Content uploaded by Beom-Seok Oh
Author content
All content in this area was uploaded by Beom-Seok Oh on Feb 20, 2018
Content may be subject to copyright.
ACCEPTED VERSION, IEEE GEOSCIENCE AND REMOTE SENSING LETTERS 1
Micro-Doppler mini-UAV Classification Using
Empirical Mode Decomposition Features
Beom-Seok Oh, Xin Guo, Fangyuan Wan, Kar-Ann Toh, Zhiping Lin∗
Abstract—In this letter, we propose an empirical mode decom-
position (EMD) based method for automatic multi-category mini-
UAV classification. The radar echo signal is firstly decomposed
into a set of oscillating waveforms by EMD. Then, eight statistical
and geometrical features are extracted from the oscillating
waveforms to capture the phenomenon of blade flashes. After
feature normalization and fusion, a nonlinear support vector
machine is trained for target class label prediction. Our empirical
results on real measurement of radar signals show encouraging
mini-UAV classification accuracy performance.
Index Terms—Micro-Doppler Signature, Empirical Mode De-
composition, Unmanned Aerial Vehicle Classification
I. INTRODUCTION
DURING the past decade, both the technology and the
industry of UAV have advanced greatly. The equipment
cost has been significantly lowered while the flying perfor-
mance of UAVs has been largely enhanced. This advancement
enables UAV to be deployed in a wide range of applications
(e.g., parcel delivery). On the other hand, UAVs may form a
threat to security-sensitive areas if they are used for spying,
reconnaissance and even for attack. This is particularly true
since existing mini-sized UAVs are difficult to detect due to
their small size, slow flying speed and low flying altitude. For
those security-sensitive areas, an accurate automatic detection
and classification of mini-UAVs is of particular interest.
The radar technology is widely adopted in surveillance sys-
tems because it has fast remote sensing capabilities regardless
of weather. Among the various signal processing techniques
for radar signals, the micro-Doppler signature (m-DS) [1] is
the most popular choice for mini-UAV classification [2]–[7].
This is because the m-DS extracted from radar echo signals
can capture the unique characteristics of micro-motion induced
by the mini-UAV propeller/rotor blades [5]–[7]. Table I shows
a summary of existing radar m-DS based state-of-the-art mini-
UAV classification methods.
In the two most recent works [6], [7], the authors have
reported very good classification performances. However, al-
though a pre-trained convolutional deep network model is
utilized in [6], collecting a large enough number of labeled
data for deep network tuning is costly. The method proposed
in [7] is for a binary-category problem only and it is not
B.-S. Oh, F. Wan and Z. Lin are with School of Electrical and Electronic
Engineering, Nanyang Technological University, 50 Nanyang Avenue, 639798
Singapore.
X. Guo is with Temasek Laboratories at Nanyang Technological University
(TL@NTU), 50 Nanyang Drive, 637553 Singapore.
K.-A. Toh is with School of Electrical and Electronic Engineering, Yonsei
University, 50 Yonsei-ro, Seodaemun-gu, Seoul, 03722, Republic of Korea.
The authors thank the Thales Nederland for providing the CW radar data.
∗Corresponding author: ezplin@ntu.edu.sg
TABLE I
EXIST ING WO RKS O N RADAR M-DS BASED MINI-UAV CLASSI FICATION
Authors Main objective Radar m-DS Classification
(year) (Num. of target class) analysis
Molchanov et mini-UAV model/type Spectrogram NBC1, linear &
al. (2014) [2] classification (5 & 11) nonlinear SVM2
Harmanny et mini-UAV type Spectrogram, Maximum a
al. (2015) [3] classification (5) Cepstrogram posteriori
Fioranelli et mini-UAV payloads Spectrogram NBC,
al. (2015) [4] classification (2) DDA3
Torvik et al. Quadcopters vs. birds Radar Nearest neigh-
(2016) [5] classification (4) polarimetry bor classifier
Kim et al. Quadcopter vs. Spectrogram + Convolutional
(2017) [6] hexacopter (2) CVD4neural network
Ren and Jiang mini-UAV vs. birds 2-D complex-log- Subspace reli-
(2017) [7] classification (2) Fourier transform ability analysis
Proposed mini-UAV model EMD Nonlinear
method classification (2 & 11) SVM
1NBC: naive Bayes classifier, 2SVM: support vector machine, 3DDA: diagonal-linear
variant of the discriminant analysis, 4CVD: cadence-velocity diagram.
straightforward to extend to the multi-category problem. Note
that classifying a mini-UAV into its type/model is an important
task as different mini-UAV may form a different level of threat.
Table I also shows that the majority of works in the literature
is relying on the short-time Fourier transform (STFT) analysis
such as spectrogram. We note that many Fourier analysis
based prior arts require signals with a longer dwell time to
capture the desired spectral information. Moreover, the time-
frequency resolution may pose a limitation for the STFT-based
techniques. A simple and yet effective resolution to address
these limitations is to analyze the m-DS using a non-Fourier
time-frequency analysis method. In this work, we propose to
deploy the EMD [8] for the m-DS analysis of mini-UAVs.
The core idea of EMD is to treat the signal as fast oscilla-
tions superimposed on slow oscillations. Essentially, the EMD
decomposes a real-valued univariate signal x∈Rdinto a set
of oscillating waves as: x=∑L
l=1ml+qL,where ml∈Rdindi-
cates the l-th intrinsic mode function (IMF) and qL∈Rdis a
residue [8]. Each of the obtained IMFs is dominated by certain
frequency components. Different from the STFT analysis, the
frequency band of each IMF is adaptively determined based
on the input signal. Another unique characteristic of the EMD
is that the signal is decomposed locally. Our idea behind using
the EMD for the mini-UAV radar m-DS analysis is to capture
the phenomenon of blade flashes locally and adaptively.
The EMD method has been successfully deployed in a
wide range of applications such as speech signal processing
[9], fault diagnosis [10], electroencephalography analysis [11],
m-DS analysis for classification of large aerial targets (e.g.,
jet-/ propeller-aircraft and helicopter) [12]–[15], and ground
moving vehicle [16]. However, none of the existing works
is on mini-UAV classification. While some EMD analysis
results on mini-UAV echo signals are presented in [17], their
ACCEPTED VERSION, IEEE GEOSCIENCE AND REMOTE SENSING LETTERS 2
analysis is limited to signals measured under indoor condition.
Moreover, no further study on EMD (e.g., feature extraction
and classification) is reported in [17].
Since the radar echo from mini-UAVs is generally much
weaker than that from the large manned aerial vehicles, the
existing EMD-based works [12]–[16] cannot be directly ap-
plied to the mini-UAV classification problem. In [14] and [16],
for example, a few statistical features (e.g., entropy, second
order moment etc.) were respectively extracted from the first
IMF and the sum of the remaining IMFs for classification of
large aerial vehicles and ground moving vehicles. However,
as shown in Figs. 1 and 2, the first IMF of mini-UAVs’
echo contains almost no target information. Our analysis (see
Section II for details) shows that the order of IMF in which
the blade flash responses appear varies over mini-UAV types.
Moreover, the order of blade flash IMF may vary within a
mini-UAV type according to the signal quality.
Motivated by the above observationson the EMD properties
and in view of the lack of study in mini-UAV classification, in
this letter, we propose an EMD based method for mini-UAV
classification. Essentially, a radar echo signal is decomposed
into a set of IMFs using EMD. Eight statistical and geometrical
features, which are designed for label discrimination, are then
respectively extracted from the IMFs. After normalization and
fusion, the extracted features are used to train a nonlinear
SVM classifier for target class label prediction. The main
contributions of this letter can be enumerated as follows:
•Proposal of an effective EMD-based method for mini-
UAV classification. Four novel geometrical features are
proposed and extracted together with four conventional
statistical features from a set of IMFs.
•Provision of an empirical analysis on radar mini-UAV m-
DS showing the relationships between the spectrogram
and the IMFs in terms of the blade flash phenomenon.
•Extraction of three sets of features from a complex-valued
signal using the real-valued EMD, and then fuse them for
mini-UAV classification performance enhancement.
II. MICRO-DOPPLER SIGNATURE ANALYSIS USING EMD
In this section, the effectiveness of using EMD for radar
m-DS analysis is empirically investigated. Particularly, the m-
DS induced by the micro-motion of three types of mini-UAVs
(e.g., helicopter, fixed-wing and quadcopter) is considered.
A physically measured continuous-wave (CW) radar echo
(see Section IV-A for the radar setup) is firstly analyzed
using the STFT which produces a spectrogram, and the EMD
which produces a set of IMFs. Both analysis results are then
compared to each other in terms of the blade flash phenomenon
[1]. The STFT is computed using a sliding Hamming window
of length 28with 90% overlapping.
A. Helicopter type mini-UAV
Fig. 1 shows two sets of a spectrogram and the first four
IMFs, obtained by analyzing the radar signals (of 50ms long)
of a helicopter mini-UAV. In Fig. 1 (a), the spectrogram
obtained using the radar echo of a helicopter mini-UAV with
strong m-DS is shown. In the subplot, the micro-Doppler
spectrum width and blade flashes are clearly visible [1]. Via
0 0.01 0.02 0.03 0.04 0.05
0
0.5
1
Norm. input
0 0.01 0.02 0.03 0.04 0.05
-0.2
0
0.2
IMF1
0 0.01 0.02 0.03 0.04 0.05
-0.2
0
0.2
IMF2
0 0.01 0.02 0.03 0.04 0.05
-0.2
0
0.2
IMF3
0 0.01 0.02 0.03 0.04 0.05
Time (s)
-0.2
0
0.2
IMF4
a b
cd
0 0.01 0.02 0.03 0.04 0.05
0
0.5
1
Norm. input
0 0.01 0.02 0.03 0.04 0.05
-0.2
0
0.2
IMF1
0 0.01 0.02 0.03 0.04 0.05
-0.2
0
0.2
IMF2
0 0.01 0.02 0.03 0.04 0.05
-0.2
0
0.2
IMF3
0 0.01 0.02 0.03 0.04 0.05
Time (s)
-0.2
0
0.2
IMF4
Micro-Doppler Signature (dB): Helicopter
0.005 0.01 0.015 0.02 0.025 0.03 0.035 0.04 0.045
Time (s)
-20
-15
-10
-5
0
5
10
15
20
Doppler (KHz)
-60
-55
-50
-45
-40
-35
-30
-25
-20
Micro-Doppler Signature (dB): Helicopter
0.005 0.01 0.015 0.02 0.025 0.03 0.035 0.04 0.045
Time (s)
-20
-15
-10
-5
0
5
10
15
20
Doppler (KHz)
-60
-55
-50
-45
-40
-35
-30
-25
-20
Fig. 1. Two sets of a spectrogram and the first four IMFs obtained using a
measurement radar signal ((a) and (c)) with strong m-DS and ((b) and (d))
with weak m-DS (due to different angle of viewing) of a helicopter mini-UAV.
a simple analysis on the blade flashes [18], several physical
mini-UAV parameters namely, the number of blades and its
length, blade tip velocity, rotation rate and blade symmetry,
can be retrieved from the subplot. It should be noted here
that, however, such information can be retrieved with a certain
accuracy only when the m-DS is strong enough. For example,
the spectrogram shown in Fig. 1 (b) is computed using a radar
signal of the same mini-UAV model, but with weak m-DS due
to different angle of viewing the mini-UAV(e.g., the rotor mast
of the helicopter mini-UAV was nearly parallel to the radar
line of sight). The mini-UAV parameters retrieved in Fig. 1
(b) might be totally different from those in Fig. 1 (a).
Different from the spectrogram, the IMFs do not possess
physical-parameter-specific information. The EMD instead
captures existence of blade flashes with time information. Fig.
1 (c) and (d) respectively show the first four IMFs obtained
from the two radar signals used to compute the spectrograms
shown in Fig. 1 (a) and (b). As shown in Fig. 1 (c), the first
IMF (IMF1) oscillates with an almost constant variance over
time. On the other hand, the local amplitude of IMF2 changes
significantly at certain time points. By synchronizing such
amplitude patterns with the spectrogram shown in Fig. 1 (a), it
is observed that those blade flashes captured by EMD appear
as high amplitude with a small number of zero-crossings
within a certain IMF (e.g., IMF2 of Fig. 1 (c)). Even for the
weak m-DS as shown in Fig. 1 (d), the IMF3 still contains
such blade flash patterns which are useful for classification.
B. Fixed-wing type and Quadcopter type mini-UAVs
We now consider two different types of mini-UAV, namely
a fixed-wing type (see Fig. 2 (a) and (c)) and a quadcopter
type (Fig. 2 (b) and (d)). Due to space constraint, only radar
signals with strong m-DS are shown here. Similar to those
IMFs shown in Fig. 1 (c), the blade flashes induced by these
two mini-UAV types appear at IMF4 (see Fig. 2 (c)) and
at IMF3 (see Fig. 2 (d)), respectively. The main reason for
observing such responses from different orders of IMF is
due to different characteristics of rotor blades. For example,
in general, the main rotor blades of a helicopter mini-UAV
are much longer and wider than that of a fixed-wing and
a quadcopter mini-UAV. Such physical differences in blades
ACCEPTED VERSION, IEEE GEOSCIENCE AND REMOTE SENSING LETTERS 3
a b
cd
0 0.01 0.02 0.03 0.04 0.05
0
0.5
1
Norm. input
0 0.01 0.02 0.03 0.04 0.05
-0.05
0
0.05
IMF1
0 0.01 0.02 0.03 0.04 0.05
-0.05
0
0.05
IMF2
0 0.01 0.02 0.03 0.04 0.05
-0.05
0
0.05
IMF3
0 0.01 0.02 0.03 0.04 0.05
Time (s)
-0.05
0
0.05
IMF4
0 0.01 0.02 0.03 0.04 0.05
0
0.5
1
Norm. input
0 0.01 0.02 0.03 0.04 0.05
-0.05
0
0.05
IMF1
0 0.01 0.02 0.03 0.04 0.05
-0.05
0
0.05
IMF2
0 0.01 0.02 0.03 0.04 0.05
-0.05
0
0.05
IMF3
0 0.01 0.02 0.03 0.04 0.05
Time (s)
-0.05
0
0.05
IMF4
Micro-Doppler Signature (dB): Plane
0.005 0.01 0.015 0.02 0.025 0.03 0.035 0.04 0.045
Time (s)
-20
-15
-10
-5
0
5
10
15
20
Doppler (KHz)
-60
-55
-50
-45
-40
-35
-30
-25
-20
Micro-Doppler Signature (dB): Quadcopter
0.005 0.01 0.015 0.02 0.025 0.03 0.035 0.04 0.045
Time (s)
-20
-15
-10
-5
0
5
10
15
20
Doppler (KHz)
-60
-55
-50
-45
-40
-35
-30
-25
-20
Fig. 2. Two sets of a spectrogram and the first four IMFs obtained using a
measurement radar signal of ((a) and (c)) a fixed-wing mini-UAV and ((b)
and (d)) a quadcopter mini-UAV.
affect the return signals, and thus micro-Doppler frequency.
This is an important property for mini-UAV classification.
III. PROPOSED METHOD FOR MINI-UAV CLASSIFICATION
We have empirically shown in Section II that the EMD can
effectively capture the unique patterns of m-DS induced by the
micro-motion of mini-UAV blades. To extract and utilize such
unique information for mini-UAV classification, in this section,
we propose to extract eight statistical and geometrical features
from the obtained IMFs. After normalization and fusion, the
obtained features are used to train a nonlinear SVM classifier.
A. Features extraction from IMFs and Normalization
Let si∈Cd,i=1,···,m,denotes the i-th univariate radar
signal, where mindicates the total number of training samples.
Since siis complex-valued, it is firstly converted to a real-
valued signal xi=|si| ∈ Rd,where |·| indicates the magnitude
function. The EMD decomposes xiinto Lnumber of IMFs
mi
land a residue qi
L. For simplicity, mi
lwill be denoted as ml
hereafter unless otherwise specified. The number of resulted
IMFs, L, varies according to the frequency contents contained
in the signal xi. In this work, only the first K<LIMFs are
used for extraction of the eight features listed in Table II.
Features 1 to 4are the conventional statistical features
adopted from the literature while Features 5 to 8are geomet-
rical features newly proposed in this work. Features 6 and 7
are designed to geometrically capture the blade flash responses
(see Fig. 1 (c) and (d)) within the IMFs while Features 5
and 8are for capturing geometrical relationships between two
consecutive IMFs. These new features are chosen because the
blade flash response and its geometrical relationship informa-
tion cannot be found in the literature.
The extracted eight feature vectors are subsequently con-
catenated: fi=fT
i,1,fT
i,2,···,fT
i,8T∈R8K−2,and then stacked
into a matrix form: FMag = [f1,f2,···,fm]∈R(8K−2)×m. For
the subsequent classification process, each row vector of FMag
is normalized to within the range of [0,1]using the min-max
normalization technique. Since the test data is not available
beforehand, its features can only be normalized with respect
to the min and max values of the training features.
B. Feature level fusion of the EMD features
It is observed from our preliminary results that the infor-
mation contained in the eight features may not be enough
to produce a good accuracy performance. Particularly, feature
vector fi∈R8K−2extracted at K=6 and from a 50ms
signal xi∈R4,800, is just of 46 dimension. To enhance the
classification accuracy performance, we propose to extract
two more sets of the eight features namely, FReal from
Real(si)and FImag from Imaginary(si), together with FMag
extracted from |si|. The fused feature is defined as follows:
FFusion = [(FMag)T,(FReal )T,(FImag)T]T∈R3(8K−2)×m.
C. Mini-UAV classification using the RBF-SVM classifier
The extracted feature matrix, FMag or FFusion, is then used
to train the SVM with a nonlinear radial basis function (RBF)
kernel. To handle the multi-categories, the SVM is trained
with the one-vs-all learning scheme. This nonlinear SVM
classifier is adopted due to its strong theoretical foundations
with guaranteed reliable classification performance. Due to
space constraint, we refer to [20] for more details on the SVM.
IV. EXPERIMENTS
The main goal of this study is to empirically verify the effec-
tiveness of the proposed method for mini-UAV classification.
To achieve this goal, a set of real radar signals returned from
11 objects, is utilized in this experimental study. The resulted
accuracy performance is then compared with that of competing
state-of-the-arts. The following subsections provide the details
of our dataset, experimental setup and results.
A. Dataset and Preprocessing
As shown in Table III, the dataset consists of eleven object
models including six commercial mini-UAVs. This dataset,
provided by Thales Nederland, was collected at outdoors
using an X-band CW radar (9.7GHz radio frequency, 192KHz
sampling rate). A horn antenna was manually adjusted towards
the nearby target object (see the fourth column of Table III
for the distance range between the radar and target object).
The radar signal is decimated by a factor of two to reduce
the sampling rate to 96KHz. The identified valid signals are
then divided into non-overlapping segments of {20, 30, ···,
100}ms long (a.k.a. dwell time) similar to that in [2]. This is
to observe the relationship among the dwell time, the accuracy
performance and the number of IMFs, K. In [2], Molchanov
et al.’s method, which was STFT-based, worked well only
for segments with longer dwell time (>100ms). Since our
proposed method is non-Fourier based, we are particularly
interested in investigating the shortest possible working range
for the dwell time. Through this range study, the relationship
between the minimum number of IMFs Kand the dwell time
can be explored. In field application, a proper value for the
dwell time can be found based on a cross-validation test using
the accumulated training data.
B. Evaluation of the proposed method (11-category problem)
1) Experimental setup: The main objective here is to clas-
sify an unseen test sample into one of the eleven categories
shown in Table III. The proposed method has two adjustable
ACCEPTED VERSION, IEEE GEOSCIENCE AND REMOTE SENSING LETTERS 4
TABLE II
THE EIG HT S TATIST ICAL (FEATURES 1TO 4) AN D GEOMETRICAL (FEATURES 5TO 8) FEATURES TO BE EXTRACTED FROM IMFS,mlWHERE l=1,··· ,K
No. Brief description Feature Details / Note
1 Number of zero crossing [19] fi,1= [Z1,··· ,ZK]T∈RKZl=∑d
n=2|sign[ml(n)] −sign [ml(n−1)]|,sign[ml(n)] = 1
if ml(n)≥0, otherwise 1. ml= [ml(1),··· ,ml(d)]T
2 Normalized signal energy [10], [16] fi,2= [E1/E,··· ,EK/E]T∈RKEl=∑d
n=1|ml(n)|2is the signal energy of ml,E=∑K
l=1El
3 Standard deviation [20] fi,3= [std (m1),··· ,std (mK)]T∈RKstd(m)returns the standard deviation of m
4 Entropy (see chapter 11 of [21]) fi,4= [En (m1),··· ,En(mK)]T∈RKEn(ml) = −∑P(ml)log2P(ml)and P(·)returns
the histogram counts of mlwith 256 bins
5 Entropy of two concatenated IMFs fi,5=En(m′
1),··· ,En m′
K−1T∈RK−1m′
l=mT
l,mT
l+1T: a concatenation of two consecutive IMFs
6 Normalized max-mean difference fi,6= [D1,··· ,DK]T∈RKDl=max
γ
(ml)−mean(ml>0)
max
γ
(ml),max
γ
(ml)returns the
γ
-th maxi-
mum value of ml, and mean(·)denotes the mean function.
The
γ
is used to remove exceptional outliers
7 Distance between two fi,7= [P1,··· ,PK]T∈RKPl=|argmax(Fl)−d/2|×2
d,Fl=∑∞
f=−∞|Ml(f)|2,and
frequency peaks Ml(f)indicates the Fourier transform of ml(n)
8 Ratio of zero-crossing number Zlfi,8= [Z2/Z1,··· ,ZK/ZK−1]T∈RK−1Refer to Feature 1 for details of Zl
TABLE III
SPECIFICATIONS OF THE R EAL MEASUREMENT DATASET
No. Target Target Distance range Num. segments
object name object type to the target†(50ms long)
1 Birds non-UAV 5m–50m 1,129
2 Yak54 Fixed-wing 20m–150m 1,480
3 EasyStar 780
4 T-REX450 1,759
5 Logo 400 Helicopter 20m–70m 1,141
6 Logo 600 717
7 Parrot 2.0 Quadcopter 3m–20m 3,534
8 / 9 1 / 2 rotors Stationary 5m, 10m, 15m, 298 / 336
10/11 3 / 4 rotors rotor 20m for each 240 / 600
†Since the CW radar cannot measure the target distance range, the distance range was
estimated based on the land size of the mini-UAV trial field, and based on those video
recordings which were acquired together with the radar data.
parameters namely, the number of IMFs Kused for feature
extraction, and the RBF kernel width
σ
for the adopted
nonlinear SVM classifier. Since the Kis closely related to the
dimension of the extracted features, the proposed method shall
be evaluated over various K∈ {3,···,7}settings. Different
from the K, an optimal value for
σ
=7 is chosen among
{0.1,1,5: 2 : 19,50,100}in terms of random 10 runs of 2 folds
cross-validation tests using only the training set. The stopping
criterion for computing each IMF is fixed at 10 iterations [22].
The test classification accuracy performance is computed
as The number of correctly matched test samples
Total number of test samples .For statistical evidence,
a test based on random 20 runs of 10 fold cross-validation
is performed, where the average is recorded. All experiments
were performed on a PC (3.2GHz, 8G RAM) under the Matlab
platform. The Matlab code for EMD is obtained from [22].
2) Results and discussions: Fig. 3 shows the test average
accuracy values of the proposed method obtained at K∈
{3,···,7}, which are plotted over different dwell times. To
position the proposed method among those competing mini-
UAV classification methods, its test accuracy is compared with
that of Molchanov et al.’s work [2]. Since the classification
results reported in [2] were obtained using the same dataset
except for certain protocol details (e.g., sampling rate, signal
normalization and cropping), the directly taken accuracy val-
ues at {50ms,75ms,100ms}from [2] are shown in the plot.
Fig. 3 shows that the proposed method before fusion
(denoted as ‘Proposed (Single)’) outperforms the compared
Molchanov et al.’s method when K∈ {6,7}IMFs are utilized
20 30 40 50 60 70 80 90 100
Dwell time, ms
0.75
0.8
0.85
0.9
Test Classification Rate (%)
Average accuracy over different dwell times, 11 class
Molchanov et al. [2]
Proposed (Single, K=3)
Proposed (Single, K=4)
Proposed (Single, K=5)
Proposed (Single, K=6)
Proposed (Single, K=7)
Proposed (Fusion, K=3)
Proposed (Fusion, K=4)
Proposed (Fusion, K=5)
Proposed (Fusion, K=6)
75ms
Fig. 3. Performances of the proposed method at K∈ {3,··· ,7}and at
σ
=7.
The extractable Kvaries according to the frequency contents contained in the
signal (e.g., 7 IMFs can be extracted only when the dwell time ≥80ms).
TABLE IV
CONFU SION MATRIX (%) OF THE ‘PRO POSED (FUSI ON,K=6 AT 70MS)’
Estimated class label
1 2 3 4 5 6 7 8 9 10 11
199 0.4 0 0 0 0 0.6 0 0 0 0
2 1.1 86.3 5.5 0.1 0.1 0 6.9 0 0 0 0
3 1.9 31.4 58.1 0 0.1 0 8.5 0 0 0 0
T 4 0.1 1.9 0.1 96.1 1.2 0.1 0.5 0 0 0 0
r 5 0.1 0.4 0.2 1.4 97.9 0 0 0 0 0 0
u 6 0.6 0 0 0.2 0.1 99.1 00000
t 7 0.7 0.1 0.3 0 0 0 98.9 0 0 0 0
h 8 1.4 0 0 0 0 0 0 95.6 3.0 0 0
9 0.9 0 0 0 0 0 1 3.3 94.6 0 0.2
10 1.8 0 0 0 0 0 2.5 0.9 6.7 66.2 21.9
11 0.7 0 0 0.4 0 0 0 0 1.3 4.4 93.2
for feature extraction. Interestingly, the performance gap is
observed to be larger when the dwell time is shorter (e.g.,
50ms). After the feature level fusion (‘Proposed (Fusion)’), the
proposed method for K>3 produces higher accuracies than
those of Molchanov et al.’s method. From the confusion matrix
of the ‘Proposed (Fusion, K=6 at 70ms)’ as shown in Table
IV, it is observed that most of the errors occur from categories
3 and 10. About 7–8.5% of the fixed-wing echos (categories 2
and 3) are also wrongly classified as a quadcopter (category 7).
Particularly, when the radar echo signals contain weak blade
flashs and/or similar blade flash patterns over categories (e.g.,
2 and 3), our method produces incorrect classification since
the eight features are only sensitive to the blade flash patterns
from the IMFs.
Fig. 4 shows the average test classification accuracies
ACCEPTED VERSION, IEEE GEOSCIENCE AND REMOTE SENSING LETTERS 5
Features 1-4
Features 1-4,5
Features 1-4,6
Features 1-4,7
Features 1-4,8
Features 1-4,6,7
Features 1-4,5,8
Features 1-4,5,6,7,8
0.8
0.85
0.9
Classification
rate (%)
Proposed (Single)
Proposed (Fusion)
Fig. 4. The impact of the proposed features on the classification accuracy
performance (%)at a dwell time of 50ms and K=6. For example, the
‘Features 1–4,6’ means five features (Features 1, 2, 3, 4 and 6) are utilized
TABLE V
COMPARIS ON BET WEEN THE PRO POSED MET HOD A ND COMPE TING
STATE-OF-THE-ARTS. AVERAGE TEST RESULTS TAK EN OVER RANDOM 20
RUNS OF 2FOLD CROSS -VALIDATION TEST S US ING 50MS SIGNAL
Method EER FARFRR=1%
Spectrogram+PCA†7.75 27.00
CVD+PCA†6.68 28.41
Cepstrogram+PCA†10.17 48.07
2D complex-log-Fourier transform+PCA†[7] 3.98 4.50
2D complex-log-Fourier transform+SRA†[7] 3.27 3.89
Proposed (Single, K=4) 3.54 4.78
Proposed (Fusion, K=4) 3.43 4.67
†The error rates are directly taken from [7]. Note that all the error rates
listed in this table were obtained using the identical data and protocols.
obtained over different combinations of the features. It is
observed from Fig. 4 that each of the proposed features
contains uncorrelated and discriminative information. Due to
the inclusion of more blade-flash-related information from
Features 5–8, our method achieves the best accuracy perfor-
mance, which is about 4–5% higher than that of using only
the four statistical features (‘Features 1–4’).
C. Comparison with state-of-the-arts (2-category problem)
As shown in Table V, the proposed method is now compared
with competing state-of-the-arts (e.g., the 2D complex-log-
Fourier transform [7] and the subspace reliability analysis
(SRA) [7]) under a binary classification problem. Note that
the dataset utilized in [7] is exactly the same with that of
ours. Following the experimental protocols reported in [7], the
equal error rate (EER) and the false acceptance rate when the
false rejection rate equals to 1% (FARFRR=1%) of the proposed
method are obtained (see Table V). These error rates are then
compared with those taken directly from [7]. Different from
the results shown in Fig. 3, the proposed method achieves
the best error rates at K=4, which appear to be similar to
(or slightly worse than) those of the 2D complex-log-Fourier
based methods [7]. However, as aforementioned, the methods
proposed in [7] were for a binary-category problem only and it
is not clear how to extend to the multi-category problem. The
table also shows that our method significantly outperforms the
spectrogram, CVD and cepstrogram based methods.
V. CONCLUSION
In this letter, an EMD based method was proposed for mini-
UAV classification. The radar echo signal was decomposed
into a set of IMFs for m-DS estimation. From our empirical
analysis, it was observed that blade flashes appeared as sharp
changes towards high magnitudes with fewer number of zero-
crossings than those time segments without such blade flashes.
Moreover, the order of IMF’s, in which such responses appear,
was observed to be different across mini-UAV types. To extract
such information, eight statistical and geometrical features
were extracted from the first few IMFs. After normalization
and fusion, the extracted features were classified using the
SVM classifier. Our empirical results on physical radar sig-
nals showed comparable or better accuracy performance than
competing state-of-the-arts for mini-UAV classification.
REFERENCES
[1] V. C. Chen, The micro-Doppler Effect in radar. Artech House, 2011.
[2] P. Molchanov, R. I. A. Harmanny, J. J. M. de Wit, K. Egiazarian, and
J. Astola, “Classification of small UAVs and birds by micro-Doppler
signatures,” Int. J. Microw. Wirel. T., vol. 6, no. 3-4, pp. 435–444, 2014.
[3] R. I. A. Harmanny, J. J. M. de Wit, and G. Premel-Cabic, “Radar micro-
Doppler mini-UAV classification using spectrograms and cepstrograms,”
Int. J. Microw. Wirel. T., vol. 7, no. 3-4, pp. 469–477, 2015.
[4] F. Fioranelli, M. Ritchie, H. Griffiths, and H. Borrion, “Classification of
loaded/unloaded micro-drones using multistatic radar,” Electron. Lett.,
vol. 51, no. 22, pp. 1813–1815, 2015.
[5] B. Torvik, K. E. Olsen, and H. Griffiths, “Classification of Birds and
UAVs Based on Radar Polarimetry,” IEEE Geosci. Remote Sens. Lett.,
vol. 13, no. 9, pp. 1305–1309, 2016.
[6] B. K. Kim, H.-S. Kang, and S.-O. Park, “Drone classification using
convolutional neural networks with merged doppler images,” IEEE
Geosci. Remote Sens. Lett., vol. 14, no. 1, pp. 38–42, 2017.
[7] J. Ren and X. Jiang, “Regularized 2-d complex-log spectral analysis
and subspace reliability analysis of micro-doppler signature for uav
detection,” Pattern Recogn., vol. 69, pp. 225–237, 2017.
[8] N. E. Huang, Z. Shen, S. R. Long, M. C. Wu, H. H. Shih, Q. Zheng, N.-
C. Yen, C. C. Tung, and H. H. Liu, “The empirical mode decomposition
and the Hilbert spectrum for nonlinear and non-stationary time series
analysis,” in Proc. Roy. Soc. London Ser. A, vol. 454, no. 1971, 1998,
pp. 903–995.
[9] A. Bouzid and N. Ellouze, “Empirical Mode Decomposition of Voiced
Speech Signal,” in Proc. Int. Symp. Control, Commun. Sig. Proc. IEEE,
2004, pp. 603–606.
[10] Y. Yu, Y. Dejie, and C. Junsheng, “A roller bearing fault diagnosis
method based on EMD energy entropy and ANN,” J. sound vib., vol.
294, no. 1, pp. 269–277, 2006.
[11] F. Riaz, A. Hassan, S. Rehman, I. K. Niazi, and K. Dremstrup, “EMD-
Based Temporal and Spectral Features for the Classification of EEG
Signals Using Supervised Learning,” IEEE Trans. Neural Syst. Rehabil.
Eng., vol. 24, no. 1, pp. 28–35, 2016.
[12] X. Bai, M. Xing, F. Zhou, G. Lu, and Z. Bao, “Imaging of Micromotion
Targets With Rotating Parts Based on Empirical-Mode Decomposition,”
IEEE Trans. Geosci. Remote Sens., vol. 46, no. 11, pp. 3514–3523, 2008.
[13] C. Cai, W. Liu, J. S. Fu, and Y. Lu, “Radar Micro-Doppler Signature
Analysis with HHT,” IEEE Trans. Aerosp. Electron. Syst., vol. 2, no. 46,
pp. 929–938, 2010.
[14] L. Du, B. Wang, Y. Li, and Hongwei, “Robust Classification Scheme for
Airplane Targets With Low Resolution Radar Based on EMD-CLEAN
Feature Extraction Method,” IEEE Sensors J., vol. 13, no. 12, pp. 4648–
4662, 2013.
[15] Y. Wang, X. Wu, W. Li, Z. Li, Y. Zhang, and J. Zhou, “Analysis of
micro-Doppler signatures of vibration targets using EMD and SPWVD,”
Neurocomputing, vol. 171, pp. 48–56, 2016.
[16] Y. Li, L. Du, and H. Liu, “Hierarchical Classification of Moving Vehicles
Based on Empirical Mode Decomposition of Micro-Doppler Signatures,”
IEEE Trans. Geosci. Remote Sens., vol. 51, no. 5, pp. 3001–3013, 2013.
[17] A. Brewster and A. Balleri, “Extraction and analysis of micro-Doppler
signatures by the Empirical Mode Decomposition,” in Proc. IEEE Radar
Conf., 2015, pp. 0947–0951.
[18] D. Tahmoush, “Detection of small UAV helicopters using micro-
Doppler,” in Proc. SPIE Defen. Secur. International Society for Optics
and Photonics, 2014, pp. 907717–907 717.
[19] S. Baykut and T. Akg¨ul, “Multiscale zero-crossing statistics of intrinsic
mode functions for white Gaussian noise,” in Proc. European Sig. Proc.
Conf., 2010, pp. 135–138.
[20] R. O. Duda, P. E. Hart, and D. G. Stork, Pattern classification. John
Wiley & Sons, 2012.
[21] R. C. Gonzalez, R. E. Woods, and S. L. Eddins, Digital Image Process-
ing Using MATLAB. Third New Jersey: Prentice Hall, 2003.
[22] P. Flandrin, “Empirical Mode Decomposition,” http://perso.ens-
lyon.fr/patrick.flandrin/emd.html, [Online; accessed 2-October-2016].