FFT data, while to handle the STFT data, theLS-SVM is combined with the global alignment (GA) kernel.Prior to the learning phase, the data were standardized such thateach dimension has zero mean and unit standard deviation.


Fig. 3. Block diagram of the sensor.

The target’s range profile is determined applying the inverseFFT (IFFT). However, this operation has as main challenge thedistinction of the target’s reflection from the effects of backscat-tering and cross-coupling between the two antennas. The latterinvolve strong reflections that overwhelm the much weakerreflected/received signal, resulting in the inability to acquireany meaningful target information. In addition, the presenceof cluttering due to furniture in a practical environment mustalso be considered. Both factors can be eliminated by a com-pensation that consists of determining an environmental rangeprofile, characterizing the total contribution from both crosscoupling and cluttering. Its magnitude is then subtracted fromthe range profiles obtained with the target in the room. Then, therange profile is shifted by a fixed value to compensate for theeffect of the fifth term in (12). This value is obtained throughcalibration using a flat metal plate, placing it at a well-knowndistance from the antennas to evaluate its range profile. Thevalue of its corresponding peak is read and subtracted from thiscalibrated distance, thus yielding .However, it must be noted that this operation can be only suc-

cessful if the backscattering and crosstalk are strongly reduced,as explained in Section III.

III. SENSOR ARCHITECTURE

Fig. 3 shows the overall sensor block, which consists of threemain parts, namely the radar module, the AT86RF231 Zigbeemodule, and the microcontroller. The sensor block is combinedwith two-element dual-band antennas, designed to operate atboth the radar and Zigbee ISM frequency bands. As will beexplained in Section III-B, the main challenge is to reduce thebackscattering and crosstalk effects while presenting also a

semispherical radiation pattern. No off-the-shelf antennas areavailable to deal with such challenges. A wideband circulatorwith one antenna could represent a more compact solution.However, although it does not experience the cross talk, themismatch between the antenna and its feed line involves astrong reflection that overwhelms the reflected signal. Thiseffect cannot be reduced below a practical value (i.e., 30 dB)in the whole radar bandwidth, as it was possible with the crosstalk in the case of two antennas.The microcontroller controls the RF switch to connect alter-

nately the radar transmitter and the Zigbee module to the trans-mitter antenna, labeled TX in Fig. 3. However, the latter is usedby the Zigbee module both to transmit and to receive frames.In fact, the sensor may receive commands from the base sta-tion. This happens to synchronize base station and sensor afterpower-on.Batteries provide the power supply to the sensor.We will now detail the radar module and the antenna design.

As the Zigbee module is an off-the-shelf component, it is notfurther described.

A. Sensor

The radar module integrates a fractional- PLL with a wide-band voltage-controlled oscillator (VCO), a 5.8–7-GHz wide-band Wilkinson power divider, an RF switch, a low-noise am-plifier (LNA), a gain block, an in-phase and quadrature (IQ)mixer, and baseband filters and amplifiers. The microcontrollerprograms the synthesizer of the phase-locked loop (PLL) to gen-erate the radar waveform, described in Section II-B, with a max-imum transmit power level of 2 dBm. A 5.8-GHz single toneis produced at every 1-s interval, immediately followed by thisSFCW waveform. The sensor requires approximately 30 togenerate a new frequency, by which the first 20 s is to programthe PLL, while another 8 s is to take into account the VCOmaximum settling time. The combined LNA and gain block ischosen to provide a total gain of approximately 30 dB to avoidsaturation of the gain block. This is considering that the two an-tennas present a cross coupling of approximately 35 dBwithinthe operation bandwidth and that the target could move veryclose to the antennas, producing a high reflection and, thus, highreceived power.The bipolar I/Q baseband signals are then acquired and digi-

tized by an analog-to-digital converter (ADC) inside the micro-controller. Since it works with unipolar signals, between 0 and3.2 V, the outputs of the mixer are bandpass-filtered and ampli-fied and then a 1.6-V dc level is added to position the signals tothe center of ADC’s dynamic range. The low-pass filter servesboth as an antialiasing filter and as a charge reservoir for theADC’s switched capacitor input stage.The 10-b ADC requires approximately 20 s to acquire the

samples and to digitize the two I/Q components before thenext frequency can be generated. These acquired and digitizedsamples, containing the target’s information, are then sent tothe base station. This transmission must be properly managedwhile simultaneously acquiring the monitoring signals. Sincethe Zigbee protocol only transmits frames containing bytes,each of the acquired 10-b samples should be ordered into twobytes. However, since the I and Q samples are acquired at the


same time, they can be mapped in 3 bytes. Sampling the signalcontaining speed information at 250 Hz for a duration of 1 s,and considering the I/Q components, the sensor is expected toaccumulate 502 samples. This translates into 753 bytes. Thetransmission of these frames must be executed in betweenspeed sampling instants. For that reason, the speed samples aremapped in frames of 75 bytes, 25 I samples, and 25 Q samples.The transmission of such frames requires about 2.6 ms each,which is much shorter than the duration of the acquisitionperiod. Each frame is therefore filled after 100 ms and thentransmitted before the next new sample is acquired. However,the transmission of the last 3 speed bytes is achieved with thetransmission of the 80 I/Q burst samples, i.e., 120 bytes. Thisinvolves the transmission of two consecutive frames, each with63 and 60 bytes, respectively. The former, containing also thelast 3 speed bytes, is transmitted immediately after the acqui-sition of the first new I/Q speed, while the latter is transmittedafter the acquisition of the second I/Q speed samples.The 5.8-GHz tone is generated immediately after the SFCW

waveform. However, the first I/Q speed samples are acquiredafter about 2 ms after the acquisition of the last burst samples.This is done to hold the 4-ms speed sampling time as a constant.

B. Antenna

Regarding the antennas for the sensor, two identical dual-band antennas based on the coplanar waveguide (CPW) bow-tiehave been designed for the sensor. According to the system re-quirements, the antennas are to be used for target monitoringbetween 5.8 and 7 GHz and for sensor-to-base station commu-nication at 2.45 GHz. Besides resonance prerequisites, a set ofstringent design requirements is imposed, which includes sizecompactness and radiation characteristics. First, the backscat-tering within the frequency band where the sensor is used tolocate the target and to determine its speed must be strongly re-duced to enable maximum forward power towards the target.A failure in doing so will obviously limit the forward transmitpower, resulting in a weak reflection from the target and in awaste of energy. Moreover, since the sensor is intended to bemounted either to the wall or to the ceiling, the relative reflec-tion of the wall or ceiling will bury the much weaker target’sreflection.Another challenge that arises when attempting to arrive at a

compact system size is the problem of the inter-element antennacross coupling. As was explained earlier, such large cross cou-pling enables the power “overflow” from the transmitting an-tenna to the adjacent receiving element involving a strong re-flection. These two undesired effects therefore involve the de-crease of the total receiver’s gain to avoid the saturation both ofthe amplifiers and of the ADC. Thus, the longer the distance tothe target is, the weaker is the reflection such that it is no longerperceptible by the ADC’s resolution, as it is buried in the noise.These considerations, plus the objective for a 5 5 m

equivalent room area coverage lead to the following designrequirements: 1) a compact antenna footprint, approximatelyequivalent to the overall sensor’s size; 2) an antenna beamwidthof approximately 60 for a wide coverage in the azimuth plane;and 3) a minimum gain of approximately 6 dB with unidirec-tional forward radiation.

Fig. 4. Fabricated prototype of the proposed two-element bow-tie antenna.

Fig. 5. Measured antenna -parameters.

The designed antennas are shown in Fig. 4. Each bow-tie el-ement is fabricated on a 1.524-mm-thick Rogers RO4003 sub-strate, with a relative permittivity of 3.38 and loss tangent

of 0.021. Two identical antennas are then secured onto18.5-mm-thick foam substrates with prior to place-ment on a common ground plane sized at 112 42 mm . Be-sides minimizing antenna back radiation, such arrangement en-ables bandwidth broadening across 6 to 7 GHz. The close-prox-imity placement of the antennas are almost certain to inducecross coupling, thus a 39 31 1 mm vertical wall is in-troduced between the two antennas, placed 19.5 mm from eachelement.The designed antenna combination has successfully achieved

its intended purpose. In fact, the use of the common groundplane reduces strongly the backscattering, while the use of themetal wall suppresses the cross coupling besides main-taining a satisfactory reflection coefficient , as shown inFig. 5. For this fabricated prototype, the results between6.6 and 7 GHz are indeed slightly above 10 dB. However,the highest/worst within this range is 7 dB, which trans-lates to a maximum of 20% reflected RF power. This is mainlydue to both the soldering and the difficulty of realizing a com-pletely flat foam spacer. Similarly for the system, this meansthat a maximum of 20% less power is transmitted, resulting in alower target-reflected signal detected by the receiver. This mayproduce an error in absolute distance detection when the targetis really far away from the antenna, i.e., more than 5 m. Atthese distances, reflections from the target received within the6.6–7-GHz band will not be perceptible by the receiver despitethe sufficient signal amplitudes received in the 6–6.6-GHz band.This then potentially causes a wrong target location interpreta-tion in the range profile. However, the measurements indicatedthis is not a problem in a 5 5 m sized room.

IV. EXPERIMENTAL RESULTS

Experimental evaluations have been conducted with realhuman volunteers who are allowed to move freely about thewhole room. The sensor has been fixed to the wall at a height of1.5 m. Furniture and metallic shelves are deliberately included


Fig. 6. Range profile of a target 4 m away the antennas (a) before and (b) aftercompensation and calibration.

to enable the existence of clutter and reflections, mimickinga typical room setting. Falls are mimicked with two differenthuman volunteers, with a similar 1.75-m height but differentweights to enable the evaluation of different fall speeds. Thefirst and second subjects’ weights are 90 and 75 kg, respec-tively. However, only frontal falls have been evaluated atdifferent locations at radial distances in the whole room, usingan inflated mattress to avoid injuries. Also, only one personwas present in the room at a time.The prototype implementation of the proposed system is

proven to be accurate, as illustrated in the following results.

A. Tagless Localization

Fig. 6(a) represents an initial range profile of a person 4 maway from the antennas, prior to the compensation and calibra-tion process. The peak, which is supposed to indicate the target’sabsolute distance, is totally overwhelmed by the undesired re-flections originating from the clutter and the antenna’s crosscoupling. In particular, the latter is the most dominant effect,as emphasized by the strong peak at 90 cm. This also indicatesthe position of the two antennas in the range profile, which intheory should be at 0 m. Therefore, the peak results shifted inrange due to the effect of the fixed frequency shift , the valueof which can be read and used during calibration.After applying the compensation and the calibration steps,

the target’s peak can be perfectly distinguished, as shown inFig. 6(b).However, it should be noted that the radar has a range resolu-

tion of 15 cm, which means that any target’s physical distancewill be rounded to the nearest location/resolution provided by

Fig. 7. Speed signal during (a) a walking movement and (b) a fall event. Thefrequency of the signal is proportional to the radial velocity of the person duringthe movement. An inflating mattress has been used when invoking falls. For thatreason, the corresponding speed signals do not stop suddenly, but there is alsothe effect of the rebounds on the mattress.

TABLE IFALSE POSITIVES AND SUCCESS RATE USING FOUR CLASSIFICATION MODELS

the radar. This also establishes the maximum error in localiza-tion of 7.5 cm.

B. Fall Detection

A data set was built containing 70 activities measured fromtwo persons. In particular, 20 walking signals have been ac-quired for each person, who was allowed free movement inthe whole room, and 30 fall signals have been acquired witheach subject located at known distances from the antennas.Fig. 7(a) and (b) shows the speed signals, respectively, duringa walking and a fall movement. The LS-SVM model is trainedusing the data of a single person (target 1) and then validatedusing the data from the other person (target 2). This process isrepeated two times since data from two persons were available.The results of the classification are shown in Table I. It canbe seen that the GA kernel, which incorporates time-depen-dent information, presents a success rate in distinguishing fall


events from normal movements of 94.3% outperforming thelinear and RBF kernels. This is expected since falls exhibit atime-dependent structure as the speed increases continuouslyuntil the sudden moment when it stops abruptly. An alternativemethod called dynamic time warping (DTW) combined witha Euclidean distance measure is frequently used to classifysequences of vectors. In order to compare the LS-SVM withGA solution to this standard method, an additional experimentwas included. The results show that the GA kernel slightlyoutperforms the DTW alternative on this data set.By further reducing the crosstalk, it will be possible to in-

crease the total receiver gain and, consequently, to improve theaccuracy of velocity detection, besides the detection range ex-tension. Also, the system’s sampling rate influences the accu-racy of the speed detection, i.e., the higher it is, the more ac-curate is its detection. However, this will significantly compli-cate the system since the duration of the SFCW pulses must beshortened, and the ADC sample rate increased. This will involvehigher power consumption and a larger number of transmittedframes to the base station. However, it should be noted that, inthis application, the main goal is to detect the changes in speedand not to determine how accurate is the value of the instanta-neous speed.

V. CONCLUSION

A radar-based system has been proposed as a new approachfor contactless fall detection and tagless localization in an in-door environment. This is inline with the growing need for novelhealthcare solutions. The full system combining radar, wirelesscommunications, and data-processing techniques has been ana-lyzed and described. The prototype implementation has been de-signed to satisfy the European and FCC UWB spectrum masks.Practical problems, such as backscattering and crosstalk, havealso been addressed. Experimental evaluations with real humansubjects have demonstrated accurate detection of the target’s ab-solute distance and fall events.The next step is to combine multiple sensors in a wireless

sensor network configuration, in order to monitor multiple per-sons and to increase the accuracy and coverage area, beyond oneroom. The final application is automated remote monitoring inprivate homes, although it can also be adopted for nurse-call so-lutions in nursing homes.

REFERENCES

[1] O. Boric-Lubecke and V. M. Lubecke, “Wireless house calls: Usingcommunications technology for health care and monitoring,” IEEEMi-crow. Mag., vol. 2, pp. 43–48, Sep. 2002, in .

[2] E. Kantoch, J. Jaworek, and P. Augustyniak, “Design of a wearablesensor network for homemonitoring system,” inProc. Federated Conf.Comput. Sci. Inf. Syst., Sep. 18–21, 2011, pp. 401–403.

[3] V. M. Lubecke and O. Boric-Lubecke, “Wireless technologies in sleepmonitoring,” in Proc. IEEE Radio Wireless Symp., Jan. 18–22, 2009,pp. 135–138.

[4] L. Changzhi, J. Cummings, J. Lam, E. Graves, and W. Wenhsing,“Radar remote monitoring of vital signs,” IEEE Microw. Mag., vol.10, no. 1, pp. 47–56, Feb. 2009.

[5] A. D. Droitcour, O. Boric-Lubecke, V.M. Lubecke, J. Lin, and G. T. A.Kovac, “Range correlation and I/Q performance benefits in single-chipsilicon Doppler radars for noncontact cardiopulmonary monitoring,”IEEE Trans. Microw. Theory Tech., vol. 52, no. 3, pp. 838–848, Mar.2004.

[6] I. Y. Immoreev, S. Samkov, and T.-H. Tao, “Short-distance ultra wide-band radars,” IEEE Aerosp. Electron. Syst. Mag., vol. 20, no. 6, pp.9–14, Jun. 2005.

[7] S. R. Lord, C. Sherrington, and H. B.Menz, Falls in Older People: RiskFactors and Strategies for Prevention. Cambridge, U.K: CambridgeUniv., 2007.

[8] Y. Xinguo, “Approaches and principles of fall detection for elderlyand patient,” in Proc. 10th Int. Conf. e-health Networking, Applicat.Services, Singapore, Jul. 7–9, 2008, pp. 42–47.

[9] N. Noury, A. Fleury, P. Rumeau, A. K. Bourke, G. O. Laighin, V.Rialle, and J. E. Lundy, “Fall detection—Principles and methods,” inProc. 29th Annu. Int. Conf. IEEE Eng. Med. Biol. Soc., Lyon, France,Aug. 22–26, 2007, pp. 1663–1666.

[10] M. Yu, S. M. Naqvi, and J. Chambers, “A robust fall detection systemfor elderly in a smart room,” in Proc. Int. Conf. Acoust. Speech SignalProcess., Dallas, TX, USA, Mar. 2010, pp. 1666–1669.

[11] Y. Zigel, D. Litvak, and I. Gannot, “A method for automatic fall detec-tion of elderly people using floor vibrations and sound—Proof of con-cept on human mimicking doll falls,” IEEE Trans. Biomedical Eng.,vol. 56, no. 12, pp. 2858–2867, Dec. 2009.

[12] L. Yun, Z. Zhiling, M. Popescu, and K. C. Ho, “Acoustic fall detectionusing a circular microphone array,” in Proc. Annu. Int. Conf. IEEE Eng.Med. Biol. Soc., Sep. 4, 2010, pp. 2242–2245.

[13] M. Mercuri, D. Schreurs, and P. Leroux, “SFCW microwave radar forin-door fall detection,” in Proc. IEEE Topical Conf. Biomedical Wire-less Technol., Santa Clara, CA, USA, Jan. 15–18, 2012, pp. 53–56.

[14] M.Mercuri, P. J. Soh, L. Boccia, D. Schreurs, G. A. E. Vandenbosch, P.Leroux, and G. Amendola, “Optimized SFCW radar sensor aiming atfall detection in a real room environment,” in Proc. IEEE Topical Conf.Biomedical Wireless Technol., Austin, TX, USA, Jan. 20–23, 2013, pp.4–6.

[15] M. Mercuri, D. Schreurs, and P. Leroux, “Optimised waveform de-sign for radar sensor aimed at contactless health monitoring,” Electron.Lett., vol. 48, no. 20, pp. 1255–1257, Sep. 2012.

[16] P. Karsmakers, T. Croonenborghs, M. Mercuri, D. Schreurs, and P.Leroux, “Automatic in-door fall detection based on microwave radarmeasurements,” in Proc. Eur. Radar Conf., Amsterdam, The Nether-lands, Oct. 31–Nov. 2, 2012, pp. 202–205.

Marco Mercuri (S’12) was born in Calabria, Italy,in 1985. He received the B.S. and M.S. degrees inelectronic engineering from the Università dellaCalabria, Arcavacata di Rende, Italy, in 2006 and2009, respectively. He is currently working towardthe Ph.D. degree at the Department of Electrical En-gineering, Katholieke Universiteit Leuven, Leuven,Belgium.His research interests include biomedical ap-

plications of microwave/RF, wireless sensors, andmicrowave/millimeter-wave measurements.

Mr. Mercuri is a student member of the IEEE Microwave Theory and Tech-niques Society (IEEE MTT-S) and the IEEE Engineering in Medicine and Bi-ology Society. He was the recipient of the 2013 IEEEMTT-S Graduate Fellow-ship Award.

Ping Jack Soh (S’09) was born in Sabah, Malaysia.He received the B.S. and M.S. degrees in electricalengineering from Universiti Teknologi Malaysia(UTM), Johor, Malaysia, in 2002 and 2005, respec-tively. He is currently working toward the Ph.D.degree at the Department of Electrical Engineering,Katholieke Universiteit Leuven, Leuven, Belgium.From 2002 to 2004, he was a Test Engineer

working on new products’ test definition for man-ufacturing purposes. In 2005, he joined MotorolaTechnology Malaysia as an R&D Engineer. There,

he worked on the characterization and testing of new two-way radio antennasand RF front-ends. In 2006, he joined the School of Computer and Communi-cation Engineering, Universiti Malaysia Perlis (UniMAP),Perlis, Malaysia, asa Lecturer, and, since 2011, a Senior Lecturer. His research interests includeplanar antennas, flexible/textile antennas, on-body communication, metamate-rials, passive microwave components, and microwave measurements.


Mr. Soh was the recipient of the CST University Publication Award in 2011and 2012, the IEEE Antennas and Propagation Society (AP-S) Doctoral Re-search Award in 2012, and the IEEEMicrowave Theory and Techniques Society(MTT-S) Graduate Fellowship for Medical Applications in 2013.

Gokarna Pandey was born in Karikot Syangja,Nepal, in 1985. He received the B.E. degree in elec-tronics and communication engineering from PokhraUniversity, Pokhara, Nepal, in 2007, and the M.Sc.degree in electrical engineering from KatholiekeUniversiteit Leuven, Leuven, Belgium, in 2012.From 2008 to 2009, he was an Assistant Lec-

turer with the Electronics and CommunicationDepartment, Lumbini Engineering College, Butwal,Nepal. His research interests include the antenna andmicrowave circuit design.

Peter Karsmakers was born on April 14, 1979. Hereceived the M.Sc. degree in electronics-ICT engi-neering from the Katholieke Hogeschool Kempen,Kempen, Belgium, in 2001, and the M.S. degreein artificial intelligence and Ph.D. degree from theKatholieke Universiteit Leuven, Leuven, Belgium,in 2004 and 2010, respectively.From 2004 to 2010, he was a Research Assistant

within the SISTA Research Group, Departmentof Electrical Engineering, Katholieke UniversiteitLeuven, Leuven, Belgium. Since 2001, he has

combined his research activities with teaching at the Thomas More University(formerly Katholieke Hogeschool Kempen), Geel, Belgium, where he iscurrently a Post-Doctoral Researcher with the MOBILAB research group. Hismain research interest are machine learning and biomedical signal processing.

Guy A. E. Vandenbosch (F’13) received the M.S.and Ph.D. degrees in electrical engineering from theKatholieke Universiteit Leuven, Leuven, Belgium, in1985 and 1991, respectively.Since 1993, he has been a Lecturer and, since

2005, a Full Professor with Katholieke UniversiteitLeuven, Leuven, Belgium. He has authored orcoauthored in approximately 150 papers in interna-tional journals and 250 presentations at internationalconferences. His research interests are in the areaof electromagnetic theory, computational electro-

magnetics, planar antennas and circuits, nano-electromagnetics, EM radiation,electromagnetic compatibility, and bio-electromagnetics.Prof. Vandenbosch leads the Working Group on Software within the Euro-

pean Association on Antennas and Propagation (EuRAAP), is chairman of theIEEE Benelux Chapter on Antennas en Propagation, and is secretary of theBelgian National Committee for Radio-electricity (URSI), where he is also incharge of commission E.

Paul Leroux (SM’10) was born in Eeklo, Belgium,in 1975. He received the M.Sc. degree and Ph.D. de-gree in electronic engineering from the KatholiekeUniversiteit Leuven, Leuven, Belgium, in 1999 and2004, respectively.From 1999 to 2004, he was a Teaching and Re-

search Assistant within the MICAS Research Group,Department of Electrical Engineering, KatholiekeUniversiteit Leuven (KU Leuven), Leuven, Belgium.Since 2004, he has been a Professor with the ThomasMore University (formerly Katholieke Hogeschool

Kempen), Geel, Belgium, where he is head of the RELIC research group.Since 2009, he is also an Associate Professor with the KU Leuven, wherehe is coordinator of the Technology Cluster Electricity-Electronics-ICT ofthe multi-campus Faculty of Engineering Technology. He has authored andcoauthored over 80 papers in international journals or conference proceedings.His current research activities focus on radiation tolerant IC design, circuits foroptical communication, and instrumentation and electronics for ultra-widebandand microwave imaging systems.Prof. Leroux received the SCK-CEN Prof. Roger Van Geen Award in 2010

from the FWO and FNRS for his work on IC design for radiation environments.

Dominique Schreurs (F’12) received the M.Sc. de-gree and Ph.D. degree in electronic engineering fromKatholieke Universiteit Leuven, Leuven, Belgium.As a Post-Doctoral Fellow, she was a Visiting

Scientist with Agilent Technologies, ETH. Zürich,and the NIST. She is now a Full Professor withKatholieke Universiteit Leuven, Leuven, Belgium.She is co-editor of two books, contributor to sevenbooks, and author and coauthor of approximately100 journal papers and 300 contributions at inter-national conferences. Her main research interests

concern the (non)linear characterization and modelling of microwave devicesand circuits, as well as (non)linear hybrid and integrated circuit design fortelecommunications and biomedical applications.Prof. D. Schreurs serves on the IEEE Microwave Theory and Techniques

Society (MTT-S) AdCom since 2009, was vice-chair of the IEEE MTT-STechnical Coordinating Committee in 2009–2010, and has been chair ofthe IEEE MTT-S Education Committee since 2012. She was the Chairof the IEEE MTT-S Technical Committee on Microwave Measurements(MTT-11) in 2005–2008. She is a Distinguished Microwave Lecturer for theperiod 2012–2014. She is also associate editor of the IEEE MICROWAVE ANDWIRELESS COMPONENTS LETTERS and International Journal of Microwaveand Wireless Technologies and reviewer for all IEEE MTT-S journals aswell as TPRC member of IMS, RWW, and EuMW. She is also reviewerfor various MTT-S (co-)sponsored conferences and regularly acts as sessionchair at conferences and as judge for student competitions. Beyond IEEE, shealso serves as Technical Chair on the Executive Committee of the ARFTGorganization and initiated the NVNA Users’ Forum as well as the IEEEWomenin Microwaves event at the European Microwave Week.