WIDESEE: Towards Wide-Area Contactless Wireless Sensing

ABSTRACTContactless wireless sensing without attaching a device to the tar-get has achieved promising progress in recent years. However, onesevere limitation in this field is the limited sensing range. This pa-per presents WIDESEE to realize wide-area sensing with only onetransceiver pair. WIDESEE utilizes the LoRa signal to achieve alarger range of sensing and further incorporate drone’s mobility tobroaden the sensing area. WIDESEE incorporates a set of novel so-lutions across software and hardware to overcome two aspects ofchallenges for wide-range contactless sensing: (1) the interferencebrought by the device mobility and LoRa’s lower sensitivity, and(2) the ambiguous target information like location when using just asingle pair of transceiver. We have developed a working prototypeof WIDESEE for human target detection and localization that areespecially useful in emergency scenarios like rescue and terroristsearch, and evaluated WIDESEE with both controlled experimentsand the field study in a high-rise building. Extensive experimentsdemonstrate the great potential of WIDESEE for supporting wide-area contactless sensing applications with a single LoRa transceiverpair hosted on a drone.

ACM Reference Format:. 2019. WIDESEE: Towards Wide-Area Contactless Wireless Sensing. InProceedings of ACM Conference (SenSys’19). ACM, New York, NY, USA,13 pages. https://doi.org/0000001.0000001

1 INTRODUCTIONBesides traditional data communication, in recent years, wirelesssignals have been employed for sensing and have enabled new appli-cations, such as indoor navigation [23, 51, 54], health care [10, 39],and human-computer interactions [16]. Wireless sensing is achievedby using one or more pairs of signal transmitter and receiver, wherethe received signal characteristics can be used to understand the con-textual information of one’s interest (e.g., localization) in a device-free manner (i.e., without additional devices attached on the targetobjects). A wide range of wireless signals have been exploited forcontactless sensing, including ultrasound [43, 50] and various typesof radio frequency (RF) signals [27, 52]. The RF signals – suchas WiFi [15, 27, 51] and RFID [28] – have the characteristics thatenable a variety of real-world sensing applications that would oth-erwise not be possible with other technologies (e.g., conventionalcamera-based [37] or acoustic signal-based approaches [49]). Forexample, RF signals do not require to secure a Line-of-Sight (LoS)between the device and targets as opposed to conventional camera-based systems. Furthermore, RF signals have a much larger sensingrange and penetration ability compared to acoustic signals [49].

ACM acknowledges that this contribution was authored or co-authored by an employee,contractor, or affiliate of the United States government. As such, the United Statesgovernment retains a nonexclusive, royalty-free right to publish or reproduce this article,or to allow others to do so, for government purposes only.SenSys’19, November 10-13, New York, NY, USA© 2019 Association for Computing Machinery.ACM ISBN 978-x-xxxx-xxxx-x/YY/MM. . . $15.00https://doi.org/0000001.0000001

Although promising, one evident issue with existing RF-basedsensing is in its limited sensing range, which hinders their appli-cations in wide-area sensing. This is mainly because the signalsreflected from the target, which contain information related to thecontext of the target, are much weaker than the direct-path signalsbetween the transmitter and receiver. The fact that wireless sensingcaptures information from the reflected signals makes the sensingrange much smaller compared to when the signals are used for com-munication purposes. For example, the current WiFi-based systemsare capable of performing sensing in a room-level range (i.e. approx-imately 3-6 𝑚) [27, 51], whereas RFID or mmWave-based systemsshow an even smaller sensing range of 1-3𝑚 [46, 55].

Recently, efforts have been made to extend the contact sensingrange of RF signals [12, 25, 56]. Ashutosh et al. introduced anapproach to employ multi-hop nodes to track the sensor attached tar-gets that are located deep inside a building structure [12]. In anotherexample, Ma et al. leveraged drones to relay sensing information,which extends the contact sensing range from 5 𝑚 to 50 𝑚 [56].Employing multiple devices or multi-hop transmission schemes canincrease the sensing coverage range. However, such approaches re-quire a complicated process of sensing infrastructure deploymentand could be vulnerable to changes or failure of even a single device.

In this paper, we present WIDESEE– a contactless wireless sens-ing system based on the emerging LoRa technology with only asingle transceiver pair. WIDESEE is designed to push the bound-ary of wide-area sensing. Our key insight is that the low-power,long-range wireless communication capability of LoRa offers a longpropagation range (i.e., several kilometers) and a strong penetrationcapability through obstacles, which in turn can be used to significant-ly increase the sensing range compared to other existing wirelesstechnologies. In this work, as a proof-of-concept, we explore theopportunities and limitations of the LoRa technology for non-contacthuman target detection and localization in wide-area scenarios. Tofurther increase the sensing area coverage, we take a novel approachto leverage the mobility of a drone to carry the transceiver andmove around the target area to perform wireless sensing. As we willdemonstrate later in this paper, we successfully realize, for the firsttime, the building-scale, through-wall sensing to detect and localizehuman targets. We believe the proposed study is particularly usefulfor human target sensing (detection and localization) for applicationsin urban search and rescue missions, such as rescue in natural disas-ters like earthquakes, hurricanes, and tsunamis in the urban areas, aswell as terrorist search and security surveillance [47].

Translating our high-level idea into a practical system, howev-er, is not trivial due to a number of technical challenges. First, thelarger sensing range of LoRa also means the interference range islarger due to the lower receiving sensitivity. Second, incorporating atransceiver pair equipped with a single antenna does not give suffi-cient information regarding the target location since the number ofunknown variables is greater than that of the constrained equationsfor localization. Third, although employing a drone equipped with a

SenSys’19, November 10-13, New York, NY, USA

signal transceiver can increase the sensing coverage, the vibration in-troduced by the drone during its operation (i.e., flying) greatly affectsthe resultant signals and accordingly the target sensing performance.

To address the aforementioned challenges in wide-range sensing,we introduce several novel solutions across the software and hard-ware stacks. To tackle the interference brought by LoRa and devicemobility, we redesign the antenna system and the sensing algorithm.Specifically, we employ a compact, reconfigurable directional an-tenna at the receiver to narrow down the target sensing region. Oursystem can quickly (i.e., within 10𝑚𝑠) switch the radiation patternwith a narrow beamwidth of 48∘. Such a design allows WIDESEE tostay focus on the area of interests to reduce the impact of interfer-ence. To further eliminate the multipath effects within the sensingarea, we take a unique approach to first extract the angle-of-arrival(AoA) related information from available time-series of amplitudesand then use the information to isolate the target path from the inter-fering multipath. As a departure from the commonly used AoA ortime-of-flight (ToF)-based methods, our design avoids the pitfall ofrelying on accurate channel phase information and large bandwidth,which are unfortunately not available on LoRa.

To reduce the ambiguities in localization, we build analytic mod-els that can predict and determine target locations. This is basedon our key observations that the speed of the moving target (e.g.,humans) is relatively constant and the resulting trajectory is smoothwithin a short period of time (e.g., < 1 𝑠). We model the signalcharacteristics of the (vibration) noise and human target movements.We use this information to filter out the vibration artifacts on thereceived signals and to improve the localization precision.

We integrate the proposed techniques to implement a workingprototype and deploy it to detect and localize human targets in threedifferent real-world environments: an open square, an undergroundparking garage, and a high-rise building structure with a size of20×42×85𝑚3. Our experimental results show that WIDESEE caneffectively detect and localize human targets using just a transceiverpair. For 90% of the test cases, the localization error of WIDESEE

is within 4.6 𝑚. Such precision would allow one to identify atwhich room the human target locates in many typical building struc-tures. This is a promising result considering we use only a singletransceiver pair, and the target was moving most of the time in a largeenvironment. We hope our study can encourage further research inexploiting wide-area wireless sensing in detecting and localizinghuman targets for applications like urban search and rescue.

This paper makes the following technical contributions:∙ We present a novel contactless system for sensing human

targets in a wide area using just one transceiver pair, by com-bining the agility of drone with the long-range propagationcharacteristic of LoRa.

∙ We introduce new algorithms and design methodologies acrossthe software and hardware stacks to effectively tackle a seriesof interference issues when applying LoRa and a flying dronefor wide-range sensing, and to address the sensing ambigu-ity issue when only one single transceiver pair is employed.Many of our proposed techniques are generally applicable,and can be applied to other wireless sensing tasks.

∙ We demonstrate, for the first time, building-scale contact-less wireless sensing can be achieved with just one LoRatransceiver pair together with a drone.

Figure 1: Motivation example: a building-scale human targetsensing scenario.

2 BACKGROUND AND OVERVIEW2.1 LoRa TechnologyLoRa offers a long communication distance for up to several kilome-ters [45] with an ability to decode signals as weak as −148 𝑑𝐵𝑚.While the ability for decoding weak signals is beneficial for longdistance communications, it makes LoRa more likely to suffer inter-ference from uninterested area. Even using a directional antenna atthe receiver end, there still exist strong multipath effects [31] withinthe detectable area that greatly affect the sensing accuracy especiallywhen using signal amplitudes. Much effort has been made to similarproblems associated with multipath effects in other types of RF (e.g.,WiFi [48] or RFID [57]) or acoustic signals [43], which are basedon AoA or ToF information of the received signals. However, suchinformation requires accurate channel phase detection and clocksynchronization between the transmitter and the receiver, both ofwhich are unavailable on LoRa.

In this work, instead of making efforts on obtaining AoA or ToF totackle the multipath effects, we consider a novel approach to leveragethe received (albeit susceptible) signal strength (i.e., amplitude) foreffectively addressing the inherent issue of multipath effects. This isdescribed in Section 3.3.

2.2 Motivation and Problem ScopeWIDESEE leverages the long-distance LoRa signal to sense targetsusing a single transceiver pair. As depicted in the conceptual illus-trative of Figure 2, WIDESEE could be used to target emergencyscenarios, such as disaster rescue and terrorist search in (high-rise)building structures. In such scenarios, accurate identification of thepresence of human targets and their locations is of high importance,but doing so is extremely challenging because (a) the localizationsensing infrastructure (e.g., surveillance cameras) may not be read-ily available or has been destructured, and (b) visual inspection ofhuman targets is restricted if not impossible. WIDESEE is designedto offer decision supports in such difficult settings.

In our motivating application of human sensing, we aim to achievethe following two goals. The first is to detect the existence of humantargets. The second is to identify the target’s location if a presence isdetected. We do not consider a multiple-device or multi-hop transmis-sion scheme, because such a strategy requires a careful and complexsetup process, which is often infeasible in emergency situations. Asa proof-of-concept, WIDESEE is designed to capable of detectingthe presence of multiple human targets located in different rooms inthe same building, but only localizing one target at a time. We leavethe simultaneous localization of multiple targets as our future work.

WIDESEE: Towards Wide-Area Contactless Wireless Sensing SenSys’19, November 10-13, New York, NY, USA

Transmitter

Reconfigurable

antenna system

(Receiver)Remote data

processing

platform

Data collection and antenna control system

Figure 2: Overview of WIDESEE. We use a drone to carry theLoRa transceiver pair and its control system. The data are sentback to the remote data processing platform to perform real-time target detection and localization.

2.3 Overview of WIDESEEWIDESEE is a prototyping wide-range contactless human target sens-ing system built upon a single LoRa transceiver pair. The transceiverpair (both the transmitter and receiver) is carried by a drone so thatby flying the drone WIDESEE can scan and sense a large area. Hav-ing a small, lightweight design for WIDESEE is essential to maintaina good endurance for the battery-powered drone.

WIDESEE operates by first transmitting the LoRa signal, and thencapturing and analyzing the resultant signals from the direct signalpath and reflections off the target and surrounding objects to detectthe presence and location of the target. To detect the presence of ahuman target, WIDESEE models how a human activity like breath-ing, waving or ambulating affects the power spectral density (PSD)of the received signal. WIDESEE then tries to locate a detectedhuman target by extracting and analyzing the target’s AoA-relatedinformation.

As depicted in Figure 2, WIDESEE consists of three innovativecomponents:

∙ A compact, reconfigurable antenna system to reduce the in-terference from uninterested areas. To ensure the sensing areaduring a certain time and adapt to the mobility of device, theantenna should be able to adjust its direction and radiationpattern in a dynamic and fast manner. Our design is detailedat Section 3.1.

∙ A data collection and antenna control system, which includesa LoRa transmitter and receiver, a data collection subsystem,and a drone. The drone carries the LoRa transceiver pairand the data collection subsystems to fly around the targetregion. The collected LoRa signal data are sent back to alaptop through a LTE mobile network) to be processed on theground. This is described in Section 3.2.

∙ A target detection and localization system, which runs ona data processing platform, i.e., a laptop in our case. Thesystem analyzes the collected data to detect and localize thehuman target. Based on the calculation, the antenna controlsystem employs an Arduino board carried by the drone toconfigure the antenna radiation pattern accordingly. This isdiscussed in detail in Section 3.3.

Top Bottom

RF line

（connected to device）

Switch

Feed

point

Parasitic

patch

Driven

patch

Figure 3: The fabricated antenna system on the receiver side.

3 SYSTEM DESIGN OF WIDESEEWIDESEE leverages LoRa’s long communication distance and highpenetration capability for sensing targets that are within a wide areaor deep inside building structures. As discussed in Section 2.1, thisadvantage also brings in more interference from uninterested objectsdue to the larger sensing range. Overcoming this limitation requiresnovel design methodologies, analysis and processing algorithms.

3.1 Reconfigurable Antenna SystemTo reduce the interference, we look for innovations at the antenna.Our first intuition is to employ a directional antenna at the receiv-er to narrow down the sensing region. However, commonly usedhorn directional antennas like RFMAX [6] have a fixed radiationpattern and mechanically rotating the antenna orientation to focuson a region (as the device is moved by a drone) is too slow. Fur-thermore, the beamwidth offered by a horn directional antenna isusually not narrow enough for obtaining a high localization accu-racy [6]. An alternative is to use a phased-array antenna that canchange the radiation pattern by adjusting the amplitude and phase ofeach antenna element, to achieve fast scanning with narrower beam-s [20, 26]. However, there is a problem for using a phased-arrayantenna. The LoRa signal has a wavelength of 33 𝑐𝑚 and to achievea 25∘ beamwidth, the linear array will have a size of approximately2𝑚. The resultant antenna design is not only expensive, but also toolarge and too heavy to be fitted on a domestic drone.

We wonder if we could bring together the advantages of horn di-rectional antennas (small size and low cost) and phased arrays (highresolution and scanning speed). In answer, we adopt a reconfigurableantenna approach [13], which is capable of switching the radiationpattern and frequency properties through adjusting its internal cur-rent flow distribution to offer a narrow beamwidth.

Specifically, we choose to use a parasitic-planar-patch anten-na [30] to implement our reconfigurable antenna. Figure 3 shows ourreconfigurable antenna implementation that is used on the receiverside, which consists of a driven patch in the center and two parasiticpatches on both sides. Beam steering is achieved by manipulatingthe status of the parasitic patches to act either as reflectors (whenshorted to ground) or directors (when not shorted to ground). Theradius of each patch is 78 𝑚𝑚. Two shorting pins are shorted toground from each parasitic patch, to ensure that the currents canflow from the parasitic patches to the ground according to the RFswitching configuration. Two SMP1345 PIN diode switches are sol-dered on the parasitic patch layer close to each of the shorting pinsand the RF/direct-current (DC) input. Each diode occupies a smallspace of around 2× 2 𝑚𝑚. The PIN diode is achieved by using aresistance (1.5 Ω) and a capacitor (1.5 𝑝𝐹 ) for ON and OFF states,respectively. The resulting antenna system is small (20×50𝑐𝑚) andhas a comparable weight to a similar-sized horn directional antenna(< 1 𝑘𝑔), but has the advantage of quickly switching the radiation

SenSys’19, November 10-13, New York, NY, USA

0.80 0.85 0.90 0.95 1Freq [GHz]

-14-12-10-8-6-4-20

dB(S

(1,1

))

up4XY Plot 1

0.8923 0.9276

(a)

0.80 0.85 0.90 0.95 1Freq [GHz]

-25

-20

-15

-10

-5

0

dB(S

(1,1

))

left4XY Plot 1

0.8968 0.9261

(b)

0.80 0.85 0.90 0.95 1Freq[GHz]

-25

-20

-15

-10

-5

dB(S

(1,1

))

right4XY Plot 1

0.8964 0.9258

(c)

d 8(Ga1nlota 1) e. 7890e+000 6. 8986e+080 s. 0081tt000 3.1177e+000 1.227Ze+00121

·6. 6S21e-091 ·2. SS37e.e00 ·loi.loiloiloile+000 ·6. 331oi6e+000 ·8. 2250e+000 ·1. 0115e+001 -1. 2'006e+001 -1 . 3896e+001 -1.5787e+001 -1 . 7677e+001 - 1. 9568e+001 -2 . 11oi58e+001

(d)

d 8(Ga1nlota 1) 9.1559e+000 7 . 2082e+080 s. 268Se+000 3. 3128e+000 1. 36S1e+00121

•S. 8258e-091 ·2. 5303e.e00 -loi.loi780e+000 -6. 1oi2S7e+000 -8. 3733e+000 -1 . 0321e+001 -1. 2269e+001 -1 .1oi216e+001 -1. 6161.te +001 -1 . 8112e+001 -2 . 0059e+001 -2 . 2007e+001

(e)

d 8(Ga1nlota 1) 9. 1S'tSe+000

7 . 2092e+080 s. 261f2tt000 3. Sl'Ue+000 1. 37'tle+00121

•S. 7996e-091 ·2. 51610e.e00 -loi.loi61le+000 -6.1oi061e+000

-8. SS12e+000 -1 . 0296e+001 -1. 221oile+001 -1 .1oi186e+001 -1. 6131e+001 -1 . 8076e+001 -2 . 0021e+001 -2 . 1967e+001

(f)

-3.00

90

60

300

-30

-60

-90

-120

-150-180

150

120

up4

Radiation Pattern 2

(g)

-3.00

90

60

300

-30

-60

-90

-120

-150-180

150

120

left4

Radiation Pattern 2

(h)

-3.00

90

60

300

-30

-60

-90

-120

-150-180

150

120

right4Radiation Pattern 2

(i)

Figure 4: Frequency and radiation properties of our reconfig-urable antenna system. Here (a-c), (d-f) and (g-i) respectivelyrepresent the frequency property, radiation pattern and nor-malized radiation pattern of mode1-3.

Table 1: Properties of our antenna system and a similar-sizedRFMAX [6] – a popular horn directional antenna.

Frequency Gain Beamwidth Sizerange (MHz) (dBic) (at 3 dB) (𝑐𝑚3)

WIDESEE (mode 1) 890-930 8.79 45∘

WIDESEE (mode 2) 896-930 9.16 48∘ 20× 50× 1WIDESEE (mode 3) 896-930 9.15 48∘

RFMAX 902-928 9 70∘ 26× 26× 5

patterns. It costs us less than 300 USD to build the antenna and itscontrol system, and we expect the price can be significantly reducedduring massive production.

Figure 4 shows the frequency and radiation properties of our an-tenna system. Our current implementation supports three differentmodes for radiation. We use an Arduino board to switch betweenthe three radiation modes in a round-robin fashion, where switchingoccurs every 10 𝑚𝑠. We empirically determined this switching fre-quency which is sufficient for sensing human targets. This is basedon the observation that the human body movements often have afrequency less than 10𝐻𝑧 [18].

Table 1 compares our antenna system (in three different modes)against a similar-sized RFMAX [6] – a widely used horn direction-al antenna. From the table, we see that the frequency range andgain of our antenna system at all modes are comparable to thoseof RFMAX, but our design has the advantages of offering quickradiation pattern switching and a narrower beamwidth (around 2/3).These advantages make our antenna system more suitable for targetsensing using LoRa. Note that, the total radiation direction range ofour system is twice RFMAX’s, and the radiation pattern switchingis much quicker ( i.e., 10 𝑚𝑠) than horn antenna which requiresmechanical rotation for direction change.

Rx

LimeSDR-mini

+ Antenna

control board

+Smartphone

Antenna

Tx

LoRa

node

(a) The transceiver pair

Arduino

board

Smart

phonePower

bank

LimeSDR

mini

(b) Data collection and control

Figure 5: We use a DJI S1000 to carry the LoRa transceiverpair (a) and the data collection/control subsystem (b).

3.2 Data Collection and Antenna Control SystemAs depicted in Figure 2, we use a consumer drone to carry thetransceiver pair and its control and data collection systems.

3.2.1 Transceiver pair. Our LoRa transceiver pair is shown inFigure 5 (a). We use an off-the-shelf device, SX1276 [7], with anomnidirectional antenna as the LoRa signal transmitter. The trans-mitter sends signals in a continuous mode at 890𝑀𝐻𝑧 frequency– the best working frequency of our reconfigurable antenna system.At the receiver end, we use LimeSDR-mini, a software defined radioboard [4], to collect the LoRa signal at a sampling rate of 250𝐾𝐻𝑧through running the GNU radio software development toolkit [2].We connect the board to our reconfigurable antenna (see Section 3.1)through one of its RF connectors, and to an Android smartphone(with 8𝐺 of RAM and 128𝐺 of storage) via a USB 3.0 port.

The receiver end works as follows. After initializing the LimeSDR-mini board, the antenna control software running on the Arduinoboard continuously switchs among the three radiation modes of ourantenna, at a frequency of every 10 ms. The LimiSDR-mini boardcollects the signal samples at each radiation mode, which are readby the phone to be transferred (labeled with the mode) to a laptopvia LTE connection for data processing. In this way, WIDESEE candetect and localize targets within an interesting area covered by eachradiation mode. Note that our target detection and localization algo-rithms can run on the smartphone or an embedded device to removethe need for data transfer, and we leave this as our future work.

3.2.2 Drone System. We use a DJI S1000 drone [8] to increasethe area a single transceiver pair can effectively cover. As illustratedin Figure 5 (b), the LimeSDR-mini, smartphone, Arduino boardare put on the top of the drone and are powered by a 5200 𝑚𝐴ℎportable power bank with a 2.4 𝐴 output. We employ the collisionavoidance system provided by DIJ to avoid drone collision withobstacles. The drone is controlled by software running on a laptop,programmed through the DJI software development kits. Our futureimplementation will look into reducing the drone’s payload by, e.g.,running the data collection and antenna control software on a singlecomputing device (e.g., the Arduino board), which can be powereddirectly by the drone’s battery.

3.3 Target Detection and Localization SystemWe develop a set of novel algorithms to process the collected LoRasignal data to detect and localize human targets. The process of

WIDESEE: Towards Wide-Area Contactless Wireless Sensing SenSys’19, November 10-13, New York, NY, USA

(a) Without human target. (b) With a moving human target.

Figure 6: Comparison of the PSD of the received signal whenthere exists no human target (a) and a moving target (b). ThePSD patterns in two scenarios differ significantly. WIDESEE ex-ploits this observation to detect the presence of human targets.

detection and localization works as follows. We first pre-process thereceived signal to remove the noises caused by the drone’s vibrationartifacts. We then exploit the PSD (Power spectrum density) of theprocessed signal to detect the presence of human targets. Note thatour detection mechanism can detect the presence of target no matterone or multiple targets are in the sensing area. After detecting amoving target presence, we apply the localization algorithm to esti-mate the location of one target whose reflection is strongest at thattime (note that during localization stage, we let the device hover inpalce). With the device mobility, we can detect and locate multipletargets successively. As we have previously discussed, WIDESEE

needs to effectively handle the multipath effects and location ambi-guities brought in by using one LoRa transceiver pair.

3.3.1 Data pre-processing. Vibrations of a flying drone inevitablyintroduce noise to the received LoRa signals. To remove the intro-duced noise, we exploit the observation that the motion artifactsbrought by a drone are within a frequency range between 60 𝐻𝑧and 150 𝐻𝑧, which is different from the lower frequency rangeof a human body movement (< 10 𝐻𝑧). Therefore, we first use alow-pass filter to remove the high-frequency motion artifacts in thefrequency domain and then convert the filtered signal back to thetime domain to be processed in the next stages. Our evaluation inSection 4 shows that this is a simple yet effective strategy.

3.3.2 Human target detection. Human activities like hand wav-ing and walking will alter the wireless propagation paths and leadto the change of signal amplitude at the receiver [11]. Prior workshows that when no human target presents, the received signal canbe approximated as the superposition of constant signals and thewhite Gaussian noise, yielding an invariant PSD with time [41]. Bycontrast, the PSD of the received signal resulted from a movinghuman target, will lead to fluctuations on the measured signal. Asan example, consider Figure 6 drawn from our own experiments.It illustrates the difference in the PSD with and without a movinghuman target. When no human target presents (Figure 6 (a)), thePSD of the received signal remains stable with time and close to0 𝐻𝑧, while when a moving target presents (Figure 6 (b)), the PSDfluctuates at low frequencies (0− 10 𝐻𝑧). Our work exploits thissignal characteristic to detect the presence of a human target – if themeasured PSD frequency and its variance are both below a threshold

0 2.5 5Frequency (Hz)

0

0.5

1

Nor

mal

ized

PS

D

(a) Ambulating.

0 1 2 3 4Frequency (Hz)

0

0.5

1

Nor

mal

ized

PS

D

(b) Waving.

0 0.5 1 1.5 2Frequency (Hz)

0

0.5

1

Nor

mal

ized

PS

D

(c) Stationary.

Figure 7: Frequency distribution of signal changes caused bythree states of a human target. Different states lead to differentsignal frequency distributions.

(empirically set to 0.1 𝐻𝑧 in our case), we consider there is no hu-man target; otherwise, we conclude that someone(with movements)is in the sensing area.

In this work, we focus on detecting human targets with largemovements: ambulating or waving in-place. We are also able todetect a stationary breathing target when there is no obstacle be-tween the transceiver pair and the target, or the obstacle is thin (seeSection 4.2.3). Figure 7 illustrates the normalized PSD of the reflect-ed signals of these three stats from a human target in a controlledenvironment. The diagram shows that different states exhibit differ-ent characteristics in the frequency domain, which can be used toidentify and differentiate these states. In particular, human breathand waving present strong, dense PSD with a frequency range of0.1-0.6𝐻𝑧 and 1-4𝐻𝑧, respectively. By contrast, the PSD distribu-tion of an ambulating human target is more spread apart, mainly dueto the human target’s randomized ambulatory trajectory pattern anduncorrelated movements of multiple body parts.

3.3.3 Ambulating Target Localization. Once an ambulating hu-man target is detected, we focus on identifying the location of thetarget. One of the biggest technical challenges in LoRa-based targetlocalization is that the multipath in LoRa is more severe than thatin other signals (e.g., Wi-Fi) as shown in Figure 8. Even thoughwe utilize the narrower beam antenna at the receiver, the multipathwithin the sensing area can be substantially strong, which negativelyimpacts target localization. To address the multipath issue, previ-ous work has investigated various techniques, such as analyzing theAoA information [52], frequency hopping based on accurate channelphase measurement [35, 40], and comparing ToF that requires largebandwidth and tight transceiver synchronization [34]. Unfortunately,these techniques are not applicable to our system because the themaximum bandwidth of LoRa is only 500𝐾𝐻𝑧, and the asynchro-nism between LoRa node (Tx) and gateway (Rx) makes it impossibleto extract stable phase readings from the received signal. Also, it isparticularlly difficult to achieve the synchronization between LoRanode and gateway due to the cheap oscillator adopted. Because ofthe chirp modulation, LoRa can tolerate high frequency offset forcommutation so high-accuracy oscillator is not needed.

In this work, we propose an amplitude-based anti-multipath methodto localize a moving target. The foundation of our method is to ex-tract AOA-related information from amplitude, inspired by a recentwork by Karanam et al. [32]. By using the AoA-related informationfor localization, we have the opportunity to remove the multipath

SenSys’19, November 10-13, New York, NY, USA

2 3 4 5 6 7 8 9Distance between Tx and Rx (m)

-28

-27

-26

-25

Am

plit

ud

e (d

Bm

)

(a) Wi-Fi.

2 3 4 5 6 7 8Distance between Tx and Rx (m)

-118

-116

-114

Am

plit

ud

e (d

Bm

)(b) LoRa.

Figure 8: Multipath severity comparison between Wi-Fi and Lo-Ra in the same outdoor environment. The figure show variationof signal amplitude with distance between Tx and Rx. For Wi-Fi, the signal amplitude is inversely proportional to the squareof distance, which matches well with the theoretical no multi-path case [5]. In contrast, the change trend of LoRa seems fluc-tuating, this is the result of the superposition of multiple paths.So we can deduce that the multipath in LoRa is much sever thanin Wi-Fi.

effect that however is not addressed in [32]. In the following sub-sections, we first describe the basic concepts behind extracting theAoA-related information when a target is moving. Then, we answerhow to obtain target location information from the extracted AoAinformation. Finally, we propose the solution to handle multipath.

AoA-related parameter estimation Based on Amplitudes. Herewe first describe the framework for estimating AoA-related parame-ter in device moving scenario (target fixed), which can be used todeduce the AoA-related parameter in target moving case (devicefixed).

Considering a scenario in which we have one receiver that canreceive signals from 𝐾 different sources. When the receiver movesalong a straight line at speed 𝑣 for time 𝑡, it can emulate a linear array.We can define the directional angles of each signal sources in far-fieldas 𝜃𝑘, 𝑘 = 1, 2, ...,𝐾. Then, the signal received from the 𝑘th sourceat time 𝑡 can be expressed as 𝑥𝑡,𝑘 = 𝑎𝑡,𝑘𝑒

𝑗(𝜇𝑘+2𝜋𝑣𝑡

𝜆𝑐𝑜𝑠𝜃𝑘) [22],

where 𝑎𝑡,𝑘 is the amplitude of the signal and 𝜇𝑘 is the signal phaseat the initial time (i.e., 𝑡 = 0). Then, the signal received at thereceiver at time 𝑡 is a superposition of the 𝐾 paths, which can bewritten as

𝑦(𝑡) =

𝐾∑𝑘=1

𝑎𝑡,𝑘𝑒𝑗(𝜇𝑘+

2𝜋𝑣𝑡𝜆

𝑐𝑜𝑠𝜃𝑘). (1)

Previous researches have proved that the autocorrelation of thereceived signal amplitude arrays at a delay 𝑅(𝜏) contains a to-tal of 𝐾(𝐾−1)

2harmonics. Each harmonic’s frequency is related

to the cosine of two available sources’ AoA (i.e., 𝜃𝑘, 𝜃𝑗 , 𝑗 =1, 2, ...,𝐾) [22, 32], which is given by𝑓𝑘,𝑗 =

𝑣

𝜆|𝑐𝑜𝑠𝜃𝑘 − 𝑐𝑜𝑠𝜃𝑗 |. (2)

These frequencies can be obtained by a frequency estimation tech-nique, such as the fast Fourier transformation of 𝑅(𝜏) followed by apeak detection of magnitudes.

For the target moving and transceiver fixed scenario as shownin Figure 9 (a). The resultant signal composes of reflection (Tx →

0

T

R

vd

LTx Rx

Ambulating

target

LoRa

(a)

Static paths

Tx Rx

Ambulating

target

Wall

LoRa

(b)

Figure 9: (a) shows the setup of moving target localization. Sim-ilar to the linear virtual array constructed by moving receiver.The moving target also can emulate a linear array. (b) presentsthe superposition signal of multiple paths. For target path es-timation, we need to remove the dynamic multipath caused bythe indirect target reflection.

0 0.5 1 1.5 2 2.5 3f

0

0.2

0.4

0.6

0.8

1

PS

D

t=[0, T)t=[W, W+T)t=[2*W, 2*W+T)t=[3*W, 3*W+T)

|v(cosT

+cosR

)|

estimates[1.38, 1.36, 1.32, 1.26]

Figure 10: Four adjacent estimates of AoA related parameter|𝑣(𝑐𝑜𝑠𝜃𝑇 + 𝑐𝑜𝑠𝜃𝑅)| obtained in PSD plots. The estimates arethe normalized frequencies 𝑓𝜆 corresponding to the peaks ofPSD plots.

target → Rx) from an ambulating human target and direct path (Tx→ Rx) that is received by the Rx at time 𝑡 can be written as:

𝑦(𝑡) = 𝑎𝑠𝑒𝑗𝜇𝑠 + 𝑎𝑑𝑒

𝑗(𝜇𝑑+2𝜋𝑣𝑡

𝜆(𝑐𝑜𝑠𝜃𝑇+𝑐𝑜𝑠𝜃𝑅)), (3)

where 𝑎𝑠 and 𝜇𝑠 are the amplitude and phase of the signal from Tx,𝑎𝑑 is the amplitude of the signal reflected from the moving target, 𝜇𝑑

is the initial phase (𝑡 = 0) of the signal from moving target, 𝑣 is themoving speed of target, 𝜃𝑇 and 𝜃𝑅 are the angle of departure (AoD)and AoA of the target reflected path, respectively.

We find that Equation (3) can be taken as a special case ofEquation (1) with 𝐾 = 2 sources, 𝑐𝑜𝑠𝜃1 = 0 and 𝑐𝑜𝑠𝜃2 =𝑐𝑜𝑠𝜃𝑇 + 𝑐𝑜𝑠𝜃𝑅. The moving target can synthesize a transmitter ar-ray. Since 𝑣, 𝜃𝑇 and 𝜃𝑅 are unknown in practice, we jointly estimatethe AoA-related parameter |𝑣(𝑐𝑜𝑠𝜃𝑇 + 𝑐𝑜𝑠𝜃𝑅)| as 𝑓𝜆 according toEquation (2).

Localization ambiguity avoidance. By utilizing the |𝑣(𝑐𝑜𝑠𝜃𝑇 +𝑐𝑜𝑠𝜃𝑅)| estimate from only one transceiver pair to localize the tar-get, there exist severe localization ambiguities that stem from threeaspects:(1) the absolute value symbol | | of 𝑣(𝑐𝑜𝑠𝜃𝑇 + 𝑐𝑜𝑠𝜃𝑅), (2)unknown distance 𝑑 between target and LoS, (3) unknown speed 𝑣and direction 𝜃0 of target movement. Here we show this problemwith a special case (𝜃0 = 0)in Figure 12 (a). We can see that even in

WIDESEE: Towards Wide-Area Contactless Wireless Sensing SenSys’19, November 10-13, New York, NY, USA

(a) Target moves towards LoS. (b) Target moves away from LoS.

Figure 11: The changing trend of 𝑣(𝑐𝑜𝑠𝜃𝑇 + 𝑐𝑜𝑠𝜃𝑅) estimatesat various target moving directions 𝜃0. It can be seen that thechanging trend of 𝑣(𝑐𝑜𝑠𝜃𝑇 + 𝑐𝑜𝑠𝜃𝑅) estimates from the 1st to𝑚th is always decreasing no matter the target moves towardsLoS (0∘ < 𝜃0 < 90∘) or moves away from LoS (90∘ < 𝜃0 <180∘).

such constraint, there still exist ambiguous location points (the areaswith red color). So it is difficult to obtain the target’s true initiallocalization.

In this paper, we solve this problem based on the facts that thetarget’s moving trajectory is smooth and the speed is constant dur-ing a short period of time (e.g., < 1 𝑠). Specifically, we utilizemultiple adjacent estimates of 𝑣(𝑐𝑜𝑠𝜃𝑇 + 𝑐𝑜𝑠𝜃𝑅) with a slidingwindow size of 𝑊 , and each estimation process requires 𝑇 timelength of samples. 𝑊 and 𝑇 are empirically set to 0.25 𝑠 and1 𝑠 in our system. We aim to solve seven unknown parameters:[𝜃𝑇 (1), 𝜃𝑅(1), 𝜃0, 𝛼, 𝛽, 𝑑(1), 𝑣], where 𝜃𝑇 (1) , 𝜃𝑅(1) and 𝑑(1) arethe initial values of 𝜃𝑇 , 𝜃𝑅 and 𝑑 respectively. 𝛼 is the difference be-tween two adjacent estimates of 𝜃𝑇 , and 𝛽 is the difference betweentwo adjacent estimates of 𝜃𝑅. For the 𝑚th estimation, we can havetwo equations as bellow:

𝑣(𝑐𝑜𝑠𝜃𝑇𝑚 + 𝑐𝑜𝑠𝜃𝑅𝑚) = ±𝑓𝑚𝜆𝑑(𝑚)(𝑡𝑎𝑛(𝜃0 + 𝜃𝑅𝑚) + 𝑡𝑎𝑛(𝜃𝑇𝑚 − 𝜃0)) = 𝐿

. (4)

Note that ± can be removed due to our observation as shown inFigure 11. We find that the changing trend of 𝑣(𝑐𝑜𝑠𝜃𝑇 + 𝑐𝑜𝑠𝜃𝑅)estimates from the 1st to 𝑚th is always decreasing. When the targetmoves towards LoS, the values of cos 𝜃𝑇 and cos 𝜃𝑅 (0∘ < 𝜃𝑇 <90∘, 0∘ < 𝜃𝑅 < 90∘) are positive and the values decrease withangles 𝜃𝑇 and 𝜃𝑅 increasing. When the target moves away fromthe LoS, the values of cos 𝜃𝑇 and cos 𝜃𝑅 (90∘ < 𝜃𝑇 < 180∘,90∘ < 𝜃𝑅 < 180∘) are negative and the values again decrease withangles increasing. So we delete another set of estimates that do notsatisfy the condition of decreasing from the 1st to 𝑚th.

Consider the fact of trajectory smoothness and speed constancy,we can add following constraint conditions:⎧⎨⎩

𝜃𝑇 (𝑚)− 𝜃𝑇 (𝑚− 1) = 𝜃𝑇 (𝑚+ 1)− 𝜃𝑇 (𝑚) ≈ 𝛼𝜃𝑅(𝑚)− 𝜃𝑅(𝑚− 1) = 𝜃𝑅(𝑚+ 1)− 𝜃𝑅(𝑚) ≈ 𝛽𝑑(𝑚) ≈ 𝑑(1)− 𝑣𝑇 (𝑚− 1)𝑐𝑜𝑠𝜃0

.

(5)We find that when 𝑚 reaches to 4, the number of equations (=8)is larger than the number of unknowns (=7). Thus we are able tosolve all of the unknowns only by using 4 times estimation resultsof |𝑣(𝑐𝑜𝑠𝜃𝑇 + 𝑐𝑜𝑠𝜃𝑅)| as shown in Figure 10.

0 5 10X (m)

0

2

4

6

8

10

Y (

m)

0

0.2

0.4

0.6

0.8

1

Ground truth

(a) Before ambiguity avoidance.

0 5 10X (m)

0

2

4

6

8

10

Y (

m)

0

0.2

0.4

0.6

0.8

1

Ground truth

(b) After ambiguity avoidance.

Figure 12: Localization ambiguity avoidance. The heatmap ex-hibiting the likelihood that how a partitioned grid is likely to bethe initial location of the target. The ground truth is highlightedwith black square. From (a), we can see all grids of the reddestcolor result in the same value of |𝑣(𝑐𝑜𝑠𝜃𝑇 + 𝑐𝑜𝑠𝜃𝑅)|, resultingsevere localization ambiguities.

Since the equations are non-linear that can not be solved directly,an intuitive choice is using Approximate Search algorithm. To avoidthe local optimum issue and reduce time overhead, we adopt particleswarm + Global Search in Global Optimization Toolbox to achieveglobal search for the system of non-linear equations. The principleis to obtain a search initial value close to the solution by usingparticle swarm, then limit the objective function to fmincon’s non-linear constraints, and finally use Globa Search to find the targetsolution. Figure 12 (b) shows that the initial localization result afterour ambiguity avoidance scheme is very close to the ground truth.

Multipath interference elimination. Consider a typical multipathscenario shown in Figure 9 (b), we can see that the received signalat the receiver is a superposition of multiple signals, which can bewritten as:

𝑦(𝑡) = 𝑎𝑠𝑒𝑗𝜇𝑠 + 𝑎𝑑𝑒

𝑗(𝜇𝑑+2𝜋𝑣𝑡

𝜆(𝑐𝑜𝑠𝜃𝑇+𝑐𝑜𝑠𝜃𝑅)

+ 𝑎𝑑,2𝑒𝑗(𝜇𝑑,2+

2𝜋𝑣𝑡𝜆

(𝑐𝑜𝑠𝜃𝑇,2+𝑐𝑜𝑠𝜃𝑅,2),(6)

where 𝑎𝑑, 𝜇𝑑, 𝜃𝑇 and 𝜃𝑅 are the amplitude, initial phase (t=0),AoA and AoD of the direct target reflection (Tx→target→Rx), re-spectively. To simplify representation, here we approximate theindirect target reflection (Tx→target→wall→Rx) to a new directreflection from a virtual target (Tx→virtual target→Rx), with ampli-tude 𝑎𝑑,2, initial phase (t=0) 𝜇𝑑,2, AoD 𝜃𝑇,2 and AoA 𝜃𝑅,2. In suchscenario, the 𝑓𝜆 estimates may represent |(𝑣(𝑐𝑜𝑠𝜃𝑇 + 𝑐𝑜𝑠𝜃𝑅))|as well as |𝑣(𝑐𝑜𝑠𝜃𝑇,2 + 𝑐𝑜𝑠𝜃𝑅,2)| and |(𝑣(𝑐𝑜𝑠𝜃𝑇 + 𝑐𝑜𝑠𝜃𝑅)) −(𝑣(𝑐𝑜𝑠𝜃𝑇,2 + 𝑐𝑜𝑠𝜃𝑅,2))|. Since the static resultant component (con-tains direct path) is much stronger than the reflections, we can lockthe two dominant 𝑓𝜆 estimates into |(𝑣(𝑐𝑜𝑠𝜃𝑇 + 𝑐𝑜𝑠𝜃𝑅))| and|𝑣(𝑐𝑜𝑠𝜃𝑇,2 + 𝑐𝑜𝑠𝜃𝑅,2)|.

To distinguish the target corresponding |𝑣(𝑐𝑜𝑠𝜃𝑇 +𝑐𝑜𝑠𝜃𝑅)| valuefrom the |𝑣(𝑐𝑜𝑠𝜃𝑇,2 + 𝑐𝑜𝑠𝜃𝑅,2)| estimate shown in Figure 13 (b-c),past solutions (e.g., Dynamic-Music) exploit the fact that the directtarget reflected path is stronger than the indirect target reflectedpath [52], so they consider the 𝑓𝜆 value of larger magnitude as thedirect target reflection resulted |𝑣(𝑐𝑜𝑠𝜃𝑇 + 𝑐𝑜𝑠𝜃𝑅)|.

This approach can be effective when the interfering object (likeat location A in Figure 13 (a)) is far away from the connecting

SenSys’19, November 10-13, New York, NY, USA

0 2 4 6 8 10X (m)

-2

0

2

4

6

Y (m

)

RxTx

Target

B

A

(a) Setup.

0 1 2 3f

0

0.2

0.4

0.6

0.8

1

PS

D

t=[0, T)t=[W, W+T)t=[2*W, 2*W+T)t=[3*W, 3*W+T)

(b) PSD plots with interfering object at lo-cation A.

0 1 2 3f

0

0.2

0.4

0.6

0.8

1

PS

D

t=[0, T)t=[W, W+T)t=[2*W, 2*W+T)t=[3*W, 3*W+T)

(c) PSD plots with interfering object at lo-cation B.

1 2 3 4Index

0.5

1

1.5

|v(c

osT+c

osR

)| es

timat

es

TargetTarget + A (Dynamic-Music)Target + B (Dynamic-Music)Target + B (Our method)

(d) |(𝑣(𝑐𝑜𝑠𝜃𝑇 + 𝑐𝑜𝑠𝜃𝑅))| estimates.

Figure 13: Multipath interference presentation. The figure shows that past AoA based multipath removal solution fails when theinterfering object is close to the connecting line between target and receiver/transmtter (like at location B). Our approach has theability to mitigate the problem.

6 8 10X (m)

0

2

4

6

8

10

Y (

m)

0

0.2

0.4

0.6

0.8

1

Ground truth

(a) Using Dynamic-Music.

6 8 10X (m)

0

2

4

6

8

10

Y (

m)

0

0.2

0.4

0.6

0.8

1

Ground truth

(b) Using our approach.

Figure 14: Localization errors of using past (5.1 𝑚) andour (1 𝑚) multipath elimination approaches, when an interfer-ing object locates at point B as shown in Figure 13 (a).

line between target and receiver/transmtter, we can see that the|𝑣(𝑐𝑜𝑠𝜃𝑇 + 𝑐𝑜𝑠𝜃𝑅)| estimates decline monotonically and matchwell with those of the scene without interfering object as shownin Figure 13 (c-d). However, when the interfering object (like atlocation B in Figure 13 (a)) is close to the connecting line betweentarget and receiver/transmtter, the |𝑣(𝑐𝑜𝑠𝜃𝑇 + 𝑐𝑜𝑠𝜃𝑅)| estimatesarise remarkable shifts as shown in Figure 13 (c-d), which causesour localization method to fail. To mitigate this problem, we take themedian of two 𝑓𝜆 values as the |𝑣(𝑐𝑜𝑠𝜃𝑇 + 𝑐𝑜𝑠𝜃𝑅)| estimate oncethe four estimates do not decline monotonously. Figure 13 (d) showsthat the estimation results by our method match better with those ofthe scene without interfering object, compared to the results obtainedby the past solution. And the localization error of our method (1 𝑚)is much lower than that of past method (5.1𝑚).

4 EVALUATION4.1 Experimental Setup and RoadmapWe performed two sets of field experiments to evaluate WIDESEE indetecting and localizing human targets: field experiments without adrone and field experiments with drone.

4.1.1 Field experiments without a drone. We would like toprovide a quantitative evaluation first to justify our design choices,and to identify the research opportunities and limitations of LoRasensing. To this end, we evaluate how the line-of-sight (LoS) lengths

Evaluation scenarios

Tx

Rx

LoS

Test area

Setup

ca

Target

b

Figure 15: Evaluation scenarios and setup. We evaluateWIDESEE in an open square (a), an underground parkinggarage (b), and a middle floor of a 17-floor building (c). Thetesting area is on one side of the Line-of-Sight (LoS).

affect the sensing distance of LoRa in Section 4.2.1. In Section 4.2.2,we report the performance of our antenna system for detecting mov-ing human targets and comparing it to two alternatives designs usingomni- and horn-directional antennas. We then evaluate LoRa’s pen-etration capability in detecting three different human activities inSection 4.2.3, before reporting the localization accuracy of our al-gorithms in Section 4.2.4. Finally, in Section 4.2.5, we evaluatethe impact of the human target’s walking speed on detection andlocalization accuracy. Controlled experiments were performed onthe ground for detecting a single moving human target. Hence, weplaced the LoRa transceiver pair 1 𝑚 above the ground. Later in thefield study, we use the drone to carry the transceiver pair.

4.1.2 File study with a drone. In the field study, we use a droneto carry the LoRa transceiver pair to detect and locate a human targetin the building shown in Figure 15 (c). We report the performancefor detecting the presence of human targets and the accuracy forlocalizing a human target with different drone speeds. The resultsare given in Section 4.3.

4.1.3 Evaluation metric. We calculate the accuracy for detectingthe presence of a human target as:

𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦 =1

𝐶

𝐶∑𝑐=1

(1− |𝐻𝑚,𝑐 −𝐻𝑡,𝑐

𝐻𝑡,𝑐|), 𝐻𝑚,𝑐, 𝐻𝑡,𝑐 = 0, 1

WIDESEE: Towards Wide-Area Contactless Wireless Sensing SenSys’19, November 10-13, New York, NY, USA

1 5 9 13 17 21 25 29 33 37 41 45 49 53 5759

Length of LoS (m)

0

10

20

30

40

50

60D

etec

tab

le d

ista

nce

(m

)

Figure 16: Impact of the LoS length on the detectable distance.WIDESEE can detect a moving object with a distance to thetransceiver pair of up to 53𝑚 in an open space.

where 𝐶 is the number of tests, 𝐻𝑚,𝑐 and 𝐻𝑡,𝑐 are the outputs ofWIDESEE and the ground-truth in the 𝑐th test respectively.

4.2 Field experiments without a drone4.2.1 Detection distance under different LoS lengths. In thisexperiment, we varied the LoS length, i.e., the distance between thetransmitter and the receiver, from 1 𝑚 to 59 𝑚 at a step size of 2 𝑚.The tests were conducted in the open square shown in Figure 15 (a).In each transceiver pair setting, we ask a target to walk along thevertical bisector of the transceiver pair 100 times with a walkingdistance of 3𝑚 each time, starting from a randomly chosen position.Note that we moved the starting point further from the transceiverpair each time, until we fail to detect the user from that point. Weconsider a position to be detectable if we can correctly identify theuser at that position for over 90% of the times. We calculate thedistance between each detectable position and the middle point ofthe transceiver pair to find the largest-possible sensing distance for agiven setting.

Figure 16 shows how the LoS setup affects the sensing distanceof WIDESEE. We see that the sensing distance in general grows asthe length of LoS increases. However, it reaches a plateau with a de-tectable distance of 53 𝑚. This suggests that WIDESEE can achievea sensing distance of 53 𝑚 for a moving target in a relatively idealscenario (an open space). Such a sensing distance is a significantimprovement over WiFi, RFID or mmWave-based systems, whichhave a sensing distance under 6 𝑚 [46, 51, 55].

4.2.2 Evaluation of our antenna system. This experiment isdesigned to evaluate the performance of our antenna system in hu-man target detection. We compare our design against two alternativedesigns which use an omnidirectional [1] and a horn directional(RFMAX [6]) antenna with a similar size. Our testing area is anopen space with a size of 42 × 48 𝑚2 depicted in Figure 15 (a).We divide the testing area into a grid of 224 blocks, 3 × 3 𝑚2 foreach block. Like the previous experiment, we ask a human targetto choose any block and then move within the block naturally. Weensure that each block was tested at least once.

Figure 17 shows that our design presents the best trade-off be-tween the detection accuracy and the area coverage. In this experi-ment, we report the number of detectable blocks. Note that this isdifferent from the evaluation in Section 4.2.1, where we are inter-ested in the longest-possible distance for detecting a target that isalways on the perpendicular bisector of the transceiver pair. In thisexperiment, most of the blocks are not on the perpendicular bisector

18 24 30 36 42 48Length of LoS (m)

0

20

40

60

80

100

Det

ecti

on

acc

ura

cy (

%) Omnidirectional antenna

Horn directional antennaOur antenna

(a) Detection accuracy.

18 24 30 36 42 48Length of LoS (m)

0

20

40

60

80

100

Det

ecta

ble

blo

cks

(%) Omnidirectional antenna

Horn directional antennaOur antenna

(b) Detectable area.

Figure 17: Comparing the detection accuracy (a) and region(b) between our approach and alternative directional antennaswith a similar size. Our antenna design gives the best trade-offbetween the detection accuracy and range.

of the transceiver pair. As we increase the LoS length beyond 36 𝑚,we see a decrease in the number of detectable blocks. This is mainlydue to the directionality of receiver antenna.

While RFMAX, the horn directional antenna achieves the second-best detection accuracy, it can detect the least number of blocks. Theomnidirectional antenna, on the other hand, can cover more blocks,but it achieves the poorest detection accuracy due to its sensitivityto the surrounding interference. The number of detectable blocksachieved by our antenna is not far from the omnidirectional antenna,but it delivers a much higher detection accuracy for all LoS settings.The better detection accuracy of our approach is largely attributed toits narrower beam, which in turn leads to a better ability for avoidingthe inference from non-targeting objects.

4.2.3 Penetration test. We also evaluate WIDESEE’s ability topenetrate the walls. Experiments were conducted in the undergroundgarage (Figure 15 (b)) and the second floor of our test building(Figure 15 (c)). Our evaluation includes four settings – no wall andwall made of reinforced concrete with three different thicknesses(26 𝑐𝑚, 52 𝑐𝑚, and 94 𝑐𝑚) between the target and the transceiverpair. In the through-wall experiment, we placed the transceiver pair1.5𝑚 away from the wall.

Figure 18 shows the results. As expected, the thinner the obstaclebetween the target and the transceiver pair, the deeper WIDESEE cansuccessfully detect the target. We observe the target’s activity alsohas a significant impact on the detectable distance. If the target isambulating or waving, WIDESEE can successfully detect the targetup to 15 𝑚 and 13 𝑚, respectively. WIDESEE can also detect astationary target with just respiration. However, the detection dis-tance is limited and depends on the thickness of the wall. This is notsurprising, as the smaller the activity and the thicker the wall is, theweaker the received signal strength will be. For our experiments, a20 𝑐𝑚 increase in the wall’s thickness would reduce the received sig-nal strength by around 29 𝑑𝐵𝑚. Nonetheless, the results show thatWIDESEE can accurately detect a human target moving or wavingdeep inside the building.

4.2.4 Localization accuracy in different multipath environ-ments. We report how multipath impacts localization accuracy inthis section. Our evaluation environments are the open square (Fig-ure 15 (a)) and basement parking garage (Figure 15 (b)). As can beseen from Figure 15 (b), the basement is supported by many pillars

SenSys’19, November 10-13, New York, NY, USA

No wall

1 1.5 2 2.5 3 3.5Length of LoS (m)

0

5

10

15

20

25

Det

ecta

ble

dis

tan

ce (

m)

Ambulating Waving Stationary

Wall thickness = 26cm

1 1.5 2 2.5 3 3.50

5

10

15

20

25Wall thickness = 52cm

1 1.5 2 2.5 3 3.50

5

10

15

20Wall thickness = 94cm

1 1.5 2 2.5 3 3.50

2.5

5

7.5

10

Figure 18: Penetration test. WIDESEE can detect a stationary human target who is behind a wall with a thickness of 52cm, andWIDESEE can detect a moving/walking target deeper inside a wall.

0 1 2 3 4 5 6Localization error (m)

0

0.2

0.4

0.6

0.8

1

CD

F

Open spaceBasement

2.1 2.7

Figure 19: CDF plots of the localization error in two large sce-narios (10× 25𝑚2) with different densities of multipath.

and hence has rich multipath. The test areas in both environmentsare of the same size (10 × 25 𝑚2), and we set the LoS length to10𝑚 to match the shorter side of the test area (see Figure 15 (right)).We divide the testing area into 125 blocks, where each block has asize of 2 × 1 𝑚2. For each block, a target was asked to walk on apredefined straight line that can have a 0∘, 30∘, 45∘, 60∘ or 90∘ de-grees towards the LoS. For each line, the user walked for around 2 𝑠,starting from the center of a block. Figure 19 plots the cumulativedistribution function (CDF) of the localization error across our 125experimental trials. This diagram shows that for over 50% of our testcases, the localization error is within 2.1 𝑚 and 2.7 𝑚 in the openspace and the basement, respectively. Such precision would be goodenough for locating a human target in many application scenarios,demonstrating the great potential of wireless sensing using a singleLoRa transceiver pair.

We conduct another experiment to show the track accuracy ofour system in a smaller area. This experiment was performed ina smaller room located in the building shown in Figure 15. Theroom has a size of 8 × 10 𝑚2. In this experiment, the LoS lengthis set as 6 𝑚. The target was asked to walk along the trajectory offive letters “BCIMO”. The recovered trajectories (dots) and the truewalking trajectories (solid lines) are shown together in Figure 21.WIDESEE achieves a median localization error of 52 𝑐𝑚, which iscomparable to the 32 𝑐𝑚 error achieved in IndoTrack - a state-of-art WiFi tracking system with dense deployment. This experimentshows that WIDESEE is able to track the target at higher accuracy ina small-sized environment.

0.5 0.2 1 0.2 2 0.2Target moving speed (m/s)

90

92

94

96

98

100

Det

ecti

on

acc

ura

cy (

%)

(a) Effect on target detection.

0.5 0.2 1 0.2 2 0.2Target moving speed (m/s)

1

1.5

2

2.5

3

Lo

caliz

atio

n e

rro

r (m

)

(b) Effect on target localization.

Figure 20: Impact of target moving speed. The target’s movingspeed has little impact on detection, but it affects localization.

4.2.5 Impact of the target’s moving speed. This experimentstudies the effect of the target’s walking speed on detection and lo-calization. We consider three walking speeds: slow (0.5± 0.2𝑚/𝑠),average (1± 0.2 𝑚/𝑠) and fast (2± 0.2 𝑚/𝑠). We conducted theexperiments in the open square shown in Figure 15 (a).

Figure 20 shows that the target’s moving speed has little impacton target detection, but it does affect the localization precision. Weobserve a localization error of 1.7 𝑚, 2 𝑚, and 2.5 𝑚 when thetarget was moving at slow, average, and fast speeds respectively.This is largely attributed to the body motions (e.g., arm swing) –the faster the walking speed is, the drastic the body movement willbe – a more drastic body movement makes it harder to satisfy theconditions that we use for localization (see Equation 5).

4.3 Building-scale Field StudyIn this field study, we employ WIDESEE with drone (see also Fig-ure 2) to perform building-scale sensing. The task is to detect andtrack a human target located on the 9th floor of a 17-floor buildingstructure with a size of 20× 42× 85𝑚3 (Figure 15 (c)). Note thatthis new building had no occupant at the time of our experiment.The thicknesses of the concrete walls and glass windows are 40 𝑐𝑚and 5 𝑐𝑚, respectively. The transceiver pair was carried by a dronein this experiment. The distance (i.e., the LoS length) between thetransmitter and the receiver is 2𝑚. Ten student volunteers participat-ed in this study, serving as the target one at a time. Figure 22 showsthe experimental setup. The students were arranged into three groupsto perform stationary (breathing) (2 students), waving (4 students)

WIDESEE: Towards Wide-Area Contactless Wireless Sensing SenSys’19, November 10-13, New York, NY, USA

0 2 4 6X (m)

0

2

4

6

Y (

m)

0 2 4 6X (m)

0

2

4

6

Y (

m)

0 2 4 6X (m)

0

2

4

6

Y (

m)

0 2 4 6X (m)

0

2

4

6

Y (

m)

0 2 4 6X (m)

0

2

4

6

Y (

m)

Figure 21: Tracking results in a smaller room with size of 6 × 8 𝑚2. The median tracking error of WIDESEE is 52 𝑐𝑚, which iscomparable with a state-of-the-art that utilizes two WiFi transceiver pair, but WIDESEE offers a much longer sensing distance witha single LoRa transceiver pair.

Figure 22: Building-scale experimental set-up.

1 0.2 1.5 0.2 2 0.2 2.5 0.2 3 0.2Flying speed of drone (m/s)

60

70

80

90

100

Det

ecti

on

acc

ura

cy (

%)

WalkingWaving

Figure 23: Target detection accuracy inbuilding-scale field study, with variousdrone’s flying speeds.

0 2 4 6 8 10 12Localization error (m)

0

0.2

0.4

0.6

0.8

1

CD

F

Our methodDynamic-Music

84.6

Figure 24: CDF plot of the localizationerror given by WIDESEE and Dynamic-MUSIC [52], the closest prior work.

and walking (4 students) activities, and located in rooms on the samefloor. We manually control the drone to fly to an initial position ofthe 9th floor, and then let our data processing platform (Section 2.3)to control the drone’s fly. We varied the flying speed of the drone inexperiments.

4.3.1 Detecting human presences. Figure 23 shows the detec-tion accuracy for each human target who was walking and waving.When the drone was flying at a low speed of 1±0.2𝑚/𝑠, WIDESEE

can successfully detect 98% and 96% of the human targets who werewalking and waving respectively. As expected, the detection accu-racy decreases as the drone’s speed increases, but WIDESEE is stillable to detect the target most of the time when the target was walkingor waving. WIDESEE is unable to detect the stationary (breathing)target in this study when the device is on a drone and the target ispretty far away (> 5 𝑚) from the device with a 40 𝑐𝑚 wall in be-tween. How to improve the detection accuracy for stationary humantargets with respiration sensing is one important direction of ourfuture work. However, our current implementation would alreadybe useful for disaster rescue to detect conscious survivors, many ofwhom tend to wave to attract the attention of rescuers.

4.3.2 Localization accuracy. Once we have detected a movingtarget, we hover the drone for 2 seconds to collect the target move-ment information, and apply the localization algorithms describedin Section 3.3.3 to estimate the target position. Figure 24 comparesthe localization error of our approach against the the state-of-art –Dynamic-Music [52] for localization. As can be seen from the figure,

our approach delivers a better localization accuracy over Dynamic-MUSIC. It reduces the localization error from 8𝑚 to 4.6𝑚 for over90% of the test cases. This is a significant improvement because inpractice such a precision improvement would allow us to identifywhich rooms of a building a human target is located. This is particu-larly useful in disaster rescue where we might want to prioritize therescue of survivors.

5 DISCUSSIONSNaturally, there is room for improvement and further work. Wediscuss a few points here.

Non-moving target localization. So far we are able to detect thenon-moving target through sensing his/her respiration or in-placeactivities like waving. Note that the sensing distance of through-wall respiration is still limited, since the signal changes caused bythe small respiration movement (around 5𝑚𝑚 chest displacement)is too small and is easy to be buried in noise. Also, the severemultipath makes it challenging to accurately localize the non-movingtarget even the virtual array constructed by device movement isemployed. In the future, we attempt to explore the feasibility ofutilizing beamforming technology [25, 53, 58] to amplify the weakreflected signal to detect non-moving target with respiration.

Target localization during device movements. In this paper, wecan localize the target by only single pair of transceiver when thetarget is moving. And we can remove the dynamic multipath inter-ference caused by target indirect reflection, according to the fact thatdynamic multipath is weaker than target direct reflection. But if the

SenSys’19, November 10-13, New York, NY, USA

transmitter/receiver is also moving, the original static componentwill also become dynamic, which is more likely stronger than thetarget direct reflection signal, so we can not get rid of the effect ofdynamic multipath in this case, resulting in large localization error.

Multiple-target identification. In this work, we are not able tolocalize multiple targets simultaneously since we can not distinguishthe dynamic multipath from multiple targets. In future work, we planto employ the method presented in [54] to separate mixed signalsfor multi-target sensing.

Despite these limitations, WIDESEE moves an important steptoward enabling wide-range contactless sensing. More importantly, itexplores the challenges in wide-range sensing and provides effectivesolutions. We believe WIDESEE provides valuable reference forresearch in this field.

6 RELATED WORKOur work is broadly related to the literature in two areas.

6.1 Human Activity RecognitionComputer-vision-based human activity sensing techniques have en-abled mature applications. For instance, Kinect [9] and Leap Mo-tion [3] can achieve fine-grained human gesture tracking. However,these systems are sensitive to lighting the condition and monitor-ing angle and cannot work when the target is blocked or behind awall. Wearable sensor-based solutions overcome the above limita-tions [14, 21, 33] but still bring in inconvenience to users as theyrequire instrumenting the users. However, it would be unrealistic toassume every target has a working wearable in emergencies.

Compared to a computer-vision or wearable solution, wireless-signal-based human activity recognition has the advantages of canpenetrate walls and does not require the user to carry a device. Ear-ly work in this area uses multiple transceiver pairs to constructa 3D lattice of wireless links to identify the presence of humanmovements [44]. Later work tries to use a single device equippedwith multiple antennas to realize activity recognition. For examples,WiSee can differentiate nine commonly seen body gestures withthe help of machine learning techniques [42]. The effectiveness ofthe learning based methods depends on the quality of the trainingdata, but obtaining good quality training data remains costly andnon-trivial. For example, CrossSense [24] requires collecting thou-sands of samples to learn a single activity recognition model for agiven environment. WIDESEE avoids the pitfalls of a learning-basedapproach by developing analytical models for activity recognition.It requires significantly less effort for collecting data samples andcan be portable to different environments.

Recent studies also show it is possible to sense the respiration [10,17, 19], heart rate [10], or even vital signs [38] using wireless signals.However, prior approaches only work at a small scale (e.g., the room-level) and would require dense deployment to work on a large area.WIDESEE builds upon these past foundations of human activitymodeling, to extend the scope of contactless human sensing for wideareas with a LoRa transceiver pair. Our work aims to close the gapof wireless sensing for disaster rescue in the urban areas, as well asterrorist search and security surveillance.

6.2 Indoor Localization and TrackingThere is an intensive body of work in localizing and tracking object-s [11, 12, 23, 29, 51, 58]. Prior work can be broadly grouped intotwo categories: device-based and device-free approaches.

A device-free approach has the advantage of not requiring theend-user to carry a device. By lifting the limitation of carrying adevice, device-free methods can target a wider range of applicationswhen compared to the device-based counterparts. WIDESEE thusfollows a device-free approach.

Object location and tracking can be realized through a range ofwireless signal characteristics, including AoA [23, 34], ToF [33, 36],and the signal amplitude [31]. An amplitude-based approach is sim-ple and cost-efficient, but it suffers from poor localization accuracy(especially in non-line-of-sight conditions) – due to additional signalattenuation resulted from obstacles and severe amplitude fluctuationdue to rich multipath indoors. Methods using phase informationsuch as AoA can effectively separate multipaths, a good resolutionand accuracy would require a large antenna array at the receiver.ToF-based methods are not ideal either, because they are limited thefrequency bandwidth.

WIDESEE is among the first attempts to realize contactless wide-area sensing with a single LoRa transceiver pair. It does so by comb-ing the long-communication LoRa signal and the mobility of thedrone. However, achieving the goal requires overcoming two chal-lenges: (1) the serve mutlitpath effects when using a single transceiv-er pair, and (2) the hurdle for not having phase or ToF informationon LoRa. Inspired by [32], WIDESEE extracts AoA-related informa-tion from amplitude measurements for localization, but it advancesprior work by relying on a single instead of multiple transceiverpairs. WIDESEE employs a set of new algorithms to remove thelocalization ambiguity caused by one transceiver pair. WIDESEE

also leverages and refines existing AoA based multipath removalmedhods [52]. The result is a promising solution for using a sin-gle transceiver pair for wide-area contactless sensing, which couldpotentially open up many new research opportunities.

7 CONCLUSIONSThis paper has presented WIDESEE, a novel hardware-software sys-tem that can perform wide-range wireless sensing using just onetransceiver pair. WIDESEE utilizes LoRa to obtain better throughwall penetration and larger range sensing. To further widen thesensing area, WIDESEE employs drones to carry the transceiverto improve the signal coverage. The combination of single LoRatransceiver pair and device mobility, however, brings new challengesof severe interference and sensing ambiguities (e.g., localization).To address these challenges, we design a set of techniques at thehardware and software layers, which can be applied to many wirelesssensing applications. We evaluate WIDESEE by developing a work-ing prototype and applying it across three different environmentsfor two representative sensing tasks – contactless human detectionand localization. Experimental results show the effectiveness ofWIDESEE for performing wireless sensing in a large site with a sin-gle LoRa transceiver pair. We believe WIDESEE marks an importantstep towards wide-range wireless sensing, and is highly attractive inreal-world scenarios like disaster rescue or terrorist search where itis impossible to deploy multiple devices.

WIDESEE: Towards Wide-Area Contactless Wireless Sensing SenSys’19, November 10-13, New York, NY, USA

REFERENCES[1] https://m.tb.cn/h.3CfHtz1?sm=3a9293/. 2017.[2] Gnuradio. https://www.gnuradio.org/. 2018.[3] Leap motion. https://www.leapmotion.com/. 2015.[4] Limesdr mini. https://limemicro.com/products/boards/limesdr-mini/. 2018.[5] Radio propagation. https://en.wikipedia.org/wiki/Radio propagation/. 2011.[6] Rfmax s9028pcr antenna. https://www.atlasrfidstore.com/

rfmax-s9028pcrj-s8658prj-rhcp-indoor-rfid-antenna-fcc-etsi/. 2015.[7] Semtech sx1276 transceiver. http://www.semtech.com/wireless-rf/rf-transceivers/

sx1276/. 2016.[8] Spreading wings s1000. https://www.dji.com/cn/spreading-wings-s1000/. 2015.[9] X-box kinect. http://www.xbox.com/. 2015.

[10] Fadel Adib, Mao Hongzi, Zachary E. Kabelac, Dina Katabi, and Robert C Miller.Smart homes that monitor breathing and heart rate. In ACM CHI, 2015.

[11] Fadel Adib, Zach Kabelac, and Dina Katabi. Multi-person localization via rf bodyreflections. In USENIX NSDI, 2015.

[12] Dhekne Ashutosh, Chakraborty Ayon, Sundaresan Karthikeyan, and RangarajanSampath. Strackio: Tracking first responders inside-out. In USENIX NSDI, 2019.

[13] Jennifer T. Bernhard. Reconfigurable antennas. Synthesis Lectures on Antennas,2(1):1–66, 2007.

[14] Chen Bo, Vivek Yenamandra, and Kannan Srinivasan. Tracking keystrokes usingwireless signals. In ACM MobiSys, 2015.

[15] Wang Chen, Liu Jian, Chen Yingying, Liu Hongbo, and Wang Yan. Towardsin-baggage suspicious object detection using commodity wifi. In IEEE CNS, 2018.

[16] Gao Chuhan, Li Yilong, and Zhang Xinyu. Livetag: Sensing human-object inter-action through passive chipless wifi tags. In USENIX NSDI, 2018.

[17] Zhang Fusang, Zhang Daqing, Xiong Jie, Wang Hao, Niu Kai, Jin Beihong, andWang Yuxiang. From fresnel diffraction model to fine-grained human respirationsensing with commodity wi-fi devices. In ACM UbiComp, 2018.

[18] Zeng Hansong and Zhao Yi. Sensing movement: Microsensors for body motionmeasurement. In Sensors, 2011.

[19] Wang Hao, Zhang Daqing, Ma Junyi, Wang Yasha, Wang Yuxiang, Wu Dan,Gu Tao, and Xie Bing. Human respiration detection with commodity wifi devices:do user location and body orientation matter? In ACM UbiComp, 2016.

[20] Aboulnasr Hassanien and Sergiy A. Vorobyov. Phased-mimo radar: A tradeoffbetween phased-array and mimo radars. IEEE Transactions on Signal Processing,58(6):3137–3151, 2010.

[21] Wang He, Tsung Te Lai, and Romit Roy Choudhury. Mole: Motion leaks throughsmartwatch sensors. In ACM MobiCom, 2015.

[22] William C. Jakes. Microwave mobile communications. Wilery-IEEE Press, 1994.[23] Xiong Jie and Kyle Jamieson. Arraytrack: a fine-grained indoor location system.

In USENIX NSDI, 2013.[24] Zhang Jie, Tang Zhanyong, Li Meng, Fang Dingyi, Nurmi Petteri, and Wang

Zheng. Crosssense:towards cross-site and large-scale wifi sensing. In ACMMobiCom, 2018.

[25] Wang Jingxian, Zhang Junbo, Saha Rajarshi, Jin Haojian, and Kumar Swarun.Pushing the range limits of commercial passive rfids. In USENIX NSDI, 2019.

[26] Gjengset Jon, Xiong Jie, McPhillips Graeme, and Jamieson Kyle. Phaser: En-abling phased array signal processing on commodity wifi access points. In ACMMobiCom, 2014.

[27] Wang Ju, Jiang Hongbo, Xiong Jie, Kyle Jamieson, and Xie Binbin. Lifs: Lowhuman-effort, device-free localization with fine-grained subcarrier information. InACM MobiCom, 2016.

[28] Wang Ju, Xiong Jie, Chen Xiaojiang, Jiang Hongbo, Rajesh Krishna Balan, andFang Dingyi. Tagscan: Simultaneous target imaging and material identificationwith commodity rfid devices. In ACM MobiCom, 2017.

[29] Wang Jue and Dina Katabi. Dude, where’s my card? rfid positioning that workswith multipath and non-line of sight. 2013.

[30] Muzammil Jusoh, Tamer Aboufoul, Thennarasan Sabapathy, Akram Alomainy,and Muhammad Ramlee Kamarudin. Pattern-reconfigurable microstrip patchantenna with multidirectional beam for wimax application. IEEE Antennas andWireless Propagation Letters, 13(1933):860–863, 2014.

[31] Wu Kaishun, Jiang Xiao, Yi Youwen, Min Gao, and Lionel M. Ni. Fila: Fine-grained indoor localization. In IEEE INFOCOM, 2012.

[32] Chitra R. Karanam, Belal Korany, and Yasamin Mostofi. Magnitude-based angle-of-arrival estimation, localization, and target tracking. In IEEE IPSN, 2018.

[33] Palanivel Kodeswaran, Ravindranath Kokku, Madhumita Mallick, and SayandeepSen. Demultiplexing activities of daily living in iot enabled smarthomes. In IEEEINFOCOM, 2016.

[34] Manikanta Kotaru, Kiran Joshi, Dinesh Bharadia, and Sachin Katti. Spot-fi:decimeter level localization using wifi. In ACM SIGCOMM, 2015.

[35] Shangguan Longfei and Kyle Jamieson. The design and implementation of amobile rfid tag sorting robot. In ACM MobiSys, 2016.

[36] Alexander Mariakakis, Souvik Sen, Jeongkeun Lee, and Kyu Han Kim. Sail:single access point-based indoor localization. In ACM MobiSys, 2014.

[37] Sara Minaeian, Jian Liu, and Young-Jun Son. Vision-based target detection andlocalization via a team of cooperative uav and ugvs. IEEE Transactions on systems,

man, and cybernetics: systems, 46(7):1005–1016, 2016.[38] Zhao Mingmin, Fadel Adib, and Dina Katabi. Emotion recognition using wireless

signals. In ACM MobiCom, 2017.[39] Rajalakshmi Nandakumar, Shyamnath Gollakota, and Nathaniel Watson. Contact-

less sleep apnea detection on smartphones. In ACM MobiSys, 2015.[40] Rajalakshmi Nandakumar, Vikram Iyer, and Shyamnath Gollakota. 3d localization

for sub-centimeter sized devices. In ACM SenSys, 2018.[41] A. Michael Noll. Principles of Modern Communications Technology. Artech

House, 2001.[42] Pu Qifan, Jiang Siyu, and Shyamnath Gollakota. Whole-home gesture recognition

using wireless signals. In ACM SIGCOMM, 2013.[43] Yun Sangki, Chen YiChao, Zheng Huihuang, Qiu Lili, and Mao Wenguang. Strata:

Fine-grained acoustic-based device-free tracking. In ACM MobiSys, 2017.[44] Li Sun, Souvik Sen, Dimitrios Koutsonikolas, and Kyu-Han Kim. Widraw: En-

abling hands-free drawing in the air on commodity wifi devices. In ACM MobiCom,2015.

[45] Vamsi Talla, Mehrdad Hessar, Bryce Kellogg, Ali Najafi, Joshua R. Smith, andShyamnath Gollakota. Lora backscatter: Enabling the vision of ubiquitous con-nectivity. 2017.

[46] Wei Teng and Zhang Xinyu. mtrack: High-precision passive tracking usingmillimeter wave radios. In ACM MobiCom, 2015.

[47] Teodor Tomic, Korbinian Schmid, Philipp Lutz, Andreas Domel, Michael Kasseck-er, Elmar Mair, Iris Lynne Grixa, Felix Ruess, Michael Suppa, and Darius Bursch-ka. Toward a fully autonomous uav: Research platform for indoor and outdoorurban search and rescue. IEEE robotics & automation magazine, 19(3):46–56,2012.

[48] Deepak Vasisht, Swarun Kumar, and Dina Katabi. Decimeter-level localizationwith a single wifi access point. In USENIX NSDI, 2016.

[49] Wang Wei, Alex X. Liu, and Sun Ke. Device-free gesture tracking using acousticsignals. In ACM MobiCom, 2016.

[50] Mao Wenguang, He Jian, and Qiu Lili. Cat: high-precision acoustic motiontracking. In ACM MobiCom, 2016.

[51] Li Xiang, Zhang Daqing, Qin Lv, Xiong Jie, Li Shengjie, Zhang Yue, Mei Hong,Li Xiang, Daqing Zhang, and Lv Qin. Indotrack: Device-free indoor humantracking with commodity wi-fi. Proceedings of the Acm on Interactive MobileWearable and Ubiquitous Technologies, 1(3):72, 2017.

[52] Li Xiang, Li Shengjie, Zhang Daqing, Xiong Jie, and Mei Hong. Dynamic-music:accurate device-free indoor localization. In ACM UbiComp, 2016.

[53] Fan Xiaoran, Ding Han, Li Sugang, Sanzari Michael, Zhang Yanyong, TrappeWade, Han Zhu, and Howard Richard E. Energy-ball: Wireless power transfer forbatteryless internet of things through distributed beamforming. ACM UbiComp,2018.

[54] Xie Yaxiong, Xiong Jie, Li Mo, and Kyle Jamieson. md-track: Leveraging multi-dimensionality inpassive indoor wi-fi tracking. 2018.

[55] Zou Yongpan, Xiao Jiang, Han Jinsong, Wu Kaishun, Li Yun, and Ni Lionel M.Grfid: A device-free rfid-based gesture recognition system. IEEE Transactions onMobile Computing, 16(2):381–393, 2017.

[56] Ma Yunfei, Nicholas Selby, and Fadel Adib. Drone relays for battery-free networks.In Proceedings of the Conference of the ACM Special Interest Group on DataCommunication, pages 335–347. ACM, 2017.

[57] Ma Yunfei, Nicholas Selby, Manish Singh, and Fadel Adib. Fine-grained rfidlocalization via ultra-wideband emulation. In ACM SIGCOMM Posters and Demos,2017.

[58] Ma Yunfei, Luo Zhihong, Steiger Christoph, Traverso Giovanni, and Adib Fadel.Enabling deep-tissue networking for miniature medical devices. In ACM SIG-COMM, 2018.