See discussions, stats, and author profiles for this publication at: [Link]

net/publication/324511855

Low-Complexity Portable Passive Drone Surveillance via SDR-Based Signal

Processing

Article in IEEE Communications Magazine · April 2018

DOI: 10.1109/MCOM.2018.1700424

CITATIONS READS

83 5,757

4 authors, including:

Samith Abeywickrama Zhang Lihao

24 PUBLICATIONS 1,367 CITATIONS

Chinese University of Hong Kong
11 PUBLICATIONS 137 CITATIONS
SEE PROFILE
SEE PROFILE

Chau Yuen
Nanyang Technological University
1,170 PUBLICATIONS 44,356 CITATIONS

SEE PROFILE

All content following this page was uploaded by Samith Abeywickrama on 26 July 2019.

The user has requested enhancement of the downloaded file.

 Low-Complexity Portable Passive Drone
Surveillance via SDR-Based Signal Processing
Hua Fu, Samith Abeywickrama, Lihao Zhang, and Chau Yuen

The authors present an AbstrAct passive detection and AoA estimation when the
overview on passive drone sends downlink signal. We discuss passive
drone detection. An SDR- This article presents an overview on passive drone classification using third-party RF signal. We
drone detection. An SDR-based, portable univer- then conclude the article.
based portable universal
sal software radio peripheral (USRP) prototype is
software radio periph- developed for detection in two scenarios. In the detectIon Methodology
eral (USRP) prototype is scenario in which a drone communicates with the A number of different methodologies for drone
developed for detection ground controller, the cyclostationarity signature detection have been proposed in academia and
in two scenarios. of the drone signal and pseudo Doppler principle industry. First, a video-based method [2] was
are employed. For the scenario in which a drone employed by using the camera. It can provide a
is not sending any signal, a micro Doppler signa- visual image. The main disadvantage is that it does
ture generated by the RF signal is exploited for not work well in bad weather conditions and it is
detection and identification. not applicable at night. Also, the size of the drone
is small, and it flies at low speed; thus, sometimes
IntroductIon it is difficult to discriminate drones from birds.
The use of civilian drones has increased dramat- For night detection, infrared sensing via a ther-
ically in recent years due to cost reduction. Like- mal camera was developed [3]. The heat emitted
wise, drones are fast gaining popularity around from a drone was used for detection by analyz-
the world. However, drone use in a problematic ing the thermal image. However, the effective
way has stirred public concerns. For example, in detection range for a small drone is usually a few
March 2016 a Lufthansa jet came within 200 ft of meters since small drones do not produce a lot of
colliding with a drone near Los Angeles Interna- heat. Audio-based detection has been employed
tional Airport; on April 2015 a drone was dropped for drone detection [4]. The acoustic signatures of
on a subway track near Commonwealth Station in different drones in the market are collected into a
Singapore; in January 2015 a drone crashed at database and compared with the recorded ones
the White House, raising concerns about security to find a match. Noisy urban environments with
risks to the government building; and drones have city traffic pose challenges for using this method.
been accused of being used to violate the privacy Furthermore, the drone noise signature can easily
and even carrying out criminal activities [1]. These be altered. A detection method based on medium
events give ample examples that developing a access control (MAC) address has been looked at
surveillance system for suspect drones is of par- as a possible candidate since many low-end com-
amount importance. Having a drone detection mercial unmanned aerial vehicles (UAVs) includ-
method and estimating its angle of arrival (AoA) ing drones have identifiable service set identifiers
are motivated by this purpose. Moreover, it is (SSIDs) and MAC addresses that are broadcast
equally important that the method should distin- [5]. The challenge lies in maintaining a database
guish the drone from other flying objects such as for all drones that are manufactured; and the
birds. This article presents an overview on passive ports may be modified, and the MAC address can
drone detection and classification and its AoA easily be spoofed to avoid detection.
estimation for two scenarios: the case in which Another technique is radar signature-based
a drone communicates to its ground controller detection [6]. In this method, the radar cross-sec-
and the case in which a drone is not sending any tion (RCS) measurements of the targets are
signal. The main contribution is threefold. First, observed for detection. Specifically, radio waves
we conduct a detailed survey of the literature on from the dedicated radar are transmitted, and the
various drone detection methodologies; second, target reflectivity characterized by RCS is used
we design a portable and practically workable to verify whether it is a drone or not. However,
prototype with low complexity by using the off- drones are challenging targets for RCS-based
the-shelf hardware; and third, we demonstrate the detection due to their size and the fact that they
effectiveness of the system through field imple- are made mostly of low-conductivity materials
mentation and experiment and discuss the practi- such as plastic, rubber, and foam, which reduce
cal challenges. their RCS. Commercial drones usually operate at
The article is organized as follows. We survey low altitude, which also makes RCS-based detec-
the various detection methodologies. We present tion challenging.

2 DRAFT
 Finally, the carrier RF-based approach provides Technique Methodology Pros and cons
a robust and effective method of drone detection
[7] as it can work day and night and in all weather (1) Provide a visual image
conditions. In general, there are two methods for Video [2] Use one or more camera(s) (2) Weather constrained
RF-based detection: active detection and passive (3) Low discrimination capability
detection. The active approach is similar to RCS-
(1) Can work at night
based detection, which requires that the dedicat- Infrared [3] Identify heat signature
(2) Short detection range
edly designed transmitter and receiver operate
in a cooperative way. The transmitter broadcasts Use external sound created (1) May not work well in a noisy environment
Audio [4]
an RF signal, whereas the receiver listens to and by flying drones (2) Acoustic signature can be altered
captures the reflected signal from the drone. The
reflected signal, which may contain some unique (1) Need to maintain database for all drones
MAC
features caused by the flying drone, is then ana- Service set identifier (SSID) (2) Ports may be modified
address [5]
lyzed to confirm the presence of a drone and its (3) MAC address can easily be spoofed
physical characteristics. One typical example is (1) Challenging task because the size of drone is
to use Doppler effects caused by air targets that usually small and most drones are made by low
provide velocity information, and hence enable Radar
RCS measurement conductivity materials and operate at low altitude
the detection of flying objects. Here, the Dop- signature [6]
(2) Difficult to distinguish from other objects like
pler measurements are passed through a series birds
of electronic filters and executed by several signal
processing algorithms to distinguish the drones RF (active) Backscattering of transmitter
(1) Transmitter is tailor-made
from the other flying targets [8]. In the passive [7] RF signal
(2) Better detection performance
approach, there is no dedicated transmitter; only (3) High cost (hardware, software, operation, etc.)
the receiver is needed to listen to and analyze
(1) Downlink video stream (1) Resilience to interception
electromagnetic (EM) waves generated by a RF (passive)
(2) EM scattering of (2) Low costs, easy deployment
drone, or generated by a third-party RF transmit- [8]
third-party RF signal (3) Poor detection performance
ter and scattered by a drone. Thus, unlike active
detection, where the transmitter is tailor-made Table 1. Summary of drone detection methods.
to facilitate detection, there is no cooperation
between the receiver and EM generator for pas-
sive detection. Compared to the active approach, and configuring the drone are generated by a
passive detection has the advantages of lower mobile phone. The controller then relays these
costs in terms of hardware and software require- instructions to the drone by sending the tele-com-
ments (e.g., instead of big power supply, only an mand data. Usually, the tele-command signal
electric battery is needed), operation and mainte- is sent regularly (e.g., 30 times/s for the Parrot
nance, portability and easy deployment, resilience [Link]).
to interception, and so on. Therefore, this article Downlink Channel: The flight information and
focuses on the passive approach, which includes onboard sensor readings such as altitude, posi-
two scenarios, namely, the drone is sending signal tion, flight speed, and path are sent by the drone
and the drone does not send any signal. Table 1 to its controller. These telemetry signals are also
recaps the various detection methods discussed. transmitted on a regular basis. For instance, the
Parrot [Link] sent the downlink information
PAssIve surveIllAnce vIA drone approximately 15 times/s and 200 times/s in
demo and full mode, respectively.
downlInk vIdeo sIgnAl Downlink First Person View (FPV)
In the case in which a drone is sending RF signal, Chan-nel: The video camera installed on the
drone surveillance can be achieved by eavesdrop- drone can transmit the live video in a digital
ping on the communication between a drone and format to the ground controller through
its ground controller. Technically, the system can downlink RF link. This video stream is then
detect a drone and estimate its AoA by process- displayed on the user’s mobile phone.
ing the data sent from the drone to its controller The communication technique used in the
using a wireless receiver. Since most commercial commercial drone is usually based on the existing
drones communicate frequently with their con- wireless standard protocols. For instance, both
trollers to update the status as well as to receive uplink and downlink channel in DJI Phantom 3
the commands, the data link between the drone drones are located in the industrial, scientific,
and its controller is always present. Hence, an RF and medical (ISM) band. The ground controller
receiver can collect wireless data samples, ana- can create a mobile wireless hotspot operating at
lyze them, and detect a drone’s presence. ISM band to which the mobile phone connects.
The modulation scheme for the uplink channel
coMMunIcAtIon between drone And uses frequency hopping spread spectrum. The
ground controller signal has 2 MHz bandwidth and hops in the 2.4–
The drone system usually includes three compo- 2.49 GHz range. The 8 default downlink chan-
nents: a drone, its controller, and a mobile smart- nels occupy the bandwidth from 2.401 GHz to
phone. The controller serves as the main source of 2.481 GHz; each has 10 MHz bandwidth. The
communication to the drone. The drone receives modulation scheme for downlink video streams
the signal from the controller to manage its flight uses orthogonal frequency-division multiplexing
status such as moving along a desired trajectory. (OFDM). This OFDM signal sent by the drone
The following are the three main communication provides a main source for performing drone
channels between a drone and its controller [9]: detection and AoA estimation.
Uplink Channel: The instructions controlling OFDM is a frequency-division multiplexing

3
 It is known that the sta-
Antenna array Signal processing using GNU radio
tistical parameters such
Drone detection
as mean, variance, and
autocorrelation for most CAF/SDC Cyclostationary
signature analysis
manmade communica-
tion signals are varying Antenna
switch AoA estimation (pseudo-Doppler)
in time with some
periodicities. This is due Phase AoA
Filter
demodulation estimation
to periodic keying of USRP B210
the amplitude, repeating
(a) (b)
spreading code, and the
preamble. Figure 1. Detection and estimation functional diagram with single-channel USRP receiver: a) USRP-based
receiver; b) detection and estimation algorithm.

scheme that employs multiple adjacent sinusoidal only one element data sample is extracted at a
tones as the carriers (i.e., the subcarriers) with- time. Nevertheless, there is a trade-off of the bet-
in the same channel bandwidth. The subcarriers ter performance of the multi-channel technique
are chosen to be orthogonal to each other in over the single-channel one. First, multi-channel
the sense that the peak of one subcarrier occurs implementation requires that the multiple receiv-
at the null of all others. OFDM uses fast Fourier ers should be coherent. In particular, the phase
transform (FFT) and inverse FFT (IFFT) to perform and the antenna gain should be synchronized,
multicarrier modulation and demodulation. At and this synchronization must be maintained over
the transmitter, IFFT is performed on the frequen- different signal processing time and a wide range
cy-domain subcarriers to produce time-domain of operating temperatures. In practice, real-time
symbols. At the receiver, FFT is carried out on the periodic calibration is used to achieve this goal.
symbols to recover the original data. Clearly, this will increase the hardware and soft-
ware complexity and power consumption. Sec-
usrP PrototyPe wIth sIngle-chAnnel receIver ond, compared to the multi-channel receiver, the
Drone detection and AoA estimation can be done single-channel system has the advantage in terms
using an antenna array with spatially displaced ele- of size, flexibility, weight, and costs. Therefore, the
ments. The physical location of the elements can multi-channel system is more suitable for applica-
geometrically be configured into linear, planar, tions with fixed site installation. For detecting a
and volumetric array. The array spacing can be flying drone, portability is crucial. Accordingly, a
uniform, non-uniform, and random. Uniform lin- handheld system with a single-channel receiver is
ear array (ULA) and uniform circular array (UCA) more desirable.
are perhaps the most popular antenna arrays. Figure 1 shows the software defined radio
Comparied to ULA, UCA is more compact spatial- (SDR)-based functional diagram used in our
ly. Therefore, UCA is more suitable for portable work. The system in Fig. 1a consists of the anten-
implementation. In general, the receiver includes na array, the universal software radio peripheral
the RF front-end circuit, which down-converts RF (USRP)-based single-channel receiver, and GNU
signal to intermediate frequency (IF) signal, the radio signal processing implemented on a laptop
analog-to-digital conversion (ADC) component computer. The antenna is composed of four ele-
where the IF signal is digitalized, and the direct ments that form a UCA. The output of one ele-
down-conversion (DDC) that transforms the sig- ment at a time is passed to the SDR, where signal
nal to the baseband. The in-phase and quadra- processing for drone detection and AoA estima-
ture-phase data samples of the baseband signal tion is implemented, as shown in Fig. 1b.
are then collected for signal processing to per-
form detection and estimation. drone detectIon vIA cyclostAtIonArIty sIgnAture
There are two types of receivers: single-chan- As discussed above, the drone signal consists of
nel receiver and multi-channel receiver. For the uplink remote control signal and downlink digi-
multi-channel receiver, each antenna element is tal telemetry/video streaming signal. The down-
followed by a receiver including RF front-end, link video signal has characteristic spectral shape,
ADC, and DDC. This means that if an array has which can be exploited for drone detection.
N elements, N receivers are needed. For sin- It is known that the statistical parameters such
gle-channel implementation, only one receiver as mean, variance, and autocorrelation for most
is used to serve multiple antenna elements. The manmade communication signals are varying in
well-known high-resolution estimation methods time with some periodicities. This is due to peri-
such as the MUSIC and ESPRIT algorithms are odic keying of the amplitude, repeating spread-
inherently multi-channel. This is because these ing code, and the preamble. The property that
algorithms require a snapshot observation; that statistical parameters of a random process vary
is, at a particular time instant, say t, the base- periodically in time is called cyclostationarity [10].
band data from all antenna elements should be A random signal is said to have second order
extracted simultaneously so that a data correla- cyclostationarity if its auto-correlation function
tion matrix can be formulated. This is not the is periodic in time. The autocorrelation is com-
case for the single-channel implementation since puted by multiplying the amplitudes at two time

4
instants t and t + t (where t is a time lag) of a 50
sample function and then averaging the product
over the ensemble (i.e., expectation operation).
In practice the expectation can be replaced by 30
the limit of a time average. The autocorrelation
function provides an analytical tool for finding 10

dB
the signal repeating pattern. Since the autocor-
relation function is periodic, it can be represent- -10
ed by a Fourier series via Fourier expansion. The
corresponding Fourier coefficients are called the -30
cyclic autocorrelation function (CAF) whose Fou- 0 2 4 6 8 10 12 14 16
rier transform is defined as the spectral correla- Time (ms)
tion density (SCD) function. Since the CAF/SCD (a)
function can be accurately measured even when -5
the signal-to-noise ratio (SNR) is low, it provides a -10
powerful tool for passive detection. -15
The downlink FPV communication of many -20
drones uses OFDM-based WiFi technology that

Power (dB)
-25
is based on the industrial IEEE 802.11 standards
[11]. The WiFi signal is generated on a frame-by- -30
frame basis. A frame includes preamble data fol- -35
lowed by OFDM symbols. The preamble used for -40
channel equalization and synchronization is trans-
-45
mitted with the known magnitude and phase. A
-50
64-point IFFT is implemented for a single OFDM 1 2 3 4 5 6 7 8 9 10
symbol with a carrier spacing of 312.5 kHz. This Frequency (kHz)
gives one IFFT output duration of 1/312.5kHz (b)
= 3.2 ms. Among the 64 subcarriers, 12 remain
unused, including the central zero subcarrier; 48 Figure 2. Cyclostationarity-based detection and Doppler-shift-based estimation:
are information data subcarriers; 4 are for pilot a) an example of cyclic autocorrelation function; b) Doppler shift due to
tones modulated by the same sequence of bina- rotating switch of antenna elements.
ry phase shift keying (BPSK) signal with known
magnitude and phase. The cyclic prefix of 0.8 er, the FM demodulator can identify the varying
ms is added as a guard interval to the beginning Doppler shift as a superimposed FM signal on the
of each data interval. The total symbol interval received signal. Therefore, the FM demodulator
is thus 4 ms. The analysis of CAF or SCD for this output includes a sine wave at the same frequen-
highly structural data frame can reveal the cyclos- cy of rotation rate. Taking the phase difference of
tationarity signature embedded in the WiFi-like this demodulated Doppler signal and a reference
OFDM signals. In particular, the pilot-induced and signal acquired from the antenna rotating circuit
cyclic prefix-induced cyclostationarity property allows us to determine the AoA.
provides very useful information and has more Rotating an antenna element mechanically is
impact. To facilitate detection, a table that lists the not desirable as well as not practical. Instead, the
carrier frequency and signal bandwidth of various element can be rotated electronically by sequen-
drones available in the market is generated, and tially switching the four UCA elements in such
the drones having WiFi-like OFDM signal on the a way that the single receiver connects to one
list can be detected. Figure 2a shows our exper- element at a time. The intuition is that since the
imental example on the pilot-induced CAF for elements are sequentially switched, the receiver
the DJI Phantom 3 signal with first peak appear- perceives it as if it were a single rotating anten-
ing at about 4.8 ms. A bigger peak appearing at na. Further, we can use the system clock — field
around 14 ms represents the periodic pattern for programmable gate array (FPGA)/USRP — to con-
the drone telemetry signal. Note that although trol the antenna array since the FPGA clock will
we have multiple peaks in Fig. 2a, for a given ensure that switching is in lockstep with samples
drone transmission, the patterns such as the sig- arriving at the host. For the RF rotating switch in
nal strength in CAF will be different for different Fig. 1, the EK42441 from Peregrine Semiconduc-
types of periodic signals. Consequently, no system tor was used in our experiment, and a Doppler
ambiguity will occur. shift of 7.8 kHz generated by an EK42441 rotat-
ing switch is shown in Fig. 2b.
AoA estIMAtIon by Pseudo-doPPler PrIncIPle As illustrated in Fig. 1, an AoA estimate
AoA estimation is based on the Doppler princi- is obtained by using an FM detector via phase
ple that causes the received signal frequency to demodulation. The received data is passed to a
increase if the antenna element is moving toward phase demodulator that extracts the phase. The
the impinging signal, and to decrease if it moves instantaneous frequency is derived by applying
away. Referring to Fig. 1, suppose that a UCA differentiation operation. A high-pass filter is then
antenna element rotates at a constant rate around employed to remove the direct current. This pro-
the central axis. As the rotation element moves duces a sinusoidal output with switch rate as the
alternately closer to and farther away the imping- frequency and AoA as the phase. Comparing the
ing signal, the received signal at the rotation ele- output to the switch reference gives the AoA.
ment is sinusoidally frequency-modulated with the
rotation frequency. If we pass the received sig-
nal through a frequency modulation (FM) receiv-

5
 The flying drone inter- be captured by the surveillance receiver. Hence,
Data sample collection in the case in which a flying drone is not send-
acts with the ambient ing any signal, third-party RF can be exploited to
RF signal. The reflected passively sense the reflected/scattered EM wave
Micro Doppler signature for drone surveillance. One directional antenna
and scattered EM wave estimation may point to the signal source to obtain a “clean”
will be captured by the reference signal, while the detection antenna will
surveillance receiver. listen to different directions to search for the pres-
Pre-processing for noise ence of EM reflections. The reason that, besides
Hence, in the case in reduction in frequency the surveillance antenna, we also need a refer-
domain
which a flying drone is ence antenna is that since there may be many
third-party RF signals, the detector has to know
not sending any signal, exactly which scattered/reflected signal will be
Drone type characterization
third-party RF can be through feature extraction used for passive signal processing. Figure 3 illus-
exploited to passively on micro Doppler signature trates the signal processing flowchart in which the
detection algorithm is based on the micro Dop-
sense the reflected/ pler signature.
scattered EM wave for Pattern classification based on
feature extracted and pattern MIcro doPPler sIgnAture
drone surveillance. recognition When a flying drone interacts with the RF
sig-nal, the properties of the scattered signal
Figure 3. Signal processing flowchart for passive reflect on the characteristics of the drone. It is
classification via a third-party RF signal. known that the carrier frequency of the
scattered sig-nal from the air target
experiences a frequency shift due to the
PAssIve detectIon vIA Doppler effect. The micro Dop-pler
thIrd-PArty rF sIgnAl phenomenon is an additional effect where any
other motion on the main body of the tar-get
The flight of a drone can also be autonomously will introduce further Doppler shift around
controlled by the onboard computer. For exam- the main shift due to such motion [12]. In this
ple, a GPS-based drone can fly on its own with sense, a micro-Doppler signature can
the destination, flight height, and speed being pre- provide extra information about the target. In
planned and configured. In this case the drone the case of a flying drone, the micro Doppler
can operate without sending any signal, which signature could be due to the vibration of the
may lead to a more serious security issue since drone body and the rotation of the drone
the drone used for criminal and terrorism activi- propeller, and is unique for different types of
ties can now fly in an autonomous manner, and flying drones since the micro-Doppler signature
hence operate beyond drone-controller range lim- associated with the way the drone operates
itation and become less noticed. Since no signal depends on the specific configuration of the
was sent, the detection has to rely on collecting drone structure and rotating propeller. For
and processing EM scattering and reflection by instance, micro-Doppler signatures for the same
a drone from the non-cooperative third-party RF propeller rotation rate but different number of
signal such as an ambient RF signal (i.e., illumi- blades can be exploited to differentiate drones.
nator of opportunity in the environment), which Moreover, experiments have shown that for the
includes the commercial broadcast such as TV, same number and speed of rotating pro-pellers,
digital audio/video broadcasting, navigation sig- different propeller material and size can lead to
nals, and cellular base station signal such as Long different micro Doppler signature.
Term Evolution. The actual micro Doppler signature
The flying drone interacts with the ambient RF created by propeller rotation is usually non-
signal. The reflected and scattered EM wave will stationary and complicated, and generally it is
difficult to find closed-form, explicit, and
exact mathematical expressions for micro
Doppler signature. To cap-
Illuminator
Drone
Illuminator
Drone

SDR receiver SDR

receiver

Figure 4. Experiment setup for micro Doppler signature.

6
 Propellor switched off Propellor switched on Since a machine can
2500 700 only “read” and handle
600 real numbers, features
2000
500 are usually represented
1500 numerically. Moreover,
400
Power

Power
FFT plot to facilitate learning
1000 300
and classification and
200
speed up this process,
500
100 features should be
0 0 highly informative and
915 915.200 915.400 915.600 915.800 916.0 916.200 916.400
Frequency (MHz)
915.200 915.400 915.600 915.800 916.000 916.200 916.400 discriminating, and the
Frequency (MHz)
feature number should
not be too large.
Frequency

STFT plot Frequency

10 20 30 40 50 60 70 80 90 100 110 120 20 40 60 80 100 120 140 160 180 200 220
Time (ms) Time (ms)

Figure 5. Micro Doppler signature of a drone propeller.

ture micro Doppler signature, some visual repre- Figure 5 shows one experimental result of
sentation can be used. One such representation is micro Doppler signature. The two images in the
obtained by processing micro Doppler signatures first row are generated by FFT plot, and the two
through spectrogram plotting. The spectrogram images in the second row are created by STFT.
is a time-frequency graph of the signal varying The two images in the first column are when the
with time. This is often generated by applying the drone propeller is switched off (i.e., only FFT and
short-time Fourier transform (STFT) to the time-do- STFT of the illuminator signal are plotted), and the
main micro Doppler signal. Briefly, the time-do- two images in the second column are the micro
main data generated by the scattered RF signal Doppler signature of the drone propeller. It can
is first segmented into short, equal time intervals be seen that the micro Doppler signature of a
called windows. Fourier transform is then applied rotating propeller is very distinctive.
to each window segment. In general, the imple-
mentation includes the following key steps: FeAture extrActIon And clAssIFIcAtIon For
1. Determine the window size and overlap drone detectIon
between successive windows. If micro Doppler signatures are distinctive, they
2. Choose windowing function. can even be separated visually by humans via
3. Multiply the data samples with windowing simply observing the spectrogram image. How-
function to generate window segments. ever, the task is not easy if a machine such as a
4. Perform FFT on each windowed segment. computer is used for automatic target recognition
Figure 4 illustrates our experiment setup for (ATR). Obviously, for drone detection, ATR imple-
measuring micro Doppler signature. A non-coop- mentation is more desirable. In this case machine
erative RF single-tone signal is used as the illumi- learning can be employed [13], which includes
nator. The experiment is carried out in an open two main components, namely, feature extraction
campus area surrounded by buildings and trees. and pattern classification. The key idea is that
The drone is a quadrocopter DJI Phantom 3. The since different types of drones have different
rotor diameter is 23.5 cm. The laptop computer micro Doppler signatures, through signal training
is HP ENVY DV6 with an Intel®CoreTM i7 pro- and classification a unique feature of a particular
cessor, and the software operating system is Linux drone can be derived, which is then used by ATR
Ubuntu 14.04. GNURadio 3.8 is adopted for the for pattern recognition.
SDR implementation. In the setup shown in the As the term suggests, feature extraction aims
left image, the illuminator, drone, and SDR receiv- at extracting the characteristic parameters that
er are roughly aligned, whereas in the right image, formulate the basis of micro Doppler signature.
the angle of illuminator-drone-receiver is set arbi- Since a machine can only “read” and handle real
trarily. numbers, features are usually represented numer-

7
 After feature extraction ically. Moreover, to facilitate learning and classifi- ly be characterized by a Rician distribution [15].
cation and speed up this process, features should The Rician factor reflects the degree of line of
the feature set can be be highly informative and discriminating, and the sight (LoS). In an urban environment, Rayleigh
passed to a classifier feature number should not be too large (using the fading is used to model the case that no LoS is
terminology, this means that the dimension of a available (i.e., the Rician factor becomes zero).
for identifying the types feature vector should be low). The first feature in Operation of a non-LoS environment can pose
of drones. The deci- micro Doppler signature is the main body velocity a serious challenge for both passive methods. It
sion about the drone of a drone. It is a parameter that describes the has been shown [15] that the drone flight alti-
motion of a drone body as a whole. The main tude may play a significant role in determining the
type can be made by Doppler shift is due to this velocity. Since most Rician factor and LoS probability, which in turn
a predefined system drones fly at relatively low speed, the main body affect the received SNR. Moreover, an adaptive
velocity may be weak. The second feature is the method can be used to enhance the detection
consisting of regularities rotation rate or period of the drone propeller. strategy in the case in which a drone loses its LoS
and rules established Due to the rotary nature of the propeller, the link [15]. These issues form promising research for
by a process of learning micro Doppler signal is periodic. The rotation our future investigation.
rate or period is a key feature parameter, as it is
through signal training. the main cause for the periodic spectrogram. The reFerences
third feature is the velocity of the rotating propel- [1] A. Morrow, “Couple Accuses Neighbor of Stalking with
ler blade tip. This parameter is the main reason for Drone,” sUAS News, 14 Dec. 2014; [Link]
com/2014/12/couple-accuses-neighbor-of-stalking-with-
the spectrum width, and thus can be estimated by drone/, accessed 7 Nov. 2016.
the spectral bandwidth. [2] Dronedefence, “Multi-Sensor Drone Warning System”;
There are several feature extraction tech- [Link] accessed 7
niques [13]. One popular one is singular value Nov. 2016.
[3] IEC, Banshee; [Link]
decomposition (SVD), which is a generalization grated-systems/banshee/, accessed 7 Nov. 2016.
of eigen-decomposition. SVD is a matrix factor- [4] [Link], Drone-Detector; [Link]
ization tool that is able to decompose any matrix [Link]/en/drone-detector-main-features/, accessed 7
into orthogonal-based spaces [14]. Further, SVD Nov. 2016.
[5] M. Richardson, “Drones: Detect, Identify, Intercept and
can find the best approximation to the original Hijack,” 2 Dec. 2015; https: //[Link]/uk/about-
data matrix using fewer dimensions. Therefore, us/newsroom-and-events/blogs/2015/december/drones-de-
SVD can offer both feature extraction and dimen- tect-identify-intercept-and-hijack/, accessed 7 Nov. 2016.
sion reduction. An STFT-based spectrogram image [6] T. Pultarova, “Radars to Prevent Drone-Aircraft Collisions
Already in Testing,” 8 Dec. 2014; [Link]
can be viewed as a 2D data set. One dimension content/articles/2014/12/radars-to-prevent-drone-aircraft-
represents the signal frequency structure, and the collisions-already-in-testing/, accessed 20 Nov. 2015.
other represents the dimension in which such [7] P. Nguyen et al., “Investigating Cost-Effective RF-Based
frequency structures are sampled. The goal of Detection of Drones,” Proc. 2nd Wksp. Micro Aerial Vehicle
Networks, Systems and Applications for Civilian Use, Singa-
feature extraction is to determine characteristic pore, June 2016, pp. 17–22.
frequency structures that vary with time. By apply- [8] DeTect, “Security and Surveillance Radar Systems,” http://
ing SVD, the spectrogram is decomposed into [Link]/[Link], accessed 7 Nov. 2016.
a product of left unitary matrix, diagonal matrix, [9] Keysight Technilogies, “RF Testing for Civil Unmanned
Aerial Vehicles,” Application Note, 18 Jan. 2016; http://
and right unitary matrix, and the frequency and [Link]/litweb/pdf/5992-1299EN.
time information are decoupled. The frequency pdf?id=2705785, accessed 27 Oct. 2016.
information is projected onto the singular vectors [10] W. A. Gardner, A. Napolitano, and L. Paura, “Cyclostation-
of the left unitary matrix, and the time information arity: Half A Century of Research,” Signal Processing, vol. 86,
no. 4, Apr. 2006, pp. 639–97.
is projected onto the singular vectors of the right [11] G. de Goeij, E. H. van Dijken, and F. Brouwer, “Research
unitary matrix. into Radio Interference Risk of Drones,” FIGO & Strict, May
After feature extraction the feature set can be 2016; [Link]
passed to a classifier for identifying the types of files/researchinto_radio_ interference_risks_of_drones.pdf,
accessed 31 Jan. 2017.
drones. The decision about the drone type can [12] C. Clemente et al., “Developments in Target Micro-Doppler
be made by a predefined system consisting of Signature Analysis: Radar Imaging, Ultrasound and Through-
regularities and rules established by a process of the-Wall Radar,” EURASIP J. Advances in Signal Processing,
learning through signal training. vol. 2013, no. 1, Mar. 2013, pp. 47–65.
[13] X. D. Zhang, Y. Shi, and Z. Bao, “A New Feature Vector
Using Selected Bispectra for Signal Classification with Appli-
conclusIon cation in Radar Target Recognition,” IEEE Trans. Signal Pro-
A comprehensive review on drone surveillance is cessing, vol. 49, no. 9, Sept. 2001, pp. 1875–85.
given. Two methods that are based on the USRP [14] S. Selvan and S. Ramakrishnan, “SVD-Based Modeling for
Image Texture Classification Using Wavelet Transformation,”
implementation and SDR signal processing for IEEE Trans. Image Processing, vol. 16, no. 11, Oct. 2007, pp.
passive drone detection and classification and 2688–96.
AoA estimation are discussed. For the method [15] M. M. Azari et al., “Ultra Reliable UAV Communication
that uses drone downlink signal, detection via Using Altitude and Cooperation Diversity,” arXiv preprint,
arXiv:1705.02877v1, [[Link]], 28 Apr. 2017, https:// arxiv.
ACF/SCD and AoA estimation by pseudo-Dop- org/abs/1705.02877, accessed 1 Aug. 2017.
pler principle are developed. For the method that
exploits the ambient RF signal, drone classification
through micro Doppler signature and machine
learning is presented. Our experiment also reveals
that in comparison to a carbon fiber propeller the
micro Doppler signature of a plastic propeller is
much less visible. The EM wave channel between
a drone and the surveillance receiver can often
be modeled by small-scale fading and large-scale
path loss, and the small-scale fading can general-

8
View publication stats