64 ACES JOURNAL, Vol. 35, No.

1, January 2020

Micro-motion Forms Classification of Space Cone-shaped Target Based on

Convolution Neural Network

Gaogui Xu 1,2, Hongcheng Yin 1,2, and Chunzhu Dong 2

1
School of Information Engineering
Communication University of China, Beijing 100024, China
[email protected]
2
Science and Technology on Electromagnetic Scattering Laboratory
Beijing 100854, China

Abstract ─ In this paper, the echo models with different besides the translation along the line of sight (LOS)
micro-motion forms (spin, tumbling, precession, and of the radar, the warhead will have spin, precession,
nutation) of space cone-shaped target are built. Different and nutation. Many researches indicated that the micro-
from the ideal point scatterers model, the radar echo Doppler signatures can be used to recognize and
contains the contribution from the complex radar cross discriminate the warhead and the decoys [4-6]. The
section (RCS) of point scatterer vs aspect angle. And a micro-motion parameters of the ballistic target reflect the
convolution neural network (CNN) model for micro- important characteristics such as its structure, size, and
motion forms classification based on the micro-Doppler micro-motion period. For radar imaging, to obtain a clear
characteristics in spectrograms is presented. The body image, the micro-Doppler component must be
simulation results show that our method can discriminate separated [7-9]. The premise of the success of these
different micro-motion forms effectively and the overall methods is to extract stable and effective micro-motion
accuracy is 97.24%. Different levels of additive white signatures.
Gaussian noise are added to simulate noise-contaminated Conventionally, the micro-motion parameters
radar echo. It has been found that the presented method estimation methods are based on the micro-motion form
has a stronger anti-noise ability than support vector [10, 11]. However, it's hard to know the micro-motion
machine (SVM). When the Signal-to-Noise Ratio (SNR) form of the target, since radar usually observes non-
of Gaussian white noise is 10 dB, the overall accuracy of cooperative targets. And the traditional classification
our algorithm is 29.79% higher than that of SVM. methods of micro-motion forms can be regarded as a
kind of feature engineering, which means that they need
Index Terms ─ Convolution neural network, micro- to extract discriminative features from the raw echo data
Doppler, micro-motion forms classification, space cone- for classification algorithms.
shaped target. In general, the domain knowledge of micro-Doppler
signals is required either in the design of time-varying
I. INTRODUCTION signatures or from the extraction of features. These
The micro-Doppler effect refers to Doppler limitations result in poor adaptability of the algorithm
modulations on the radar echo due to micro-motion and low anti-noise capability. In this paper, we propose
dynamics (such as mechanical vibrations or rotations) of a deep convolution neural network (CNN) model, which
the target or any structure on the target, in addition to the is based on the classical CNN devised by LeCun et al
constant Doppler frequency shift induced by the bulk [12]. It can effectively discriminate the micro-motion
motion of a radar target [1, 2]. Micro-Doppler is a unique forms from the radar echo, which makes the application
signature of the target with movements and provides of the parameter estimation methods more successful.
additional target features that are complementary to CNN is an end-to-end learning that automatically learns
existing methods [2]. The micro-Doppler effect can be features and discriminates patterns from data. Compared
used to classify and identify the target, for example, with the traditional method, CNN takes feature design,
human detection and activity classification [3] and the feature selection, feature extraction, and pattern
warhead and decoys of ballistic target [4]. recognition as a whole by optimizing a certain loss
Due to the interference by releasing decoys and the function. Through automatic learning features, it has
wobble under the action of the Earth's atmosphere, achieved state-of-the-art results in many fields, such as

Submitted On: July 24, 2019

Accepted On: December 28, 2019 1054-4887 © ACES
XU, YIN, DONG: MICRO-MOTION FORMS CLASSIFICATION OF SPACE CONE-SHAPED 65

image classification [13, 14] and speech recognition high-frequency electromagnetic waves, the scatterers of
[15, 16]. Thus, the influence of the preorder algorithm a cone-shaped target are mainly composed of cone
module on the subsequent algorithm module is reduced. top 𝑃0 and two cone bottoms 𝑃1 and 𝑃2 . If the radar
This paper is organized as follows: in Section II, we transmits a sinusoidal waveform with a carrier frequency
develop the echo models of space cone-shaped target 𝑓0 , then the baseband of the returned signal from the
with four different micro-motion forms, and the details cone-shaped target is:
of the classification algorithm based on CNN are 2
 2 Ri 
described. Next, in Section III, the simulation results are s  t     i exp{ ji }exp  j 2 f 0 
i 0  c 
(1)
given, followed by the conclusion in Section IV. 2
 2 Ri 
=  RCSi exp  j 2 f 0 ,
II. MICRO-MOTION FORMS i 0  c 
CLASSIFICATION where 𝑐 is the speed of the electromagnetic wave
A. Mathematical model propagation, 𝑅𝑖 is the range from the radar to the
With the development of modern penetration scatterer 𝑃𝑖 , and 𝑅𝐶𝑆𝑖 = √𝜎𝑖 𝑒𝑥𝑝{𝑗𝜑𝑖 } is the complex
technology, cone-shaped decoys have appeared. Since radar cross section (RCS) of the scatterer 𝑃𝑖 , which will
the attitude control of the decoy is weak, the angle of vary in different attitudes, as shown in the Appendix.
precession and nutation will be larger, and even the It is worth noting that the location of the scatterer
decoy will tumble. Therefore, the echo models of 𝑃𝑖 (𝑖 = 0, 1, 2) is not fixed, and it will slide on the surface
four different forms of micro-motion (spin, tumbling, of the target. For example, scatterer 𝑃0 is a moving
precession, and nutation) are given. scatterer that slides on the surface of the spherical crown
Figure 1 gives the geometry of radar and cone- with the change of incident direction. The location of 𝑃0
shaped target with spin, tumbling, precession, and is the point at which the incident ray passes through the
nutation. Three coordinate systems are introduced. The center of the sphere and intersects with the sphere.
radar is stationary and located at the origin of the radar Suppose a scatterer 𝑃 is located at ⃗⃗⃗ 𝑟0 =
coordinate system (𝑈, 𝑉, 𝑊) . The cone-shaped target (𝑥0 , 𝑦0 , 𝑧0 )𝑇 represented in the target local coordinate
is located at the local coordinate system (𝑥, 𝑦, 𝑧) and system (𝑥, 𝑦, 𝑧) at instant of time 𝑡 = 0 , where the
spinning around its symmetric axis ⃗⃗⃗⃗⃗⃗
𝑂′ 𝑧 . The origin of superscript T means transposition. Generally, the local
the local coordinate system is at the center of mass of coordinate system and the reference coordinate system
the target. The reference coordinate system (𝑢, 𝑣, 𝑤) is do not coincide at the initial time. For the reference
parallel to the radar coordinate system and shares the coordinate system, the position of point 𝑃 at time 𝑡 = 0
origin with the local coordinate system. The origin 𝑂′ is can be calumniated as 𝑅𝑖𝑛𝑖𝑡 ⋅ ⃗⃗⃗ 𝑟0 , where 𝑅𝑖𝑛𝑖𝑡 is an initial
assumed to be at a distance 𝑅0 from the radar. rotation matrix determined by Euler angles. After
rotation, the point 𝑃 in the radar coordinate system at
w z A time 𝑡 = 0 is located at 𝑅 ⃗⃗⃗⃗0 + 𝑅𝑖𝑛𝑖𝑡 ⋅ ⃗⃗⃗
𝑟0 .
Since the space cone-shaped target is rotationally
symmetrical, its spin does not affect radar echo. Thus,
the echo is obtained as:
P v
O
y
s1 (t )   RCS P exp  j
P
 4
 0
  
R0  Rinit  r0 , (2)

P
W R0 where 𝜆0 is the wavelength and ‖∙‖ represents the
Euclidean norm.
x u In the process of unfolding, the force of decoy is
more complex, which will cause its tumbling motion to
 V
O be irregular. For simplicity, suppose that the target will

tumble around the center of mass about its axes 𝑥, 𝑦, 𝑧
with an angular velocity 𝜔 ⃗ = (𝜔𝑥 , 𝜔𝑦 , 𝜔𝑧 )𝑇 . Therefore,
U
the unit vector of 𝜔 ⃗ is 𝜔 ⃗ ′ = (𝜔𝑥′ , 𝜔𝑦′ , 𝜔𝑧′ )𝑇 = 𝜔
⃗ /‖𝜔⃗ ‖.
According to Rodrigues formula [17], at time 𝑡 the
Fig. 1. Geometry of radar and cone-shaped target with
rotation matrix becomes:
spin, tumbling, precession, and nutation.
Rt  I  ˆ  sin t  ˆ 2 (1  cos t ), (3)
′
For simplicity, a target can be represented as a set of where Ω = ‖𝜔 ⃗‖, 𝜔 ̂ is a skew symmetric matrix
point scatterers. In the interest of brevity, the scatterer in constructed by 𝜔 ⃗ ′.
the following text refers to the point scatterer. Under After tumbling, the echo can be expressed as:
66 ACES JOURNAL, Vol. 35, No. 1, January 2020

 4
   extract time-varying signatures for micro-motion forms
s2 (t )   RCS P exp  j R0  Rt  Rinit  r0 . (4) classification. In this paper, the Wigner-Ville distribution
P  0  (WVD) [18] is used to analyze the micro-Doppler.
Precession refers to the coning motion of a target Different micro-motion forms present their
along the axis ⃗⃗⃗⃗⃗⃗𝑂′ 𝐴 , in addition to spinning around its unique micro-Doppler signatures. The micro-Doppler
modulation caused by different micro-motion forms can
symmetry axis ⃗⃗⃗⃗⃗⃗ 𝑂′ 𝑧 , as shown in Fig. 1. The angle
be seen in the high-resolution spectrogram. By using
between axis ⃗⃗⃗⃗⃗⃗ 𝑂′ 𝐴 and axis ⃗⃗⃗⃗⃗⃗
𝑂′ 𝑧 is called the precession spectrograms, we convert the micro-motion forms
angle. Spinning and coning are both rotating motions, classification problem to an image recognition problem.
except that the axis of rotation is different. Therefore, Figure 2 shows the hierarchical structure of the CNN,
Precession can be regarded as the spin of the target in the which stacks multiple layers of simple neural network
local coordinate system and then the coning motion in architecture, thus extracting the representation of data
the reference coordinate system. At time 𝑡, the location layer by layer. It takes color spectrograms as the input
of the scatterer 𝑃 in the reference coordinate system is and learns the mapping between the spectrogram and
𝑅𝑐 ⋅ 𝑅𝑖𝑛𝑖𝑡 ⋅ 𝑅𝑠 ⋅ ⃗⃗⃗𝑟0 , where 𝑅𝑐 is the rotation matrix of its corresponding micro-motion form. CNN is mainly
coning motion and 𝑅𝑠 is the rotation matrix of spinning composed of four components. The first component is
motion, both of which are determined by equation (3). the convolution layer which consists of a set of learnable
Thus, the echo of micro-Doppler modulations induced filters. Each filter has small receptive fields of input data
by precession is: and can be seen as a specific feature extractor. However,
 4
   in convolution operation, if there is no nonlinear
s3 (t )   RCS P exp  j R0  Rc  Rinit  Rs  r0 . (5) transformation in the network, the network will only
P  0  learn the linear characteristics of the input, resulting in
If the precession angle does not remain at a constant, the network can’t deal with nonlinear problems, and
the target will oscillate up and down between two limits. the application scene is limited. The second component
This motion is called nutation. After precession, the is the nonlinear activation function that carries on the
transient position vector of symmetry axis ⃗⃗⃗⃗⃗⃗ 𝑂′ 𝑧 is ⃗⃗⃗ 𝑧𝑡 = nonlinear transformation to the result of the linear
𝑇 transformation. The commonly used activation functions
𝑅𝑐 ⋅ 𝑅𝑖𝑛𝑖𝑡 ⋅ [0,0,1] . The target will then oscillates in the
plane composed of vector ⃗⃗⃗ 𝑧𝑡 and unit vector 𝑥̂ = ⃗⃗⃗⃗⃗⃗ 𝑂′𝐴 . are the Sigmoid function, the Restricted Linear Units
⃗⃗⃗⃗⃗⃗ (ReLU) function, and their variants. In this paper, we
It is equivalent to rotating about the axis 𝑧̂ = 𝑂′𝐴 × ⃗⃗⃗ 𝑧𝑡 ,
choose ReLU as the activation function, the formula of
where × represents cross multiplication.
which is 𝑓(𝑥) = 𝑚𝑎𝑥( 0, 𝑥). The third component is the
Therefore, 𝑥′ = 𝑥̂ , 𝑦′ = 𝑧̂ × 𝑥̂/‖𝑧̂ × 𝑥̂‖ , and 𝑧′ =
pooling layer, which subsamples the data according to
𝑧̂ /‖𝑧̂‖ form a new coordinate system. The transition
certain rules while retaining the main information. By
matrix from the reference coordinate system (𝑋, 𝑌, 𝑍)
performing 2 × 2 pooling operation with max pooling,
to the new coordinate system is 𝐴 = (𝑥′, 𝑦′, 𝑧′) . The the data size will be halved. Thus, The pooling layer
rotation matrix of the target rotating around the 𝑧′ axis is greatly reduces the network parameters and prevents
𝐵, which is determined by oscillating frequency 𝑓𝑜 and overfitting.
oscillating amplitude 𝜃𝑜 . Then, the echo is obtained as
[2]:
Conv. 128@5×5/ReLU

s4 (t )   RCS P exp  j
P
 4
 0
 
R0  Rv  Rc  Rinit  Rs  r0 ,

 Conv. 128@5×5/ReLU
Max pool 2×2
(6) Max pool 2×2
Conv. 64@3×3/ReLU
where 𝑅𝑣 = 𝐴 ⋅ 𝐵 ⋅ 𝐴𝑇 .
Max pool 2×2 Conv. 256@5×5/ReLU
B. Micro-motion forms classification based on
convolution neural network Conv. 32@3×3/ReLU Max pool 2×2
As shown in equations (2), (4), (5), and (6), the
echoes of the target with different micro-motion forms Conv. 32@3×3/ReLU FC 1024
are very different. In frequency domain, the common
method for analyzing echo signal is Fourier transform.
Spectrograms Label distribution
Due to lack of localized time information, it is not
suitable for analyzing a signal whose frequency changes
over time. Compared with Fourier transform, a joint Fig. 2. The hierarchical structure of the CNN.
time-frequency analysis can provide localized time-
dependent frequency information, which is used to The last component is the full connection layer
XU, YIN, DONG: MICRO-MOTION FORMS CLASSIFICATION OF SPACE CONE-SHAPED 67

(FCN), where all the input nodes are connected to all the 8𝜋 rad/s with an interval of 𝜋/4. Compared with spin,
output nodes. In the classification problem, Softmax precession has rotation around the coning axis, and the
function is typically used as an activation function for the frequency of rotation is from 2 Hz to 5 Hz with an
output layer to obtain the probability of the category and interval of 0.5 Hz. Nutation is a periodic transformation
can be expressed as: of the precession angle based on precession. Assume
L that the wobble frequency is from 5 Hz to 7 Hz with an
ai  e zi /  e j ,
z
(7) interval of 0.5 Hz and that the oscillating amplitude
j 1
ranges from 1° to 5° with an interval of 1°.
where 𝐿 represents the total number of target categories.
Suppose that the radar pulse repetition frequency
The category of the target is determined by the neuron
(PRF) is 2000. The spectrograms are obtained by WVD
with the highest probability, which is 𝑎𝑟𝑔𝑚𝑎𝑥( 𝑎).
analysis of 2s echo data. Since the RCS is related to
Suppose the difference between the network output
polarization, the horizontal to horizontal (HH) and
and the real value is 𝑙𝑜𝑠𝑠(𝑥), which is called the error
vertical to vertical (VV) polarization are considered. If
function. The training of CNN can be divided into
the simulation is carried out according to the above
two processes: forward propagation and backward
conditions, the samples of each micro-motion form
propagation. In the forward propagation, the output of
are uneven. Combined with the limited information
the network is calculated according to the input image.
contained in the spin, for precession, we randomly select
And in the backward propagation, the network parameters
21 frequencies from the spin frequencies, resulting in
are automatically updated by the error function. This
66,150 data. For nutation, we randomly select one
parameter update method is called gradient descent
frequency from the spin frequencies, resulting in 78,750
(GD). It is an optimization method to find a local
data. The size of the spectrogram is normalized to 256 ×
(preferably global) minimum of a function. For the
256.
network parameter 𝑤𝑖𝑗 , it is updated using:
In order to further approach the actual case, we
wij 1  wij  wij , (8) consider not only the contribution from the complex
where 𝜏 is the gradient descent iteration and 𝛥𝑤𝑖𝑗𝜏 is the RCS of point scatterer vs aspect angle, but also the
change of the scatterer's location with target motion.
weight update.
They will have an impact on micro-Doppler modulation,
𝛥𝑤𝑖𝑗𝜏 is determined by:
making it difficult for the feature-based classification
loss ( x)
wij   , (9) algorithms to discriminate different micro-motion forms.
wij Figure 3 shows an example of spectrograms of a
where 𝜂 is the learning rate and 𝑥 is the input cone-shaped target with precession for HH and VV
spectrogram. polarization. There are two visible scatterers in the time
According to the number of input samples used for of 2 seconds, as shown in Fig. 3 (a). The curves of
each update of the weights, two variants of GD are scatterer 𝑃0 and scatterer 𝑃1 are sinusoidal curve. The
generally used, namely on-line learning and batch RCS of scatterer 𝑃1 is smaller for VV polarization than
learning. In on-line learning, the weights are updated for for HH polarization, resulting in discontinuity of its
every input. In batch learning, the error function is curve, as shown in Fig. 3 (b). Different polarization only
calculated for a batch of input samples. If the selection affects the continuity of the curve and does not change
of the input samples is random, the GD optimization the micro-Doppler characteristics of the target. Among
method is called stochastic gradient descent (SGD). In the 176,400 data, 80% of spectrograms of each micro-
this paper, the SGD is used and the batch size is 256. motion form are used as the training set and the rest as
the test set. The convolution neural network shown in
III. SIMULATION RESULT Fig. 2 is constructed.
The echoes of four micro-motion forms are
simulated and WVD is used to obtain the spectrograms.
Assume that the operating frequency of radar is 10 GHz
and the center of mass of the target is located at
(𝑈 = 1000𝑚, 𝑉 = 5000𝑚, 𝑊 = 5000𝑚). The azimuth
angle 𝛼𝑃 and elevation angle 𝛽𝑃 of the coning axis ⃗⃗⃗⃗⃗⃗
𝑂′ 𝐴
range from 10° to 80° with an interval of 5°. Thus, there
are 225 combinations of (𝛼𝑃 , 𝛽𝑃 ) pairs. The initial
Euler angles are (30°, 30°, 45°). For spin, the spinning
frequency is from 1 Hz to 20 Hz with an interval of
0.5 Hz. Assume that the range of each component of the
angular velocity of rotation 𝜔 ⃗ is from 𝜋/2 rad/s to (a)
68 ACES JOURNAL, Vol. 35, No. 1, January 2020

number of spectrograms to extract the common features

in the same class and ignore the differences. It
automatically extracts the features related to the
classification tasks, so as to achieve higher accuracy than
the traditional algorithm.
With the increase of training epochs, the error on the
training set becomes smaller and smaller, while the error
on the test set decreases first and then increases. At this
time, the network is overfitted. After 200 epochs, the
parameters of the network are approximately optimal,
and the overall accuracy is 97.24%. The confusion
matrix for the micro-motion forms classification based
(b)
on CNN is listed in Table 1. Each column in the table is
the predicted label, each row is the true label, and the
Fig. 3. Micro-Doppler modulations of cone-shaped
diagonal element is the number correctly classified.
target induced by precession for (a) HH and (b) VV
The last column is the classification accuracy of each
polarization.
category. In order to measure the overall performance of
the algorithm in multiple classification problems, the
Different micro-motion parameters, such as the overall accuracy is introduced, which can be expressed as:
rotation frequency, will have effect on the amplitude and
1 n
the period of the curve in spectrograms. The distinctions OA   Cii , (10)
of different micro-motion forms are the curve N i 1
characteristics of scatterers. The rotationally symmetrical where 𝑁 is the number of total samples, 𝑛 is the number
target with spin does not effect micro-Doppler of categories, and 𝐶𝑖𝑖 is the number correctly classified
modulation. The micro-Doppler frequency shift is for class 𝑖.
approximately 0. Thus, there is not much information For spinning and tumbling, due to their simple
contained in spectrogram, which is often a widened micro-motion, their classification accuracy is higher.
straight line. As far as tumbling is concerned, the When the oscillating amplitude of the target is small or
amplitudes of different scatterers are the same. The the wobble frequency is slow, the difference between
difference between precession and tumbling is the nutation and precession will become smaller, and it is
amplitude of the curves. Although they are both easy to misjudge the nutation as precession. However,
sinusoidal curves, for precession, the amplitudes of the classification accuracy of nutation is still above 95%.
curves are different. For nutation, there will be no more In contrast, the support vector machine (SVM) which is
sinusoidal curves. The micro-Doppler of scatterers is a feature-based classification algorithm is also used
modulated by the mixing of spinning, coning, and for micro-motion forms classification. Since the size of
oscillating. At the same time, different polarizations spectrograms is 256 × 256, the number of features is
will affect the continuity of the curve. Therefore, the too large for SVM. Therefore, we use the principal
method based on feature engineering is difficult to component analysis (PCA) as a pre-processing. The
extract effective curve features from spectrograms, overall accuracy of SVM is 72.69%. It is 24.55% lower
especially under noise. Different from other feature than that of CNN, which is identical to the theoretical
based classification algorithms, CNN learns from a large analysis.

Table 1: Confusion matrix for the micro-motion forms classification based on CNN
Spin Tumbling Precession Nutation Accuracy (%)
Spin 3510 0 0 0 100
Tumbling  2715 63 12 97.31
Precession  59 12704 467 96.02
Nutation  139 548 15063 95.64

Additive Gaussian white noise of various Signal-to- modulation of scatterer 𝑃0 is a sinusoidal curve. The
Noise Ratio (SNR) levels (5, 10, 15, and 20 dB) are added sinusoidal curve will be destroyed in a low SNR, thus the
to simulate noise-contaminated radar echoes. Figure classification error will be increased. But it is clearly seen
4 demonstrates the anti-noise capability of the two that the performance of our algorithm is very good. When
algorithms. With decreasing SNR, the performance of these the SNR is 10 dB, the overall accuracy of our algorithm
algorithms decreases. The key point used to distinguish is still 91.03%, which is 29.79% higher than that of SVM.
precession from nutation is that the micro-Doppler Therefore, our method has strong anti-noise capability.
XU, YIN, DONG: MICRO-MOTION FORMS CLASSIFICATION OF SPACE CONE-SHAPED 69

The three scatterers are spherical crown 𝑃0 and two

scatterers 𝑃1 , 𝑃2 on the bottom edge (the intersections of
the incident surface and the bottom edge). Since the blunt
cone is axisymmetric, its attitudes can be characterized
by the included angle 𝜃 between the incident ray and the
symmetry axis of cone. Here, 𝜃 ranges from 0 to 𝜋. For
scatterer 𝑃0 , it is only visible in 0 ≤ 𝜃 < 𝜋/2 − 𝛼. Thus,
the complex RCS of 𝑃0 is [19]:
 1 1

 2 b 1  sin(2k0 b(1  sin  ))  ,  0

2

  
 k0 b cos 
2


 1

 0 =   2 b, 0    
 2 (11)
Fig. 4. The anti-noise capability with Gaussian noise.
 
0   
IV. CONCLUSION  2
In this paper, we describe the complex RCS 
variation in the target’s different attitudes; develop the 0  0,
echo models of space cone-shaped target with different where 𝑘0 = 2𝜋/𝜆0 .
micro-motion forms (spin, tumbling, precession, and
nutation); propose a CNN model for micro-motion forms For scatterer 𝑃1 , its RCS is not only related to 𝜃, but
classification based on spectrograms. Compared with also affected by polarization, which can be described as:
SVM, the method avoids artificial feature selection and 
sin  1
feature extraction. Simulated results illustrate that CNN n a  
1    cos  1
can exploit micro-Doppler features of space cone-shaped n k0 sin   n 
target and effectively discriminate different micro-motion
3  2  
1
forms. Since precession is a special kind of nutation with  
 cos  cos   (12)
the oscillating amplitude 𝜃𝑛 = 0, the overall accuracy of  n n  
them has decreased somewhat. However, compared with
SVM, ours has a higher overall accuracy. As a result, we 
1 = .
found that our algorithm effectively suppresses noise. 4
When the SNR is higher than 10 dB, the overall accuracy Here 𝑛 = 3/2 + 𝛼/𝜋. The upper sign applies when the
of our algorithm is higher than 90%. polarization is VV polarization while the lower sign
applies if it is HH polarization.
APPENDIX
The sharp cone is a special case of the blunt cone, at When the electromagnetic wave is incident
which the radius of the spherical crown of the blunt cone perpendicular to the conical surface, the divisor of
is 0. Suppose the bottom radius of the blunt cone is 𝑎, the equation (12) is 0. The RCS of scatterer 𝑃1 is mainly
half cone angle is 𝛼, and the spherical crown radius is 𝑏, caused by the specular reflection of the conical surface.
as shown in Fig. 5. √𝜎1 should be calculated by the physical optical method
[20, 21] and the maximum value is reached. At this point,
P2 𝜃 is equal to 𝜋/2 − 𝛼 , and the equation (12) can be
modified to:
1
a

b
 
 4 cos   2
3 3 3  2
2

 1   k0  a   b cos   2
  . (13)
  9 sin 2
   
P0  
iˆ Near 𝜋/2 − 𝛼, if √𝜎1 calculated by equation (12) is
h
greater than the maximum, it should be smoothed.
P1
Unlike scatterer 𝑃1 , scatterer 𝑃2 is visible only when
Fig. 5. Geometry of blunt-cone target. 0 < 𝜃 ≤ 𝛼 or 𝜃 ≥ 𝜋/2. Its RCS is:
70 ACES JOURNAL, Vol. 35, No. 1, January 2020

 sin( / n) a   
1
  3  2   
1

   cos  1  cos  cos  , 0    

 n k0 sin    n   n n   

 
 2  0,    .
 2
(14)
 sin( / n) a   
1
    2  
1

   cos  1  cos  cos   
 n k0 sin    n   n n   2


2 =-
4
The equations (12) and (14) give the variation of the will become meaningless and need to be modified. When
RCS of the bottom scatterers 𝑃1 and 𝑃2 with different 0 ≤ 𝜃 ≤ 𝛼, the contribution from scatterers 𝑃1 and 𝑃2
attitudes. When 𝜃 is approximately 0 or 𝜋 , 𝑠𝑖𝑛 𝜃 will can be expressed as :
approach zero, at which point the equations (12) and (14)

2  a sin  1
 3 
1 e   2 e  n   cos  cos  J 0  2k0 a sin  
j 1 j2

n  n n 
2 tan  3 (15)
j sin 1

n   
n J  2k a sin   cos  1 J  2k a sin    ,
   2
  3 
2 1 0
 n 
0


 cos  cos 
 n n 
where 𝐽𝑖 (𝑖 = 0, 1, 2) represents the Bessel function of Aerospace and Electronic Systems, vol. 46, no.4,
first kind of order 𝑖. pp. 1969-1982, 2010.
When 𝜃 > 𝜋 − 𝜃𝑐𝑎 and 2𝑘0 𝑎 𝑠𝑖𝑛 𝜃𝑐𝑎 = 2.44 , the [5] J. Lei, “Target classification based on micro-
contribution from scatterers 𝑃1 and 𝑃2 can be expressed doppler signatures,” Proceedings of the IEEE
as [22]: International Radar Conference, pp. 179-183, 2005.
J  2k a sin    j 2  [6] Y. Yang and J. Lei, “Target classification and
1 e j1   2 e j2   a 1 0 e . (16) pattern recognition using micro-doppler radar
sin  signatures,” Proceedings of the Seventh ACIS
International Conference on Software Engineering,
ACKNOWLEDGMENT Artificial Intelligence, Networking, and Parallel/
This work was supported by The Major Research Distributed Computing, pp. 213-217, 2006.
Plan of National Natural Science Foundation of China [7] J. Li and H. Ling, “Application of adaptive chirplet
under Grant no. 61490690, 61490695. representation for ISAR feature extraction from
targets with rotating parts,” IEE Proceedings of
REFERENCES Radar Sonar and Navigation, vol. 150, no. 4, pp.
[1] V. C. Chen, F. Y. Li, S. S. Ho, and H. Wechsler, 284-291, 2003.
“Analysis of micro-doppler signatures,” IEEE IET [8] S. Ljubisa, D. Igor, and T. Thayananthan,
Proc.-Radar Sonar Navig., vol. 150, no. 4, pp. 271- “Separation of target rigid body and micro-doppler
276, 2003. effects in ISAR imaging,” IEEE Transactions on
[2] V. C. Chen, F. Y. Li, S. S. Ho, and H. Wechsler, Aerospace and Electronic Systems, vol. 42, pp.
“Micro-Doppler effect in radar: Phenomenon, 1496-1506, 2006.
model, and simulation study,” Trans. Electron [9] A. Bultan, “A four-parameter atomic decomposition
Syst., vol. 42, no. 1, pp. 2-21, 2006. of chirplets,” IEEE Transactions on Signal
[3] Y. Kim and T. Moon, “Human detection and Processing, vol. 47, pp. 731-745, 1999.
activity classification based on micro-doppler [10] K. Li, Y. Liu, and K. Huo, “Estimation of micro-
signatures using deep convolutional neural motion parameters based on cyclostationary
networks,” IEEE Geoscience and Remote Sensing analysis,” IET Signal Processing, vol. 4, no. 3, pp.
Letters, vol. 13, no. 1, pp. 1-5, 2015. 218-223, 2010.
[4] H. W. Gao, L. G. Xie, S. L. Wen, and Y. Kuang, [11] Y. Liu, K. Li, and Z. Zhuang, “Estimation of micro-
“Micro-doppler signature extraction from ballistic motion parameters based on micro-doppler,” IET
target with micro-motions,” IEEE Transactions on Signal Processing, vol. 4, no.3, pp. 213-217, 2010.
XU, YIN, DONG: MICRO-MOTION FORMS CLASSIFICATION OF SPACE CONE-SHAPED 71

[12] Y. LeCun, B. Boser, J. S. Denker, D. Henderson, Gaogui Xu received the B.S. degree
R. E. Howard, W. Hubbard, and L. D. Jackel, and the M.S. degree in Computer
“Handwritten digit recognition with a back- Application Technology from China
propagation network,” Advances in Neural Three Gorges University, Hubei
Information Processing Systems, vol. 2, no. 2, pp. Province, China, in 2014 and 2017
396-404, 1990. respectively.
[13] A. Krizhevsky, I. Sutskever, and G. Hinton, He is currently pursuing the
“ImageNet classification with deep convolutional Ph.D. degree of Electromagnetic
neural networks,” Adv. Neural Inf. Process. Syst., Field and Microwave Technology at Communication
vol. 25, pp. 1090-1098, 2012. University of China, Beijing, China. His research
[14] C. Szegedy, W. Liu, Y. Jia, et al., “Going deeper interests include micro-Doppler, micro-motion, target
with convolutions,” 2015 IEEE Conference on recognition, and deep learning.
Computer Vision and Pattern Recognition (CVPR)
IEEE, 2015. HongCheng Yin was born in
[15] G. Hinton, et al., “Deep neural networks for Yujiang, Jiangxi Province, China, in
acoustic modeling in speech recognition,” IEEE 1967. He received the B.S. degree in
Signal Process. Mag., vol. 29, no. 29, pp. 82-97, 1986 from Northwest Telecommun-
2012. ication Engineering Institute, Xi’an,
[16] T. Mikolov, A. Deoras, D. Povey, L. Burget, and China; the M.S. degree in 1989 from
J. Cernocky, “Strategies for training large scale the Science and Technology on
neural network language models,” Automatic Electromagnetic Scattering Labora-
Speech Recognition & Understanding, pp. 196- tory, and the Ph.D. degree in 1993 from Southeast
201, 2011. University, Nanjing, China, all in electromagnetic field
[17] R. M. Murray, Z. Li, and S. S. Sastry, A and microwave technique.
Mathematical Introduction to Robotic Manipulation. He is now a Research Scientist of Science and
CRC Press, Boca Ration, 1994. Technology on Electromagnetic Scattering Laboratory,
[18] J. M. O’Toole and B. Boashash, “Fast and memory- Beijing, China. He has coauthored four books and
efficient algorithms for computing quadratic time– published more than 130 papers in technical journals
frequency distributions,” Applied and Computa- and conferences. His research interests include electro-
tional Harmonic Analysis, vol. 35, no. 2, pp. 350- magnetic scattering, radar target signature, and radar
358, 2013. target identification.
[19] L. C. Potter, D. M. Chiang, R. Carriere, and M. J. He is a fellow of Chinese Institute of Electronics.
Gerry, “A gtd-based parametric model for radar
scattering,” IEEE Transactions on Antennas and Chunzhu Dong was born in 1981.
Propagation, vol. 43, no. 10, pp. 1058-1067, 2002. He received the B.S. degree in
[20] A. B. Gorji, B. Zakeri, and R. C. Janalizadeh, Zhengzhou University, Henan
“Physical optics analysis for RCS computation of a Province, China, in 2004; the M.S.
relatively small complex structure,” The Applied degree and the Ph.D. degree in
Computational Electromagnetics Society Journal, Electromagnetic Field and Micro-
vol. 29, no. 7, pp. 530-540, 2014. wave Technology from the Comm-
[21] J. Perez and M. F. Catedra, “Application of unication University of China,
physical optics to the RCS computation of Beijing, China, in 2007 and 2015 respectively.
bodies modeled with NURBS surfaces,” IEEE Currently he is a Senior Engineer of Science and
Transactions on Antennas and Propagation, vol. Technology on Electromagnetic Scattering Laboratory,
42, no. 10, pp. 1404-1411, 1994. Beijing, China. His research interests include radar target
[22] R. A. Ross, Investigation of Scattering Principles. signature, electromagnetic scattering, and microwave
Vol III. Analytical Investigation General Dynamics, technique.
Forth-worth, Texas, 1969.