See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/333923767

Micro-UAV Detection and Classification from RF Fingerprints Using Machine

Learning Techniques

Conference Paper · March 2019

DOI: 10.1109/AERO.2019.8741970

CITATIONS READS
138 1,304

5 authors, including:

Martins Ezuma Fatih Erden

North Carolina State University North Carolina State University
33 PUBLICATIONS 490 CITATIONS 74 PUBLICATIONS 1,215 CITATIONS

SEE PROFILE SEE PROFILE

Chethan Kumar Anjinappa Ozgur Ozdemir

North Carolina State University North Carolina State University
49 PUBLICATIONS 526 CITATIONS 109 PUBLICATIONS 1,342 CITATIONS

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Non-Orthogonal multiple access for drone based communication networks View project

Handover Performance Analysis and Enhancement View project

All content following this page was uploaded by Fatih Erden on 22 July 2019.

The user has requested enhancement of the downloaded file.

 Micro-UAV Detection and Classification from RF
Fingerprints Using Machine Learning Techniques
Martins Ezuma, Fatih Erden, Chethan Kumar Anjinappa, Ozgur Ozdemir, and Ismail Guvenc
Department of ECE
North Carolina State University
Raleigh, NC 27606
{mcezuma,ferden,canjina,oozdemi,iguvenc}@ncsu.edu
arXiv:1901.07703v2 [eess.SP] 10 Apr 2019

Abstract—This paper focuses on the detection and classification Administration (FAA) and National Aeronautics and Space
of micro-unmanned aerial vehicles (UAVs) using radio frequency Administration (NASA) to integrate commercial micro-UAVs
(RF) fingerprints of the signals transmitted from the controller into the national airspace (NAS) [6].
to the micro-UAV. In the detection phase, raw signals are split
into frames and transformed into the wavelet domain to remove Even though so many beneficial civilian applications of
the bias in the signals and reduce the size of data to be processed. micro-UAVs abound, there is an associated risk to the public
A naive Bayes approach, which is based on Markov models safety. In recent times, there have been reports of micro-UAVs
generated separately for UAV and non-UAV classes, is used to violating public privacy and the security of sensitive facilities
check for the presence of a UAV in each frame. In the classifica- such as nuclear power plants and airports [7]. In 2018, a
tion phase, unlike the traditional approaches that rely solely on
time-domain signals and corresponding features, the proposed drone was intentionally crashed into a nuclear power plant in
technique uses the energy transient signal. This approach is France [8]. According to the FAA, reports of safety-incidents
more robust to noise and can cope with different modulation involving drones now average about 250 a month [9]. Some
techniques. First, the normalized energy trajectory is generated of these events involve micro-UAVs crashing into commercial
from the energy-time-frequency distribution of the raw control airplanes, military helicopters, the White House, and outdoor
signal. Next, the start and end points of the energy transient
are detected by searching for the most abrupt changes in the public events. Apparently, most of these events occur when
mean of the energy trajectory. Then, a set of statistical features drone pilots intentionally violate no-fly-zone restrictions. In
is extracted from the energy transient. Significant features are addition, micro-UAVs have been exploited by terror groups
selected by performing neighborhood component analysis (NCA) for the placement of improvised explosive devices (IED)
to keep the computational cost of the algorithm low. Finally, and chemical, biological, radiological, nuclear, and explosives
selected features are fed to several machine learning algorithms
for classification. The algorithms are evaluated experimentally (CBRNE) [10]. Recently, two armed commercial drones car-
using a database containing 100 RF signals from each of 14 rying powerful explosives detonated close to the Venezuelan
different UAV controllers. The signals are recorded wirelessly president during an outdoor event [11]. Therefore, there is
using a high-frequency oscilloscope. The data set is randomly an urgent need to secure the national airspace against such
partitioned into training and test sets for validation with the unconventional threats. This can be achieved by accurately
ratio 4:1. Ten Monte Carlo simulations are run and results are
averaged to assess the performance of the methods. All the micro- detecting and identifying non-compliant micro-UAVs.
UAVs are detected correctly and an average accuracy of 96.3% Several techniques have been proposed for micro-UAV
is achieved using the k-nearest neighbor (kNN) classification. detection and classification so far. Conventional radar-based
Proposed methods are also tested for different signal-to-noise techniques, which are widely deployed for detecting and
ratio (SNR) levels and results are reported. identifying aircrafts, mostly fail to detect micro-UAVs [12].
Alternative techniques like sound and video-based detection
are only suitable for short-range scenarios due to ambient
I. INTRODUCTION
noise [13]. Some of these challenges can be addressed by radio
In recent years, non-military micro-unmanned aerial vehi- frequency (RF) fingerprints-based techniques. However, the
cles (micro-UAVs) have proliferated conspicuously. Besides current trend on RF fingerprint classification of micro-UAVs
the recreational use by hobbyists, there is a growing interest focuses mostly on time-domain techniques which are not very
in the use of micro-UAVs for commercial applications. One effective. This is because time domain techniques are based
of the major areas of use is in precision agriculture, where on the assumption that there is an abrupt change at the start
micro-UAVs make it easy to map and survey farmlands point of the signal. However, this assumption is not always
for crop variability and phenology, crop dusting/spraying for true when the transition between the transient and noise is
weed and pest control, irrigation management, and livestock more gradual [14]. Consequently, time domain techniques may
monitoring [1]. Other commercial applications of micro-UAVs delay the detection of the transient of the signal. In worst case
include infrastructure health monitoring, package delivery, scenario, this may increase the probability of missed target
media & entertainment, and ad hoc access point Internet detection at low signal-to-noise-ratio (SNR).
connectivity [2]–[5]. Due to the potential benefits of micro- This paper is motivated by the need to address the afore-
UAVs, there is a collaborative plan by the Federal Aviation mentioned challenges. Due to the problems associated with

1
the use of the time-domain transient analysis, a new approach [21]. These flapping-induced micro-Doppler frequencies ap-
for the micro-UAV classification is proposed in this paper. In pear at a much lower frequency-band as compared to the
this approach, the time-domain signal is first transformed into micro-Doppler frequencies induced by the rotating propellers
the energy-time-frequency domain and the energy trajectory is of the micro-UAVs. Sparsity-based techniques can also be
computed. Then, a set of statistical features is extracted from used to extract features from radar micro-Doppler signatures.
the energy transient instead of the time-domain transient. The In [22], [23], orthogonal matching pursuit (OMP), a sparse-
dimensionality of the feature set is reduced using neighbor- coding based dictionary learning algorithm, is used to extract
hood component analysis (NCA) and the significant features features from the radar micro-Doppler signatures for automatic
are classified using several machine learning algorithms. It is target recognition.
shown that the discriminating features can still be extracted Although radar-based detection has been one of the main-
even when the time-domain signal waveform is distorted by stream approaches to the problem, their performance, i.e.,
noise. Moreover, a micro-UAV detection method is described maximum detectable range, is highly limited if the target
in this paper. RF signals are transformed into the wavelet has a low radar cross-section [24]. This also explains why
domain to remove the bias and reduce the size of the data. stealth aircrafts, designed to avoid radar detection, must have
Then, the naive Bayes approach is used based on the Markov very low radar cross section (RCS). A conventional aircraft
models to differentiate between the noise and micro-UAV (without stealth coating) has an average RCS value of 426.58
signals. m2 (26.3 dBsm) at lateral incidence to a millimeter wave
The paper is organized as follows. Section II gives an radar signal [25]. Therefore, in order to evade detection,
overview of the state-of-the-art techniques for micro-UAV de- military aircrafts are coated with radiation-absorbent materials
tection and classification; Section III provides a description of (RAM) to reduce the RCS. On the other hand, experimental
the process model for the approach; Section IV and Section V measurements at millimeter wave frequencies show that many
describes the proposed detection and classification techniques commercially available micro-UAVs have an average RCS
respectively; Section VI describes the experimental setup and value of about 0.02 m2 (-16.98 dBsm) [26]–[28]. This very
presents the results; and Section VII provides the concluding low RCS value is due to the shape and design material of
remarks. these micro-UAVs. Therefore, many micro-UAVs are naturally
stealth to conventional radars [24]. This explains the failure of
II. RELATED WORK the United States White House’s surveillance radar to detect
Existing techniques for micro-UAV detection and classifi- a micro-UAV flying across the fence and crash-landing into
cation can be categorized under four headings, namely, radar- the lawn [12]. This challenge with radar-based detection has
based, vision-based, sound-based, and RF fingerprinting. motivated researchers to investigate other detection techniques
for micro-UAVs.

A. Radar-based Techniques
B. Vision-based Techniques
Micro-UAV detection using radars have been widely stud-
ied. Radars transmit electromagnetic signals which interact In [29], a computer vision-based technique is described
with the target, in particular, the micro-UAVs. This interaction for micro-UAVs detection. This approach uses high resolution
causes a shift in the carrier frequency of the received signal due cameras to capture micro-UAVs in different background en-
to the Doppler effect. In addition to the main Doppler shift, if vironments. Several features such as Haar-like, histogram of
the target has vibrating or rotating structures, e.g., propellers, gradients (HOG), and local binary patterns (LBP) are extracted
vibrating platforms, and engines, these micro-motions will from the images. These features are fed into cascades of
induce time-varying frequency modulation on the received boosted classifiers for target detection. The cascaded boosted
signal [15]. These additional frequency modulations, called the classifiers perform detection at multi-stage sequences with
micro-Doppler effect, generate side-bands (or spectral lines) increasing complexity. In this system, only test sets that pass
around the main Doppler frequency shift. Analysis of the the previous stage are allowed into the next stage.
micro-Doppler signature may review some of the dynamic Deep learning networks have also been explored for the
characteristic of the target that can be used for target detection micro-UAV detection problem. Usually, deep learning tech-
and identification [16], [17]. niques do not rely on the human crafted features for target
In [18], the micro-Doppler signature of a quad-copter is detection. They autonomously learn the optimal features from
compared with that of a walking human. The study concludes the captured micro-UAV images. In [30], [31], convolution
that the unique micro-Doppler signature of the micro-UAVs neural networks (CNN) are investigated for micro-UAV detec-
are useful for the design of an automatic target recognition tion. These deep learning based techniques show fairly good
(ATR) system. These unique features can be used to dis- performance. However, training CNN networks requires huge
tinguish micro-UAVs from fixed-wing airplanes, helicopter, amount of data making real-time application computationally
and birds. In [19], micro-UAVs and small birds are classified expensive.
based on the eigenpairs extracted from the decomposition of In [32], [33], the authors described a computer-vision ap-
their micro-Doppler signatures. Larger birds can be readily proach based on a generic Fourier descriptor (GFD). The tech-
recognized and discriminated from small-UAVs because of the nique uses speeded-up robust features (SURF) for keypoint
frequency modulation induced by their flapping wings [20], detection on grayscale images of micro-UAVs. The keypoints

2
of interest are shape-descriptors of the micro-UAVs. In com- help detect and classify micro-UAVs. Unlike the radar-based
parison to CNN, this method provides much faster detection techniques, the RF sensing device/receiver is a passive listener
for micro-UAVs. However, all vision-based techniques suffer and does not transmit any signal. This makes RF fingerprint-
from one common problem. The performance of the camera based detection energy efficient. In addition, the challenge of
sensors depends on the ambient condition such as lighting. detecting micro-UAVs (with extremely low RCSs) is solved
In addition, vision-based detection of micro-UAVs may not since all that is required is to intercept the transmit signal
perform well if the surveyed area is large. from the micro-UAV controller. The range problem associated
with the vision and acoustic-based techniques can be solved
C. Sound-based Techniques by using high-gain receiver antennas together with a highly
sensitive receiver system to listen for the micro-UAV controller
Acoustic or sound-based techniques use arrays of micro-
signals. The issue of environmental noise can be suppressed
phones to extract the unique acoustic signature of micro-UAVs.
by employing several de-noising techniques, e.g., wavelet
Typically, micro-UAVs produce hissing or buzz-like sounds
decomposition and band pass filtering. These advantages make
in frequencies ranging from 400 Hz to 8 kHz [34]. This
RF fingerprint detection techniques a promising solution.
unique acoustic signature is due to the brushless DC motor of
In [40], GMMs are to detect the transient start points of
micro-UAVs. Using different audio analysis techniques, micro-
signals transmitted by a micro-UAV controller. This time-
UAVs can be separated from the background noise. In [35],
domain technique uses the expectation maximization (EM)
micro-UAV localization and tracking using an acoustic array is
algorithm for estimating the detection threshold. However,
described. The micro-UAVs are localized based on estimation
EM algorithm is not guaranteed to converge to a global
of the time difference of arrival (TDOA) of the received audio
optimal. Furthermore, in order to justify the use of GMM to
signals. In order to accurately compute TDOA, the authors
model the RF signal sampled with a high frequency, many
proposed an algorithm based on the Gauss priori probability
Gaussian components are needed. This will definitely increase
density function (GPDF).
the computational cost of the detection algorithm. In [41],
In [36], time and frequency domain acoustic features are
RF hash fingerprints are used to detect the micro-UAVs. The
extracted from micro-UAV audio recordings. These features
RF hash fingerprints are generated by calculating the distance
are used to train a multi-class support vector machine (SVM)
between peak locations in the envelope of the time-domain RF
for micro-UAV identification. In [37], the authors investigate
signal. The extracted fingerprints are used to train a distance-
the effectiveness of Gaussian mixture model (GMM) and deep
based support vector data description (SVDD) algorithm for
learning algorithms for drone sound detection. The problem,
target detection. However, since most micro-UAV controller
modeled as a binary classification problem, is based on the
signals have similar time-domain waveforms with random
detection of sound events. The study concludes that a long
spikes, time domain peak classification is not very effective.
short-term memory (LSTM) recurrent neural network (RNN)
shows the best micro-UAV sound detection performance. In
[38], micro-UAV detection using hybrid advanced acoustic III. UAV DETECTION/CLASSIFICATION SCENARIO
cameras is described. The system comprises 120 elements AND ASSUMPTIONS
microphone array and a video camera. The microphones are
Fig.1 shows the system setup for RF-based UAV detection
spherically arranged. Thus, allowing the system to simulta-
and identification. The detection system is an RF sensing
neously detect multiple micro-UAVs in 2D (angular position)
device which can capture signals from both the UAV and its
or 3D dimensions. The angular direction of a micro-UAV is
controller. In addition, the RF receiver also captures other
estimated using the phase of the acquired audio signals from
signals in the environment which co-exist with the UAV
the micro-UAVs.
transmissions in the same frequency band. Our overall goal
In [39], a similar hybrid audio-assisted detection system for
in this paper is to develop an algorithm which is capable of
micro-UAVs is described. The system consists of thirty high-
detecting a micro-UAV and, if present, identifying the type of
definition cameras and three microphones. In order to perform
it based on the extracted RF fingerprint of the UAV controller
micro-UAV detection, HOG features are extracted from image
signal. Thus, the process is divided into two major tasks,
data and mel-frequency cepstral coefficients (MFCCs) features
namely, the detection and classification. The detection stage
from the audio data. An SVM is trained to perform detection
makes a decision of whether the captured RF signal belongs to
on the test data set. The fundamental challenge of the audio-
a micro-UAV or the noise signal. If a micro-UAV is detected,
based systems is the practical range of the commercial micro-
then the classification stage is invoked to make a decision
phones. Most of them have a range of 25-30 ft [13] and are
regarding the type of the micro-UAV. Fig. 2 is a flowchart
highly sensitive to environmental noise.
that provides a graphical description of the overall process
model that involves UAV detection and classification.
D. RF fingerprinting The starting point for target classification is the target de-
RF fingerprint-based detection relies on the characteristics tection. This is achieved by continuously sensing the 2.4 GHz
of the RF signals of the micro-UAV controllers. Experimental channel for the presence of transmissions by a non cooperative
investigations show that the micro-UAV controllers have a UAV controller. Experimentally, it was observed that it is
unique RF signature due to the circuitry design and modulation much effective to detect the signal from the UAV controller
techniques employed. Therefore, RF fingerprint analysis can as against the transmission from the UAV itself because the

3
 Keysight DSOS604A High-
Definition Oscilloscope: 6 GHz, 4
Analog Channels
UAV
Antenna

UAV remote
controller

Captured waveform from UAV

remote controller

Fig. 1. System setup for RF-based UAV detection.

RF signal with multiple transients.

Fig. 3 shows typical RF signal waveforms captured from
data
six different micro-UAV controllers. These waveforms look
No distinct with well defined transitions. Therefore, Markov based
models can be effectively used to detect the UAV controller
Detection signal from the background noise. In this work, the back-
ground noise is modelled as an additive white Gaussian noise.
Yes According to [42], the maximum intercept range of the RF
sensing system can be modelled as:
Spectrogram r
λ Pt Gt GI
RImax = , (1)
Energy trajectory 4π LδI
where λ, Pt and Gt represents the transmit wavelength, trans-
NCA mit power and antenna gain of the UAV controller respectively,
GI is the antenna gain of the RF receiver (intercept) system,
L is the combined losses between the controller and receiver
Classifier (transmission and system losses) and δI is the sensitivity of the
receiver. Moreover, the RF receiver sensitivity, δI , is defined
... as:
UAV1 UAV2 UAV3 UAVn δI = kTo F Bρi , (2)

Fig. 2. Flowchart of the classification algorithm. where k = 1.38×10−23 Joules/k is Boltzmann constant, To =
290 k is the standard noise temperature, F is the noise factor of
the receiver, B is the bandwidth of the receiver and ρi is the
former has higher energy than the latter. For this work we SNR at the input of the receiver. In practice, the maximum
focus exclusively on the detection/classification of the RF range can be increased by using directional antenna for far
signal from the controller; RF signal detection/classification field detection.
from the UAV is left as a future work. The next task after signal detection is the classification pro-
Contrary to the classical energy detector system, which cess. This is achieved by using machine learning algorithms.
sets a single fixed threshold for target detection, the proposed First, the energy trajectory features of the captured RF signal
system is based on state transition probabilities. This approach waveform is extracted from the spectrogram representation
reduces the probability of false alarm due to random burst in of the signal. Thereafter, feature selection is then performed
the background noise. Moreover, this approach is motivated by using Neighborhood component analysis (NCA). The selected
the fact that actual RF signal waveforms from UAV controllers features are used for training and testing of the machine
can be considered as a time-varying spectral vector sequences learning models used for the classification.

4
 In the following, first, Section IV will provide a detailed the detection capability. The latter aspect is important for
description of the proposed RF-based UAV detection approach, low complexity system design which result in faster detection
while Section V will focus on the UAV classification problem. algorithms.
In this study, a three-stage wavelet decomposition is used as
IV. UAV DETECTION USING RF SIGNALS shown in Fig. 4. Low-pass g[n] and high-pass h[n] filters of
First step in Fig. 2 is capturing RF signal data and the the Haar transform are chosen due to their simplicity. Each
detection of the presence of a UAV signal in the data. In filter is followed by a down sampler. Wavelet coefficients
our experiments, the RF signals to be detected are captured obtained after the third level are considered as inputs to
from different micro-UAV controllers using a high-frequency the detection algorithm both for training and testing the RF
oscilloscope. The details of the experimental setup and data signals. An example of the RF signal received from the micro-
collection are given in Section VI. Each RF signal is recorded UAV controller of DJI Matrice 100 and the corresponding
such that it is a vector of the same size. wavelet transformed signal is shown in Fig. 5. It is clear from
Fig. 3 shows the typical RF signals received from six the figure that the wavelet transform removes the bias and
different micro-UAV controllers. As it is clear from the figure, reduces the number of samples while preserving the structure
each micro-UAV signal has a different waveform which can associated with the original raw signal.
be attributed to the unique characteristics of the transmitter
circuits, modulation techniques, and the packet structure. This B. Bayesian Decision Making
makes it unreliable to use simple thresholding techniques to We use a Bayesian approach for the decision making
detect micro-UAVs especially in noisy environments. It is process. To state the general problem, let C ∈ {0, 1} be an
also difficult to detect the time-domain transient to obtain the index denoting the class of the measured RF signal y. When
fingerprint of the corresponding control signal because the end y ∈ UAV class then C = 1; otherwise C = 0. Let yT be
point of the transient is not clear unless there is an overshoot a vector containing the transformed RF signal y. Then, the
in the signal. posterior probability of the UAV class given yT is
In order to perform detection, all non-UAV signals are
P (yT |C = 1)P (C = 1)
classified as noise. This includes background noise in the P (C = 1|yT ) = , (3)
receiver itself and interference from all other wireless sources P (yT )
such as Wi-Fi, Bluetooth, and microwave oven which operate where P (yT |C = 1) is the likelihood function conditioned
in same frequency band. Thus, the overall goal of the detection on C = 1, P (C = 1) is the prior probability of UAV class,
algorithm is to classify the received signals as belonging to and P (yT ) is the evidence. A similar expression holds for
the UAV or non-UAV class. However, due to the lack of data P (C = 0|yT ). In terms of posterior probabilities, we decide
for the non-UAV class at this point of a time, we restrict C = 1, if
our discussion to the classification of UAV and noise signals, P (C = 1|yT ) ≥ P (C = 0|yT ). (4)
albeit, the proposed method is developed for a more general
If we assume the number of samples from each of the classes
case. Other wireless sources will be explored in our future
in the training set are equal, the prior probabilities of the UAV
work. Therefore, the detection problem boils down to detecting
and noise class become equal. Thus, the decision is favored
the presence/absence of a UAV based on the received RF
to C = 1, if
signals in a noisy environment.
A pre-processing process is applied to the RF signals before P (yT |C = 1) ≥ P (yT |C = 0). (5)
proceeding with the detection algorithm. The RF signals are
transformed into the wavelet domain using a pre-defined While this simplifies the problem of making a decision, there
wavelet tree. The rationale behind the adapted technique [43] is still the problem of computing the likelihood P (yT |C =
is that it helps in detecting the possible RF signals even in {0, 1}). This calculation is central to any Bayes decision prob-
the low SNR regime. This leads to a better detection ability lem because it reflects the interdependence of the classes of
which is a necessity for applications like micro-UAV threat nature. In order to capture the dependency between the states,
detection. we incorporate the method discussed in the next subsection.
We will get back to the calculation of the likelihood after we
introduce the concept of states.
A. Pre-processing Step A close inspection of the collected data revealed that most of
The SNR of the received RF signal varies with the distance the UAV signals are differently structured (see Fig. 3). This is
and transmitted power of the wireless source. Higher SNRs true for signals from other wireless sources as well. However,
can be achieved in multiple ways [43], [44]. One such way the same cannot be said about the noise data. That is, often
is the use of the wavelet transform for removing the bias in the UAV data changes smoothly resulting in consecutive states
the received RF signals and de-noising up to a certain extent. of the signal that are not statistically independent. This is
[43], [45], [46]. The use of wavelets provides the two-fold particularly useful when the SNR is low because at low SNRs
advantage compared to the traditional time domain and Fourier the measured signal looks like a random noise.
domain analysis [47]. One is the improvement in the SNR In order to exploit the dependency between the adjacent
(de-noising) and other in the data compression without loss states, we define 3-state Markov models for each class. We
of information. The former aspect is important to improve define two thresholds T1 and T2 based on the amplitude of the

5
 0.2 0.2 0.2

Amplitude (Volts)

Amplitude (Volts)

Amplitude (Volts)
0.1 0.1 0.1

0 0 0

-0.1 -0.1 -0.1

-0.2 -0.2 -0.2

0 0.5 1 1.5 2 2.5 0 0.5 1 1.5 2 2.5 0 0.5 1 1.5 2 2.5
-4 -4
Time (s) 10 Time (s) 10 Time (s) 10-4
0.2 0.2 0.2
Amplitude (Volts)

Amplitude (Volts)

Amplitude (Volts)
0.1 0.1 0.1

0 0 0

-0.1 -0.1 -0.1

-0.2 -0.2 -0.2

0 0.5 1 1.5 2 2.5 0 0.5 1 1.5 2 2.5 0 0.5 1 1.5 2 2.5
Time (s) -4 Time (s) -4 Time (s) -4
10 10 10

Fig. 3. Examples of micro-UAV control signals for six different controllers: (a) DJI Matrice 100, (b) DJI Phantom 3, (c) Hobby King T6A V2, (d) DX6e
Spektrum, (e) JR X9303, and (f) Jeti Duplex DC-16 (from top left to bottom right).

With the help of the above thresholds, we define three-states

S1 , S2 , and S3 . Based on the amplitude of the signal at each
index n, a state is allotted to that index based on the following
decision rule:

S1 , yT [n] > T1

SyT (n) = S2 , T2 ≤ yT [n] ≤ T1 , (6)

Fig. 4. Three-stage wavelet decomposition. S3 , yT [n] < T2


where SyT (n) is a set containing the states of the signal. Based
wavelet transformed noise signal, which are used to distinguish on the above rule, it is straightforward to obtain the states
the three states. The values of T1 and T2 (with T1 > T2 ) are associated with any time series signal. Once the state sequence
fixed based on the training data. We present the rationale of SyT is obtained, the probability of a transition between any
the choice of T1 and T2 in the subsequent subsections and two states is calculated. The transition probability matrix are
validate the choice in the numerical simulation section. generated based on the transitions from adjacent indexed states
as seen in Fig. 6. The transition number and probability matrix

0.2 0.2

0.15 0.15

0.1 0.1
Amplitude (Volts)

Amplitude (Volts)

0.05 0.05

0 0

-0.05 -0.05

-0.1 -0.1

-0.15 -0.15

-0.2 -0.2
0.5 1 1.5 2 2.5 1 2 3
Time (s) 10-4 Time (s) 10-5

Fig. 5. The raw signal from DJI Matrice 100 (left) and the corresponding
output (right) obtained at the third stage of the wavelet decomposition. Fig. 6. Definition of transition numbers and probabilities between the states.

6
are defined as follows:
 
N11 N12 N13
45 -20
TN = N21 N22 N23  ,
40
 N31 N32 N33 -40
p11 p12 p13

Power/frequency (dB/Hz)
TN 35
TP = p21 p22 p23  = P , -60
i,j Nij
30
p31 p32 p33
25 -80
respectively, where TN is a matrix capturing the number of
20
-100
transitions between any two states, Nij is the number of
15
samples transiting from state Si to Sj ; the matrix TP captures -120
the transition probabilities and is obtained upon normalizing 10

the TN matrix with the total number of samples. Here, pij 5 -140

is the transition probability from state Si to Sj for the signal

0 2 4 6 8 10
of interest, i.e., pij = p(Si → Sj ). Note that the transition
Frequency (GHz)
probabilities generated for the UAV and the noise data will be
different when the considered SNR level is modest. Also, note Fig. 7. Spectrogram of the RF signal shown in Fig. 5.
that the choice of T1 and T2 dictates the transition probabilities
for both the noise and UAV class data.
The choice of the thresholds has a direct impact on the values. If the signal of interest is from the UAV class, then
decision making process. T1 and T2 are set to ±3σ of the the classification stage is invoked.
wavelet transformed environmental noise signal, where σ
represents the standard deviation. The basis for this choice is V. UAV CLASSIFICATION USING RF
that the environmental noise often is modeled as a Gaussian FINGERPRINTS
noise and the noise samples will be within the ±3σ band We propose a technique based on the energy-time-frequency
with a very high probability (≈ 0.993). The validation of the domain. Energy transients extracted in this domain can be used
choice is presented in the numerical simulation section. Based as the fingerprints of the corresponding signals.
on these settings, the UAV and noise class training transition For the representation of the RF signals in the energy-time-
probabilities are calculated. frequency domain, we use the spectrogram method. The spec-
The following procedure is followed to obtain the UAV trogram of any signal is computed using the squared magni-
transition probability matrix. The desired UAV packets from tude of the discrete time short-time Fourier transform (STFT)
all the classes are appended and the states and transition 2
∞
probabilities are generated. A similar approach is employed X
−jωn
Spectrogram(n, ω) = y[m]w[n − m]e , (10)
to calculate the noise transition probabilities based on the
n=−∞
collected environmental noise data.
For a given test signal, the signal goes through the similar where y[n] is the captured signal, m is discrete time, ω is
pre-processing steps. Based on the pre-processed output, the the frequency, and w[n] is a sliding window function that
states are defined and the TN is calculated. Finally, the acts as a filter. In addition, the spectrogram analysis of the
likelihood of the class being a UAV is calculated as follows: captured RF signals can reveal the transmit frequency of the
X signal as well as the frequency hopping patterns. These are
P (yT |C = 1) = P (Si → Sj |C = 1), vital detection information. The spectrogram of the RF signal
i,j={1,2,3}: in Fig. 5 is shown in Fig. 7. It can be clearly seen that the
Y
= TPC=1 (i, j)TN (i,j) , transmit frequency of the signal is 2.4 GHz.
(7) The spectrogram, by definition, displays the energy/intensity
i,j={1,2,3}:
Y distribution of the signal along the time-frequency axis.
= pN N12 N33
11;C=1 p12;C=1 . . . p33;C=1 .
11
Therefore, the energy trajectory can be computed from the
i,j={1,2,3}:
spectrogram by taking the maximum energy values along
The log-likelihood of the above expression results in the time-axis. From this distribution, we estimate the energy
X transient by searching for the most abrupt change in the mean
log (P (yT |C = 1)) = Nij log(pij;C=1 ). (8) or variance of the normalized energy trajectory. The energy
i,j={1,2,3}:
transient defines the transient characteristics of the signal in
Similarly, the log-likelihood of the signal coming from a noise energy domain and is represented by fE (n), n = 1, . . . , N .
class is calculated by For the RF signal in Fig. 5, the normalized energy trajectory
X computed from the spectrogram, and the corresponding energy
log (P (yT |C = 0)) = Nij log(pij;C=0 ). (9) transient are shown in Fig. 8.
i,j={1,2,3}:
Once the energy transient is detected, the RF fingerprints
The decision will be favored to C = 1, if log(P (yT |C = are extracted. These fingerprints are the statistical moments
1)) ≥ log(P (yT |C = 0)); otherwise, C = 0. We discuss the that describe the energy transient. The extracted features are
detection results in the simulation section for different SNR skewness (γ), variance (σ 2 ), energy spectral entropy (H), and

7
 1.5 (NN) [48]. Since some of the features may be correlated,
and so redundant, we perform feature selection to reduce
1 the computational cost of the classification algorithm. This
Normalized energy trajectory

is discussed next.
0.5

0 Energy transient A. Feature Selection Using NCA

In practice, it is often required to reduce the dimensionality
-0.5
of a feature set by removing correlated features. By this
means, computational cost of the classification algorithm can
-1
be reduced. Most often, a feature selector is a linear operator
that projects the original data or feature set into a lower
-1.5
0 10 20 30 40 50 dimensional space. Neighborhood component analysis (NCA)
Time ( s) is such a linear projector.
NCA is a non-parametric, embedded, and supervised learn-
Fig. 8. Energy trajectory computed from the spectrogram in Fig. 7.
ing method for feature selection. NCA learns a matrix
by which the primary data are transformed into a lower-
kurtosis (k). Physically, γ is a measure of the asymmetry of dimensional space [49]. In this lower-dimensional space, the
the energy distribution around the mean value; σ 2 measures features are ranked according to a weight metric, with the
the spread of the energy trajectory about the mean value; H more important features receiving higher weight values. NCA
provides a measure of the Shannon entropy (energy spectral achieves this goal by maximizing the regularized objective
power), and k is a measure of the sharpness or flatness of the function f (w) with respect to the weight variable w. The
energy transient. These features are defined in terms of the regularized objective function is defined as:
mean (µ) and standard deviation (σ) of fE as follows N p
1 X X
1
N f (w) = pi − λ wr2 , (12)
N i=1
X 3
γ(fE ) = (fE (n) − µ) r=1
N σ3 n=1
N where λ is the regularization term, N is the number of
1 X 2
σ 2 (fE ) = (fE (n) − µ) samples in the feature set, and pi is the average leave-one-out
N n=1 probability (LOO). In other words, pi is the probability that
(11)
N
X NCA correctly learns an observation in the feature set. In order
H(fE ) = − fE (n) log2 fE (n) to perform feature selection, NCA uses the regularization term
n=1 to drive to zero all the weights corresponding to the redundant
N or correlated features. In [49], NCA is compared with the
1 X 4
k(fE ) = (fE (n) − µ) . linear dimensionality reduction (LDA) on several dataset. It
N σ 4 n=1
is observed that if the classes are not convex and cannot be
The feature sets consisting of the above statistical parameters linearly separated, then LDA result will be inappropriate. In
are used to train four popular machine learning algorithms: contrast, NCA adaptively finds the best project matrix without
kNN, discriminant analysis (DA), SVM, and neural networks assuming any parametric structure in the lower dimensional
space. In the same experiment, NCA was shown to outperform
relevant components analysis (RCA) and principal component
7
analysis (PCA).
6 NCA ranks the features according to their importance. Fig. 9
shows the results of the NCA performed for the data to be
5 mentioned in Section VI. Fig. 9 shows the result of NCA. The
RF fingerprints are ranked according to their weight value. As
Feature weight

4 can be seen, kurtosis has the highest weight and so is the most
important RF fingerprint for this test case. On the other hand,
3
skewness has the lowest weight and so is the least important
2
RF fingerprint. This behavior is reasonable because there is
a correlation between the features skewness and kurtosis.
1 Consequently, for the training and testing, the classifiers can
discard skewness and still produce good results. This can
0 prevent the over-fitting problem when training the classifiers.
Skewness Variance Entropy Kurtosis
Feature index In addition, for large-scale classification problems, there can
be huge computational saving in training the classifiers with
Fig. 9. NCA results showing the relative importance of the statistical features. fewer number of significant features.

8
 TABLE I
UAV C ONTROLLER C ATALOGUE .

ID 1 2 3 4 5 6 7
Model DJI Inspire 1 DJI Matrice DJI Phantom DJI Phantom 3 DX5e DX6e FlySKy FS-T6
Name 100 4Pro Spektrum Spektrum

ID 8 9 10 11 12 13 14
Model Futaba T8FG Graupner Hobbie JR X9303 DX6i Turnigy 9X Jeti Duplex
Name MC-32 King-T6A V2 Spektrum DC-16

TABLE II
D ETECTION P ERFORMANCE .

SNR (dB)
0 2 4 6 8 10 12 14 16 18 20 22 24
Detection 13 19 23 46 61 84 100 100 100 100 100 100 100
Accuracy (%)

100
Kurtosis
Entropy
Variance
80 Skewness

Accuracy (%)
60

40

20

0
Fig. 10. Near-field measurement setup. kNN DA SVM NN
Method

Fig. 12. Impact of the features on the classification accuracy of the different
machine learning methods.

100
kNN
DA
SVM
80
NN
Accuracy (%)

60

40

Fig. 11. Far-field measurement setup. 20

VI. EXPERIMENTS AND RESULTS 0

25 20 15 10 5 0
SNR(dB)
A. Experimental Setup
RF signals are collected indoor from 14 micro-UAV con- Fig. 13. Classification accuracy versus SNR.
trollers operating at 2.4 GHz in near-field. Table I gives
the catalogue of the micro-UAV controllers used for data
collection and their class label (ID). The indoor experimen- Figs. 10 and 11, respectively. The RF sensing and detection
tal environment is very noisy with strong interference from system consists of a 6 GHz bandwidth Keysight MSOS604A
several sources operating in the 2.4 GHz frequency band: Wi- oscilloscope with the highest sampling frequency of 20 Gsa/s,
Fi, Bluetooth and micro-wave oven. The experimental setup 2 dBi omnidirectional antenna, and 24 dBi Wi-Fi grid antenna.
for both near-field and far-field indoor scenarios are shown in The antennas operate in the 2.4 GHz frequency band. The

9
 100 Once a UAV controller is detected, the received RF signal
should be classified to identify which UAV it is. In order to val-
90
idate the classification methods, 10 Monte Carlo simulations
are run. The average accuracy of each method is calculated for
a number of cases. Fig. 12 shows the classification accuracy
Accuracy (%)

80 of each method as well as the impact of different feature

selections on the performance of that method. kNN and SVM
70
perform similarly with a classification accuracy of 96.3% and
96.84%, respectively, and are followed by DA with an accu-
kNN
racy of 88.15%. NN can only achieve a classification accuracy
60 DA of 58.49% when there are 14 micro-UAV controllers. Fig. 12
SVM
NN
also verifies the results of NCA given in Fig. 9 where we
50
see the relative importance of the features in the classification
6 8 10 12 14 accuracy. From the figure, it is obvious that kurtosis is the most
Number of controllers
significant feature, that is, the feature contributing the most to
Fig. 14. Classification accuracy versus the number of controllers at SNR of the classification accuracy. As predicted by NCA, considering
25 dB. skewness in addition to the other three features contributes the
least to the classification accuracy. This observation holds for
all the methods.
omnidirectional antenna is used to capture the UAV controller
RF signal at close distance while the grid antenna is used Fig. 13 shows the classification performance of the methods
for far-field signal capture (reduced SNR scenarios). At the at different SNRs. As expected, the classification accuracy
near-field, the SNR is about 30 dBi and decreasing with the decreases as the distance between the UAV controller and the
distance from the receiver. receiver system increases. At an SNR of 25 dB, corresponding
The RF signals from the micro-UAV controllers are captured to a distance of about 5 m, kNN achieves a classification
by the antenna and fed into the oscilloscope (receiver system). accuracy of about 97.29%. At this distance, DA and SVM
The collected data are automatically saved in a cloud database show similar performance. However, NN achieves a classi-
for post processing. For each controller, 100 RF signals are fication accuracy of only 57.14%. When the SNR is 10 dB,
collected. Each RF signal is a vector of size 5000000 × 1 and corresponding to a distance of about 80 m, kNN, DA and SVM
has a time span of 0.25 ms. The database are partitioned with achieves a classification accuracy between 60-70% while NN
the ratio p = 0.2. That is, 80% of the saved data is randomly performs below 50%.
selected for training and the remaining 20% is used for testing In Fig. 14, we investigate the robustness and stability of the
(4:1 partitioning). classifiers as the number of UAV controllers increases. At SNR
of 25 dB, we see that the performance of the kNN, DA, and
SVM remains almost the same with the change in the number
B. Results
of the controllers. However, NN shows instability when there
During the experiments, the environmental noise was fairly are 8 or more controllers. It is clear that the NN is not a good
static. However, there were significant scattering and absorp- choice unless there are 6 or less controllers at least with the
tion from objects in the hallway. We took measurements in feature set used in this study.
the hallway at various distances up to 130 m. We observed the
It is obvious from the observations so far that kNN performs
signal vanishes beyond 130 m. Thus, the hallway behaves like
the best and NN performs the worst. This is probably because
a lossy rectangular waveguide. In addition, it was observed that
we did not optimize the hyper parameters of the NN algorithm.
the polarization planes of the transmitter and receiver antennas
In general, NN algorithms are very sensitive to the choice of
significantly affected the received signal strength. It was
hyper parameters used. On the other hand, hyper-parameter
noticed that ±3σ of the Wavelet transformed environmental
optimization was built into the kNN, DA and SVM classifiers
noise was around 0.0098 volts. Thus, throughout the work, T1
used in this study. Hyper-parameter optimization for the NN
and T2 was set to ±0.0098 volts, respectively.
method will be investigated in our future works. Therefore,
The performance of the detection algorithm is presented
because of the superior performance of kNN in this study, it
in Table. II. As expected, we see that the detection ac-
will be considered as the base classification method.
curacy increases with the SNR. When SNR=10 dB, which
corresponds to a distance of 80 m, the detection accuracy Table III shows a sample confusion matrix obtained for the
is 84%. The detection accuracy increases as we reduce the kNN classifier. This table describes the performance of the
distance between the UAV controller and receiver antenna. The kNN model on a set of test data (RF signals) for which the
system is able to detect all the UAVs at any SNR values of true labels are known. The test data were captured at an SNR
beyond 12 dB. The performance could be further improved by of 25 dB. The confusion matrix shows that kNN achieves
attaching an external low-noise power amplifier to the input an accuracy of 97.1%. Except one or two samples from 4
of the oscilloscope. This will reduce the input noise from the controllers, the classsifier is not confused between the micro-
environment and improve the detection accuracy. UAVs.

10
 TABLE III
C ONFUSION M ATRIX FOR THE K NN M ETHOD C OMPUTED F ROM 280 T EST S IGNALS O BTAINED F ROM 14 D IFFERENT M ICRO -UAV C ONTROLLERS AT
SNR OF 25 D B.

Target Class
ID 1 2 3 4 5 6 7 8 9 10 11 12 13 14 Out (%)
1 20 0 1 0 0 0 0 0 0 0 0 0 0 0 95.2
2 0 20 0 0 0 0 0 0 0 0 0 0 0 0 100
3 0 0 19 0 0 0 0 0 0 0 0 0 0 0 100
4 0 0 0 20 0 0 0 0 0 0 0 0 0 0 100
Predicted Class

5 0 0 0 0 20 2 0 1 0 0 0 2 0 0 80.0
6 0 0 0 0 0 18 0 0 0 0 0 0 0 0 100
7 0 0 0 0 0 0 20 0 0 0 0 0 0 0 100
8 0 0 0 0 0 0 0 18 0 0 0 0 0 1 94.7
9 0 0 0 0 0 0 0 0 20 0 0 0 0 0 100
10 0 0 0 0 0 0 0 0 0 20 0 0 0 0 100
11 0 0 0 0 0 0 0 0 0 0 20 0 0 0 100
12 0 0 0 0 0 0 0 1 0 0 0 18 0 0 94.7
13 0 0 0 0 0 0 0 0 0 0 0 0 20 0 100
14 0 0 0 0 0 0 0 0 0 0 0 0 0 19 100
Out 100 100 95 100 100 90 100 90 100 100 100 90 100 95 97.1
(%)

VII. CONCLUSION [3] E. Ackerman and E. Strickland, “Medical delivery drones take flight in
east africa,” IEEE Spectrum, vol. 55, no. 1, pp. 34–35, 2018.
In this paper, we investigated the problem of detecting and [4] V. Gatteschi, F. Lamberti, G. Paravati, A. Sanna, C. Demartini,
classifying micro-UAV control signals. The detection algo- A. Lisanti, and G. Venezia, “New frontiers of delivery services using
rithm uses a Bayesian approach based on the Markov models drones: A prototype system exploiting a quadcopter for autonomous drug
shipments,” in Proc. IEEE 39th Annual Comp. Soft. Applications Conf.
of UAV and non-UAV classes while the classification method (COMPSAC), vol. 2, Taichung, Taiwan, July 2015, pp. 920–927.
relies on the energy-time domain RF signal and uses features [5] W. Shi, J. Li, W. Xu, H. Zhou, N. Zhang, S. Zhang, and X. Shen,
extracted in this domain. We show that the kNN classifier “Multiple drone-cell deployment analyses and optimization in drone
assisted radio access networks,” IEEE Access, vol. 6, pp. 12 518–12 529,
performs the best while NN has the the worst performance 2018.
when considering lower SNR levels and increased number of [6] P. Kopardekar and S. Bradford, “UAS traffic management (UTM)
controllers. We obtain an accuracy of above 80% with the research transition team (RTT) plan: FAA and NASA collaborative
kNN classifier up to SNR of 15 dB for 14 controllers. We also efforts planned through September 2020,” accessed: 10-31-2018.
[Online]. Available: https://utm.arc.nasa.gov/docs/2017-FAA_NASA_
show that it is possible to increase the number of controllers UTM_RTT_Plan.pdf
up to a certain level without compromising the performance [7] A. Solodov, A. Williams, S. Al Hanaei, and B. Goddard, “Analyzing the
using kNN and SVM methods where both result in accuracy of threat of unmanned aerial vehicles (UAV) to nuclear facilities,” Security
J., vol. 31, no. 1, pp. 305–324, 2018.
above 95%. In the future work, we will perform experiments in [8] Reuters, “Greenpeace slams superman-shaped drone into nuclear plant,”
outdoor environment using multiple sensors and UAV signals 2018. [Online]. Available: https://nypost.com/2018/07/03/greenpeace-
for micro-UAV detection and classification. This approach will slams-superman-shaped-drone-into-nuclear-plant/
[9] A. Levin, “FAA warns of drone collision risks with airplanes,” 2017.
be more effective in modern electronic warfare environment, [Online]. Available: https://www.bloomberg.com/news/articles/2017-11-
where autonomous military UAVs employ low probability of 28/faa-warns-of-drone-collision-risks-with-airplanes-as-use-grows
intercept (LPI) emitters which are difficult to detect by a [10] M. Hutter and R. Scurek, “Possibilities of misuse of unmanned aerial
single RF sensing platform due to their low peak power. This vehicles (UAV) to terrorist targets,” Prace Naukowe Akademii im.
Jana Długosza w Cz˛estochowie. Technika, Informatyka, Inżynieria Bez-
problem can be addressed in our future work by using netted pieczeństwa, vol. 4, pp. 195–202, 2016.
sensor fusion system and deep learning algorithms for cluster [11] C. Colin, “Approaching a new normal: What the drone
fingerprinting based detection and classification of these LPI attack in venezuela portends,” 2018. [Online]. Avail-
able: https://www.rand.org/blog/2018/08/approaching-a-new-normal-
emitters. Moreover, in such an advanced system, techniques what-the-drone-attack-in-venezuela.html
for specific emitter identification (SEI) such as the formation [12] M. Schmidt and M. Shear, “A drone, too small for radar
of 3D feature cluster map could be investigated for improved to detect, rattles the white house,” 2015. [Online]. Available:
https://www.nytimes.com/2015/01/27/us/white-house-drone.html
classification. These are beyond the scope of the current work.
[13] I. Bisio, C. Garibotto, F. Lavagetto, A. Sciarrone, and S. Zappatore,
“Unauthorized amateur UAV detection based on wifi statistical finger-
ACKNOWLEDGMENT print analysis,” IEEE Communications Magazine, vol. 56, no. 4, pp.
This work has been supported by the NASA grant 106–111, 2018.
[14] J. Hall, M. Barbeau, and E. Kranakis, “Detection of transient in radio
NNX17AJ94A. frequency fingerprinting using signal phase,” Wireless and Optical
Communications, pp. 13–18, 2003.
R EFERENCES [15] V. C. Chen, The micro-Doppler effect in radar. Artech House, 2011.
[1] G. Sylvester, “E-Agriculture in Action: Drones for Agriculture,” [16] Q. Zhang, Y. Luo, and Y.-a. Chen, Micro-Doppler characteristics of
2018, accessed: 11-31-2018. [Online]. Available: http://www.fao.org/3/ radar targets. Elsevier, 2016.
I8494EN/i8494en.pdf [17] T. Thayaparan, S. Abrol, and E. Riseborough, “Micro-doppler radar
[2] D. Kang and Y.-J. Cha, “Autonomous uavs for structural health mon- signatures for intelligent target recognition,” Defence Research and
itoring using deep learning and an utrasonic bacon system with geo- Development, Canada Ottawa (Ontario), Tech. Rep., 2004.
tagging,” Computer-aided Civil and Infrastructure Engineering, vol. 33, [18] M. Ezuma, O. Ozdemir, C. Kumar, W. A. Gulzar, and I. Guvenc, “Micro-
pp. 885–902, 2018. UAV detection with a low-grazing angle millimeter wave radar,” in Proc.

11
 IEEE Radio Wireless Week (RWW) Conf., Orlando, FL, (to appear in [41] Z. Shi, M. Huang, C. Zhao, L. Huang, X. Du, and Y. Zhao, “Detection
2019). of LSSUAV using hash fingerprint based svdd,” in Proc. IEEE Int. Conf.
[19] P. Molchanov, R. I. Harmanny, J. J. de Wit, K. Egiazarian, and J. Astola, Commun. (ICC), Paris, France, May 2017, pp. 1–5.
“Classification of small UAVs and birds by micro-doppler signatures,” [42] P. E. Pace, Detecting and classifying low probability of intercept radar.
International Journal of Microwave and Wireless Technologies, vol. 6, Artech House, 2009.
no. 3-4, pp. 435–444, 2014. [43] O. A. Alyt, A. S. Omar, and A. Z. Elsherbeni, “Detection and local-
[20] B. Torvik, A. Knapskog, . Lie-Svendsen, K. E. Olsen, and H. D. ization of rf radar pulses in noise environments using wavelet packet
Griffiths, “Amplitude modulation on echoes from large birds,” in Proc. transform and higher order statistics,” Progress In Electromagnetics
11th European Radar Conference (EuRAD), Rome, Italy, Oct 2014, pp. Research, vol. 58, pp. 301–317, 2006.
177–180. [44] A. Ciancio, S. Pattem, A. Ortega, and B. Krishnamachari, “Energy-
[21] B. Torvik, K. E. Olsen, and H. D. Griffiths, “X-band measurements of efficient data representation and routing for wireless sensor networks
radar signatures of large sea birds,” in Proc. IEEE Int. Radar Conf., based on a distributed wavelet compression algorithm,” in Proc. In-
Lille, France, 2014, pp. 1–6. ternational conference on Information processing in sensor networks,
[22] S. Vishwakarma and S. S. Ram, “Classification of multiple targets based Nashville, Tennessee, 2006, pp. 309–316.
on disaggregation of micro-doppler signatures,” in Proc. Asia-Pacific [45] J. Wang, S. Tang, B. Yin, and X.-Y. Li, “Data gathering in wireless
Microwave Conference (APMC), New Delhi, India, Dec 2016, pp. 1–4. sensor networks through intelligent compressive sensing,” in Proc. IEEE
[23] G. Li, R. Zhang, M. Ritchie, and H. Griffiths, “Sparsity-based dynamic Int. Conf. Comput. Commun. (INFOCOM), Orlando, FL, 2012, pp. 603–
hand gesture recognition using micro-doppler signatures,” in Proc. IEEE 611.
Radar Conference (RadarConf), Seattle, WA, May 2017, pp. 0928–0931. [46] K. Hassan, I. Dayoub, W. Hamouda, and M. Berbineau, “Automatic
[24] İ. Güvenç, O. Ozdemir, Y. Yapici, H. Mehrpouyan, and D. Matolak, “De- modulation recognition using wavelet transform and neural networks in
tection, localization, and tracking of unauthorized UAS and jammers,” wireless systems,” EURASIP J. Adv. Sig. Processing, vol. 2010, p. 42,
in Proc. IEEE/AIAA Digital Avionics Syst. Conf. (DASC), St. Petersburg, 2010.
FL, 2017, pp. 1–10. [47] G. Strang, “Wavelet transforms versus fourier transforms,” Bulletin of
[25] L. S. C. Dos Santos, L. A. de Andrade, and A. M. Gama, “Analysis of the American Mathematical Society, vol. 28, no. 2, pp. 288–305, 1993.
radar cross section reduction of fighter aircraft by means of computer [48] S. Theodoridis and K. Koutroumbas, Pattern recognition. Academic
simulation,” Journal of Aerospace Technology and Management, vol. 6, Press, 2009.
no. 2, pp. 177–182, 2014. [49] J. Goldberger, G. E. Hinton, S. T. Roweis, and R. R. Salakhutdinov,
[26] C. J. Li and H. Ling, “An investigation on the radar signatures of small “Neighbourhood components analysis,” in Proc. Advances in Neural
consumer drones,” IEEE Antennas and Wireless Propagation Letters, Information Processing Systems, 2005, pp. 513–520.
vol. 16, pp. 649–652, 2017.
[27] Á. D. de Quevedo, F. I. Urzaiz, J. G. Menoyo, and A. A. López, “Drone
detection and RCS measurements with ubiquitous radar,” in Proc. IEEE
Int. Conf. on Radar, Brisbane, Queensland, Australia, 2018, pp. 1–6.
[28] R. Nakamura and H. Hadama, “Characteristics of ultra-wideband radar
echoes from a drone,” IEICE Communications Express, vol. 6, no. 9,
pp. 530–534, 2017.
[29] F. Gökçe, G. Üçoluk, E. Şahin, and S. Kalkan, “Vision-based detection
and distance estimation of micro unmanned aerial vehicles,” Sensors,
vol. 15, no. 9, pp. 23 805–23 846, 2015.
[30] C. Aker and S. Kalkan, “Using deep networks for drone detection,” in
Proc. IEEE Int. Conf. Adv. Video Sig. Based Surveillance (AVSS), Lecce,
Italy, 2017, pp. 1–6.
[31] J. Peng, C. Zheng, P. Lv, T. Cui, Y. Cheng, and L. Si, “Using images
rendered by PBRT to train faster R-CNN for UAV detection,” Computer
Science Research Notes, pp. 13–18, 2018.
[32] E. Unlu, E. Zenou, and N. Rivière, “Using shape descriptors for UAV
detection,” Electronic Imaging, vol. 2018, no. 9, pp. 1–5, 2018.
[33] Unlu, Eren and Zenou, Emmanuel and Rivière, Nicolas, “Generic
Fourier descriptors for autonomous UAV detection,” in Proc. 7th Int.
Conf. Pattern Recognition Appl. Methods (ICPRAM), Funchal, Madeira-
Portugal, 2018, pp. 550–554.
[34] L. Hauzenberger and E. Holmberg Ohlsson, “Drone detection using
audio analysis,” M.S. thesis, Department of Electrical and Information
Technology, Lund University, Sweden, June 2015.
[35] X. Chang, C. Yang, J. Wu, X. Shi, and Z. Shi, “A surveillance system
for drone localization and tracking using acoustic arrays,” in Proc.
IEEE Sensor Array and Multichannel Signal Process. Workshop (SAM),
Sheffield, UK, 2018, pp. 573–577.
[36] A. Bernardini, F. Mangiatordi, E. Pallotti, and L. Capodiferro, “Drone
detection by acoustic signature identification,” Electronic Imaging, vol.
2017, no. 10, pp. 60–64, 2017.
[37] S. Jeon, J. Shin, Y. Lee, W. Kim, Y. Kwon, and H. Yang, “Empirical
study of drone sound detection in real-life environment with deep neural
networks,” in Proc. 25th European Signal Process. Conf. (EUSIPCO),
Aug 2017, pp. 1858–1862.
[38] J. Busset, F. Perrodin, P. Wellig, B. Ott, K. Heutschi, T. Rühl, and
T. Nussbaumer, “Detection and tracking of drones using advanced
acoustic cameras,” in Proc. SPIE Unmanned/Unattended Sensors and
Sensor Networks XI; and Advanced Free-Space Optical Communication
Techniques and Applications, vol. 9647, 2015, pp. 1–8.
[39] H. Liu, Z. Wei, Y. Chen, J. Pan, L. Lin, and Y. Ren, “Drone detection
based on an audio-assisted camera array, Laguna Hills, CA,” in Proc.
IEEE Third Int. Conf. on Multimedia Big Data (BigMM), April 2017,
pp. 402–406.
[40] C. Zhao, M. Shi, Z. Cai, and C. Chen, “Detection of unmanned aerial
vehicle signal based on gaussian mixture model,” in Proc. 12th Int. Conf.
Comput. Sci. Edu. (ICCSE), Aug 2017, pp. 289–293.

12
View publication stats