Download as pdf or txt
Download as pdf or txt
You are on page 1of 13

IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL.

70, 2021 8001113

A Target Detection Scheme With Decreased


Complexity and Enhanced Performance for
Range-Doppler FMCW Radar
Qiangwen Zheng , Lijie Yang , Yaping Xie , Junjie Li , Tang Hu , Jiang Zhu , Member, IEEE,
Chunyi Song , Member, IEEE, and Zhiwei Xu , Senior Member, IEEE

Abstract— In conventional fast-ramp-based target detection to its superior performance in target detection and local-
schemes for the frequency modulated continuous wave (FMCW) ization [1]–[3]. Compared with lidar, ultrasound, and video
radar systems, all chirp sequences are performed fully on cameras, which are also the widely used sensors, the FMCW
two-dimensional fast Fourier transforms (FFTs) to estimate
range and velocity of a target, and all range-Doppler cells are radar is more robust for target detection under bad weather
scanned by the two-dimensional (2-D) cell-averaging constant conditions of heavy rain, snow, and dense fog [4], and is,
false alarm rate (CA-CFAR) detector to extract valid targets therefore, widely used in applications of automotive [5],
against background noise. This results in an extremely high security [6], and surveillance [7].
computational complexity. To decrease the complexity while The fast-ramp-based two-dimensional (2-D) fast Fourier
maintaining a desired performance, this article proposes a target
detection scheme based on distributed one-dimensional (1-D) transform (FFT) is used for estimating range and velocity
CA-CFAR and region-of-interest (ROI) preprocessing, which of targets in FMCW radar systems, and all chirp signals
performs Doppler-FFT only for range ROI cells and scans only are performed with the 2-D FFT operations regardless of
Doppler ROI cells to extract pretargets. Simulation results indi- targets numbers. This has resulted in extraneous computa-
cate that the proposed scheme significantly decreases complexity tional complexity [8]. A peak-detection algorithm needs to be
and simultaneously improves detection performance in terms of
detection probability, figure of merit (FoM), and CFAR loss, applied for identifying presence or absence of valid targets
as compared with the conventional schemes. The advantages in range-Doppler matrix (RDM) [9] and further extracting
of the proposed scheme over the conventional ones are further the corresponding location and motion parameters against
verified by field tests under multitarget scenario, using our background noise [10], too.
recently developed FMCW radar with high angular resolution. Recently, some parameter estimation schemes with low
Index Terms— Detection analysis, frequency modulated con- complexity have been studied [11]–[13], which are only
tinuous wave (FMCW) radar, low complexity, region of interest applicable to specific scenarios. A pedestrian detection scheme
(ROI), target detection. is developed in [11] for weak moving target detection under
strong clutter, which first detects the coherent phase dif-
I. I NTRODUCTION ference from the spectrum data obtained from range-FFT
operation, and then only the range profiles of moving pedes-
F REQUENCY modulated continuous wave (FMCW) radar
has increasingly attracted interests in recent years, due trians are extracted and further performed Doppler-FFT oper-
ation. The scheme is thus invalid for weak stationary target.
Manuscript received May 25, 2020; accepted September 16, 2020. Date Hyun et al. [12] propose a region-of-interest (ROI) processing
of publication September 28, 2020; date of current version December 7,
2020. This work was supported in part by the National Natural Science scheme to decrease the complexity of blind spot detection,
Foundation of China (NSFC) under Grant 61971379, in part by the Zhejiang which, however, is only applicable to limited number of
Lab through the Major Scientific Project under Grant 2018DD0ZX01, in part targets. The improved moving target indicator (MIT) scheme
by the Leading Innovative and Entrepreneur Team Introduction Program of
Zhejiang under Grant 2018R01001, in part by the Fundamental Research in [13] solves the blind-speed problem and reduces the com-
Funds for the Central Universities, and in part by the Zhejiang Institute putational complexity by using random dual chirps, at the
of Standardization under Project 2019SY05WX0007. The Associate Editor expense of losing the location information of stationary targets
coordinating the review process was Shoaib Amin. (Corresponding author:
Chunyi Song.) and the Doppler information of moving targets.
Qiangwen Zheng, Yaping Xie, and Junjie Li are with the Institute of To realize accurate identification of valid targets against
Marine Electronic and Intelligent System, Ocean College, Zhejiang University, background noise, the constant false alarm rate (CFAR)
Zhoushan 316021, China.
Lijie Yang and Tang Hu are with the Zhejiang Lab, Hangzhou 310006, detection has been developed, which mitigates the impact
China. of background noise by automatically adjusting the detec-
Jiang Zhu, Chunyi Song, and Zhiwei Xu are with the Institute of Marine tion threshold [14]–[16]. To fully utilize the characteristics
Electronic and Intelligent System, Ocean College, Zhejiang University,
Zhoushan 316021, China, also with the Engineering Research Center of of targets in both range and Doppler domains, a 2-D rec-
Oceanic Sensing Technology and Equipment, Ministry of Education, Zhejiang tangle reference window is employed to derive the local
University, Zhoushan 316021, China, and also with the Key Laboratory of noise level for the conventional 2-D cell-averaging CFAR
Ocean Observation-Imaging Testbed of Zhejiang Province, Zhejiang Univer-
sity, Zhoushan 316021, China (e-mail: cysong@zju.edu.cn). (CA-CFAR) detection, realized by averaging the power of
Digital Object Identifier 10.1109/TIM.2020.3027407 multiple range-Doppler cells over the reference window [17].
1557-9662 © 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.

Authorized licensed use limited to: INSTITUTO MILITAR DE ENGENHARIA. Downloaded on October 11,2022 at 21:08:40 UTC from IEEE Xplore. Restrictions apply.
8001113 IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 70, 2021

where A T is the amplitude of the transmitted chirp signal, f 0


is the starting frequency, μ is the frequency modulated slope,
i.e., μ = B/T with B representing the bandwidth of the chirp
signal, and rect((t − T v2/T )) is the normalized rectangular
window and is defined as
  
t − T /2 1, |t − T /2| ≤T /2
rect = (2)
T 0, |t − T /2| >T /2

Therefore, one frame of transmitted signals x(t) is given by


Fig. 1. Illustration of the transmitted and received signals of FMCW radar.

L−1
x(t) = x 0 (t−lT ) (3)
A 2-D cross reference window is proposed and applied in [18] l=0

to reduce the number of extraneous reference cells at the where l = 0, 1, . . . , L −1 and L is the number of chirp signals
expense of degrading a certain detection accuracy. To fully in one frame. Then, we have the total time duration TF = LT .
utilize the energy distribution features of targets [19], a 2-D In the FMCW radar system, the transmitted chirp signals are
circular reference window is proposed to improve estimation received by the receiver antennas after propagation. Suppose
of background noise and detection of valid targets [20]. How- that the echo signals received by a single receiver antenna are
ever, the abovementioned 2-D CA-CFAR detectors need to reflected by M distinctive targets, i.e., the superposition of
scan every range-Doppler cell to identify valid targets against the received signals for the lth transmitted chirp signal can be
background noise, resulting in extraneous computation. expressed as
Aiming at decreasing complexity while guaranteeing detec-
tion performance, we propose a novel target detection scheme 
M

for range-Doppler FMCW radar in this work. With conduct- r (t, l) = A(m) −1
R A T x 0 (t−τm − lT ) + ω(l, t) (4)
ing ROI preprocessing in range and Doppler domains, only m=1

range-Doppler cells representing existence of valid targets are where m = 1, 2, . . . , M and M is the total number of
filtered out and processed by Doppler-FFT and CA-CFAR targets, A(m)
R and τm denote the received signal’s amplitude
operations. Meanwhile, distributed 1-D CA-CFAR operations and the round-trip time delay for the mth specific target,
are performed three times to fully utilize the information of respectively, and ω(l, t) is the additive white Gaussian noise
targets in both range and Doppler domains, to improve the (AWGN). Then, the amplitude of the received signal A(m)R can
detection performance. In the end, a peak focus algorithm is be expressed as [22]
carried out on the produced pretargets to further identify the √
main-lobe cells of valid targets while decreasing the ghost σm G T G R
A(m) = A (5)
(2Rm )α
target cells. The ghost target here represents a target that does R T

not truly exist; the detection of a ghost target then means the
occurrence of false alarm [21]. where σm is the reflection coefficient of radar cross section
The structure of this article is organized as follows. (RCS) for the mth target, G T and G R , respectively, are the
Section II briefly introduces the system model and conven- gain of transmitter antenna and that of receiver antenna, Rm
tional detection scheme. The proposed detection scheme is is the range between the mth target and the radar sensor, and
presented in Section III. Simulation results are demonstrated α is the attenuation coefficient decided by the propagation
and analyzed in Section IV. Section V shows the field test environment.
results in multitarget situation. Finally, the conclusion is drawn The received signals are then conjugately mixed with the
in Section VI. transmitted signals to transform the wideband nonstationary
signals into the superposition of a series of low-frequency
single-tone signals, the so-called beat signal, followed by a
II. FMCW R ADAR S YSTEM M ODEL AND low-pass filtering. The operation is repeated for L consecu-
C ONVENTIONAL D ETECTION S CHEME tive chirp signals. The beat signal corresponding to the lth
transmitted chirp signal then can be expressed as
A. FMCW Radar System Model
This section introduces the FMCW radar system model. 
M

Fig. 1 shows the transmitted and received signals of FMCW s(l, t) = x(t)r ∗ (l, t) = A(m)
R AT
radar. The frequency of the FMCW chirp signal, which   m=1  
is a wideband nonstationary signal, increases linearly with Bτ
× exp j 2πμτm t + 2π f 0 τm 1 −
transmitted time during the modulation duration T . A single 2 f0 T
FMCW chirp signal x 0 (t) is expressed as + ws (l, t) (6)
   
t − T /2 μ  where (·)∗ is the complex conjugate, and
x 0 (t) = A T rect exp j 2π f 0 t + t 2 (1)
T 2 ws (l, t) =x(t)ω∗ (l, t). Since f 0 B and T  τ in reality [23],

Authorized licensed use limited to: INSTITUTO MILITAR DE ENGENHARIA. Downloaded on October 11,2022 at 21:08:40 UTC from IEEE Xplore. Restrictions apply.
ZHENG et al.: TARGET DETECTION SCHEME WITH DECREASED COMPLEXITY AND ENHANCED PERFORMANCE 8001113

(6) is simplified as

M
s(l, t) = A(m)
R A T exp( j (2πμτm t +2π f 0 τm ))+ws (l, t).
m=1
(7)
The round-trip time delay of the lth transmitted chirp signal
for the mth target can be further extended as
2Rm +2v m lT
τm,l = (8)
c
where c is the speed of light and v m is the radical velocity of
the mth target. Substituting (8) into (7), s(l, t) can be rewritten
as Fig. 2. Schematic of the 2-D CA-CFAR detector.

M
s(l, t)= am exp( j (2π( fr,m t + fd,m lT +γm )))+ws (l,t) (9)
where q is the index of Doppler bins in the Doppler domain
m=1
and 1 ≤ q ≤ NC , NC is the number of Doppler-FFT points.
where am is the mth complex amplitude term, i.e., am = Therefore, applying the 2-D FFT operation, an RDM is
A(m)
R A T exp( j (4π f 0 Rm /c)), f r,m = 2μR m /c is the range fre- produced containing range and velocity information of targets
quency of the mth target, fd,m = 2 f 0 v m /c is the Doppler and undesired clutter or noise background, as follows:
frequency of the mth target, γm = 2μv m lT t/c is the range- ⎡ ⎤
Doppler-coupling, and can be neglected in the 2-D FFT S̃(1, 1) S̃(2, 1) · · · S̃(NC , 1)
operation [24]. ⎢ S̃(1, 2) S̃(2, 2) · · · S̃(NC , 2) ⎥
⎢ ⎥
The corresponding continuous beat signal s(l, t) is sam- RDM = ⎢ .. .. .. .. ⎥. (13)
⎣ . . . . ⎦
pled with f s to produce N samples within the duration T ,
S̃(1,NS ) S̃(2,NS ) · · · S̃(NC ,NS )
as follows:

M To extract targets’ parameters against background noise,
s(l, n) = am exp( j (2π( fr,m nTs + f d,m lT +γm )))+ws (l, n) the 2-D CA-CFAR detector is executed to estimate the local
m=1 noise power based on the arithmetic mean of the neighborhood
(10) areas. Fig. 2 depicts the schematic of 2-D CA-CFAR detector.
In Fig. 2, the reference window consists of a cell under test
where fs is the sampling frequency, Ts = 1/ f s is the sam- (CUT), guard cells, and reference cells. The reference cells
pling interval, and N = T /Ts is the numbers of samples, are regarded as the samples of noise estimate around the CUT.
n= 0, 1, . . . , N−1. First, the RDM is carried out the square-law operation, i.e.,
2
Z (q, k) = | S̃(q, k)| . Then, the local noise power around the
B. Conventional Target Detection Scheme current CUT is estimated as
This section briefly illustrates the conventional target detec- 1 
Nr
tion scheme for FMCW radar, which consists of a 2-D FFT P(q, k) = Z (nr ) (14)
Nr n =1
and a 2-D CA-CFAR algorithm [3], [18]. The beat signal can r

be denoted by a 2-D sinusoidal waveform across two time


where (q, k) is the coordinates of the current CUT, Nr
indices [25], which contain the fast time index n and the
is the number of reference cells in the reference window,
slow time index l. To estimate the range information, L times
and Z (nr ) is the power of each reference cell. Accordingly,
N S -point range-FFTs across the fast time are performed on the
the detection threshold is denoted by T (q, k) = α · P(q, k),
beat signal sequences s(l, n). Thus, the kth range-FFT output −1/N
with α = Nr (Pfa r − 1) representing the threshold scaling
of the lth sampling sequence s(l, n) is expressed as
factor and Pfa denoting the desired false alarm probability. The
Ns  threshold can dynamically adopt to the change of background

S(l, k) = s(l, n)exp − j n(k − 1) (11) noise caused by temperature variation or some other physical
NS
n=1 effects [20]. Finally, whether the current CUT is a target can
where k is the index of range bins in the range domain; for be determined based on the following conditions:
1 ≤ k ≤ N S , N S is the number of range-FFT points. Mean- H0
while, to estimate the velocity information, N S times NC -point Z (q, k) ≷ T (q, k) (15)
H1
Doppler-FFTs across the slow time are performed on S(l, k),
and the qth Doppler-FFT output is denoted by S̃(q, k), where H0 and H1 denote the hypothesis of presence and
as follows: absence of a target in the current CUT, respectively. The
 NC  conventional scheme results in an extremely high computa-

S̃(q, k) = S(l, k)exp − j (l−1)(q − 1) (12) tional complexity because two-stage full-dimensional FFT is
l=1
NC performed and all range-Doppler cells are scanned.

Authorized licensed use limited to: INSTITUTO MILITAR DE ENGENHARIA. Downloaded on October 11,2022 at 21:08:40 UTC from IEEE Xplore. Restrictions apply.
8001113 IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 70, 2021

−1/N
with
 Nr1 the detection threshold T(k) =(Pfa r1 −
1) nr1 =1 Z (nr1 ), with Nr1 denoting the number of reference
cells in the linear reference window, and Z (nr1 ) representing
the power of each reference cell.
The range ROI processing is accomplished as soon as all
range ROI cells are determined. Consequently, the super-
position of determined range ROI cells in all range-profile
sequences can be described as S(l, kn u ) for 1 ≤ n u ≤ Nu ,
where kn u and Nu are the index of the range bins corresponding
to range ROI cells and the number of detected targets in range
domain, respectively.

B. Doppler ROI Detection


Fig. 3. Structures of detection schemes. (a) Conventional scheme. (b) Pro- As the second main step, Nu times NC -point Doppler-FFTs
posed scheme. across the slow time are performed on the range ROI detection
results S(l, kn u ) for velocity estimation. The operation times of
NC -point Doppler-FFT are decreased from N S to N u for Nu 
N S . Meanwhile, the qth Doppler-FFT output for S(l, kn u ) is
denoted by S̃(q, kn u ), for k1 ≤kn u ≤ k Nu , and is expressed as


NC 

S̃(q, kn u ) = S(l,kn u )exp − j (l − 1)(q − 1) . (17)
l=1
NC

The RDM based on range ROI cells is obtained, as follows:


⎡ ⎤
S̃(1,k1 ) S̃(2,k1 ) · · · S̃(NC ,k1 )
Fig. 4. Schematic of the 1-D CA-CFAR detector. ⎢ S̃(1,k2 ) S̃(2,k2 ) · · · S̃(NC ,k2 ) ⎥
⎢ ⎥
RDMROI = ⎢ .. .. .. .. ⎥. (18)
⎣ . . . . ⎦
III. P ROPOSED D ETECTION S CHEME S̃(1,k Nu ) S̃(2,k Nu ) ··· S̃(NC , k Nu )
This section proposes a low-complexity target detection It is possible that multiple targets exist at the same range bin
scheme based on distributed 1-D CA-CFAR detection and ROI with different velocities. Thus, to extract the effective Doppler
preprocessing for FMCW radar, aiming to decrease the compu- bins corresponding to the targets with various velocities, i.e.,
tational complexity and improve the detection performance in Doppler ROI cells, the 1-D CA-CFAR detection is executed
multitarget environment. As illustrated in Fig. 3, the detection for the second time (the second 1-D CA-CFAR detection).
process of the conventional scheme primarily consists of In the same manner, the 1-D CA-CFAR detector is employed
analog-to-digital conversion (ADC), 2-D FFT operation, and on all range-Doppler cells in RDMROI along the Doppler
2-D CA-CFAR detection; the detection using the proposed domain. After Doppler ROI detection, the effective Doppler
scheme is accomplished by the following three main steps. bins, the so-called Doppler ROI cells, at the kn u th range bin
are expressed as S̃(qm p , kn u ) for 1≤m p ≤M p , with M p denoting
A. Range ROI Detection the number of the effective Doppler bins at the kn u th range bin
and qm p denoting the index of effective Doppler bins at the
As the first main step, in the same manner, L times
kn u th range bin. In reality, the number of effective Doppler
N S -point range-FFTs across the fast time are performed for
bins is not necessarily the same at various rang bins, i.e., M p
range estimation. Some and even the most range bins represent
is a function of kn u as M p = F(kn u ). Thus, the total number
background noise in multitarget circumstance [26]. In order to
of Doppler ROI cells produced by Doppler ROI detection can
determine the valid range bins that represent valid targets, the
be denoted by N P , i.e., N P = nNuu=1 F(kn u ).
l1 th range-profile sequence Sl1 (k) is chosen to be executed
the first 1-D CA-CFAR detection. Fig. 4 shows the schematic
of the 1-D CA-CFAR detector. The 1-D CA-CFAR detector C. Pretargets Estimation and Peak Focus Detection
uses a linear reference window to estimate the local noise
level and determine the valid range bins. Similarly, the l1 th The second 1-D CA-CFAR detection only employs the
range-profile sequence is executed the square-law operation, features of targets in Doppler domain. To fully utilize the
i.e., Z (k) = |Sl1 (k)|2 =|S(l1 , k)|2 , l1 ∈ l. Then, the valid range characteristics of targets in both range and Doppler domains
bins, i.e., range ROI cells, can be determined via to obtain better detection performance, the 1-D CA-CFAR
detection is performed for the third time (the third 1-D
H0 CA-CFAR detection) in range domain on the Doppler ROI
Z (k) ≷ T (k) (16)
H1 cells. Above all, the partial DFT operation in Doppler domain

Authorized licensed use limited to: INSTITUTO MILITAR DE ENGENHARIA. Downloaded on October 11,2022 at 21:08:40 UTC from IEEE Xplore. Restrictions apply.
ZHENG et al.: TARGET DETECTION SCHEME WITH DECREASED COMPLEXITY AND ENHANCED PERFORMANCE 8001113

is carried out on the reference cells around the Doppler ROI


cells S̃(qm p , kn u ) [27], as follows:
 NC 

S̃(qm p , kr ) = S(l,kr )exp − j (l − 1)(qm p − 1) (19)
l=1
NC
with kn u − (Nr1 + Ng1 + 2/2) ≤ kr ≤kn u +
(Nr1 + Ng1 + 2)/(2), q1 ≤ qm p ≤ q M p , where kr is the
index of the reference cells in range domain, Nr1 is the
number of reference cells, and Ng1 is the number of guard
cells. The total times of the partial DFT operation is denoted
by N D , then N D = N P Nr1 . The results of the third 1-D
CA-CFAR detection performing on the Doppler ROI cells, the
so-called pretargets, are expressed as S̃(q p , k p ) for q p ∈ qm p
and k p ∈ kn u , with (q p , k p ) representing the coordinates of
determined pretargets.
To further improve the detection accuracy of valid targets
while avoiding the detection of ghost targets, the peak focus
algorithm is executed for the detected pretargets, as follows:
2 H0 2
| S̃(qm p ,kr )| ≷ max{| S̃(i, j )| } (20)
H1
2
where qm p −1 ≤ i ≤q m p +1 and kr −1 ≤ j ≤kr +1, | S̃(qm p , kr )| Fig. 5. Architectures of the reference window for the conventional three
2
is the power of pretargets, and | S̃(i, j )| is the power of schemes. (a) Rectangle window. (b) Cross window. (c) Circular window.
range-Doppler cells around a pretarget. The proposed scheme
finishes detection when all pretargets have been performed
with the peak focus algorithm and determine valid targets. The proposed scheme executes L times N S -point range-FFT
operations, Nu times NC -point Doppler-FFT operations, Nr N p
IV. P ERFORMANCE E VALUATION times NC -point partial DFT operations, and three times 1-D
In this section, performance of the proposed scheme is CA-CFAR detection. The total number of multiplications and
evaluated and compared with conventional ones through sim- that of additions are, respectively, expressed as
ulation analysis. The three conventional schemes chosen here NS NC
are all based on 2-D FFT and 2-D CA-CFAR algorithms, C M2 = L log2 N S +2N S +Nu log2 NC +2Nu NC
2 2
differing only in the reference window: a 2-D cross refer- +Nr N p NC + Nr N p gr N p
ence window [18], a 2-D circular reference window [20], C A2 = L N S log2 N S +N S (Nr −1)rNu NC log2 NC
and a 2-D rectangle reference window [17], respectively.
They, respectively, are named “FFT plus cross scheme,” “FFT +Nu NC (Nr −1)+Nr N p (NC −1)+N p (Nr −1). (22)
plus circular scheme,” and “FFT plus rectangle scheme” In (21), Nr is the only parameter that may slightly vary
for convenience. The architectures of the reference win- among the three schemes. Without losing generality, the FFT
dows for the three conventional schemes are severally shown plus cross scheme is selected as the representative of the three
in Fig. 5(a)–(c). conventional schemes and used for computational complexity
comparison. The comparison is made based on the settings of
A. Computational Complexity Analysis L = 128, N S = 1024, N P = 35, and Nr = 28.
In this work, the required number of multiplications and Fig. 6 shows the comparison results between the proposed
additions is used as the metrics to analyze the computational scheme and the FFT plus cross scheme when fixing Nu = 15,
complexity. The three conventional schemes have consis- with the y-axis representing the number of multiplication C M
tent expression of complexity since they conduct the same and the number of addition C A defined in (21) and (22), and
operation procedure. Each conventional scheme executes L C M + C A . The results demonstrate that the proposed scheme
times N S -point range-FFT operations and N S times NC -point has lower computational complexity than the conventional
Doppler-FFT operations for obtaining the 2-D RDM. Mean- ones, and the advantage could be enhanced by increasing NC .
while, the 2-D CA-CFAR detection, in which the number of Fig. 7 shows the comparison results between the two
reference cells is Nr , is employed to scan all range-Doppler schemes under various Nu /Ns . The gap of computational
cells. Thus, the required number of multiplications and addi- complexity between the proposed scheme and the conventional
tions, denoted by C M1 and C A1 , respectively, is calculated as one varies significantly according to the ratio of Nu /Ns .
follows: The proposed scheme obtains significant advantage over the
NS NC conventional one when the ratio of Nu /Ns is small, which
C M1 = L log2 N S +N S log2 NC +2N S NC is the most common case in reality. The advantage decreases
2 2
C A1 = L N S log2 N S +N S NC log2 NC + N S NC (Nr −1). (21) fast with the increase of the ratio of Nu /Ns ; when the ratio

Authorized licensed use limited to: INSTITUTO MILITAR DE ENGENHARIA. Downloaded on October 11,2022 at 21:08:40 UTC from IEEE Xplore. Restrictions apply.
8001113 IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 70, 2021

Fig. 6. Comparison of computational complexity under the fixed Nu = 15. Fig. 8. Comparison of the normalized ADT among the 1-D CA-CFAR
detector and the three 2-D CA-CFAR detectors.

Fig. 8 demonstrates the comparison results of the nor-


malized ADT among the 1-D CA-CFAR detector and the
three 2-D CA-CFAR detectors in terms of the noise estima-
tion accuracy, based on the same settings of the number of
reference cells Nr = 14, the number of guard cells Ng = 4
in both range and Doppler dimensions, and the false alarm
probability Pfa = 10−6 . The 1-D CA-CFAR detection has two
choices of performing the noise estimation in range dimension
and in Doppler dimension. By increasing SNR, the impact of
the extended target power to the noise estimation accuracy
becomes more severe, which has resulted in the increased
ADT. Based on the same number of reference cells, the 1-D
CA-CFAR detector in range dimension achieves the lowest
(the best) ADT, while the 1-D CA-CFAR detector in Doppler
dimension achieves the largest (the worst) ADT, among the
Fig. 7. Comparison of computational complexity under various Nu /Ns . all five detectors. In particular, the advantage obtained by
the 1-D CA-CFAR detector in range dimension over the
three 2-D CA-CFAR detectors is more significant than the
approaches to one, the proposed scheme lose the advantage gap between the three 2-D CA-CFAR detectors and the 1-D
and obtains a slightly higher computational complexity than CA-CFAR detector in Doppler dimension. The comparison
the conventional one. The three crossing points are also results indicate that the target power extended in Doppler
marked in Fig. 7. dimension has degraded the noise estimation accuracy.
The proposed detection scheme in this work is based on
the distributed 1-D CA-CFAR and ROI preprocessing. The
B. Noise Estimation Accuracy Analysis
procedure is mainly accomplished by performing the following
In this section, the noise estimation accuracy is analyzed 1-D CA-CFAR detections for three times: 1) range ROI
and discussed for the 1-D CA-CFAR detector and the three detection: the first 1-D CA-CFAR detection is performed in
2-D CA-CFAR detectors defined in Fig. 5. range dimension to extract the range-ROI cells; 2) Doppler
The normalized average decision threshold (ADT) [28], [29] ROI detection: the second 1-D CA-CFAR detection is per-
is defined and applied to evaluate the noise estimation accuracy formed in Doppler dimension to obtain Doppler-ROI cells; and
and detection performance of CA-CFAR detectors, shown as 3) pretargets estimation: the third 1-D CA-CFAR detection
E{T } is performed in range dimension to extract the pretarget
ADT = (23) cells. The detection performance of the proposed scheme is
μ
resulted from the whole procedure consists of the above three
where {E T } = αC A · μ̄ is the estimated decision threshold, operations, and could not be directly derived from the noise
αC A is the threshold factor, μ̄ is the estimated noise power estimation accuracy of the 1-D CA-CFAR detection either
value, and μ is the true noise level. in range dimension or in Doppler dimension. Accordingly,

Authorized licensed use limited to: INSTITUTO MILITAR DE ENGENHARIA. Downloaded on October 11,2022 at 21:08:40 UTC from IEEE Xplore. Restrictions apply.
ZHENG et al.: TARGET DETECTION SCHEME WITH DECREASED COMPLEXITY AND ENHANCED PERFORMANCE 8001113

Fig. 9. Simulation results: detection probability versus SNR when


Pfa = 10−6 . Fig. 10. Simulation results: the ROC curve at SNR = 10 dB.

the detection performance should be evaluated in terms of D. CFAR Loss Analysis


detection probability Pd [30]. The CFAR loss describes the difference of SNR to reach the
same detection probability for a specific false alarm probabil-
ity, between a CFAR detector and the ideal Neyman–Pearson
C. Detection Probability Analysis detector [32]. The required SNR to achieve a target detection
The detection probability Pd is another indicator to evaluate probability Pd for the given false alarm probability Pfa is
the detection performance. The detection scene and parame- denoted by
ters are set as follows. In a homogeneous multitarget scene, (Pd /Pfa )1/Nr −1
16 Swerling I targets [31] are randomly distributed in a 2-D SNR = 1/Nr
. (24)
1 − Pd
RDM with 1024 × 1024 size, which contains a Rayleigh
distributed interference background with an average power Accordingly, the ideal SNR can be described as
of 30 dB; impact of the multitarget shadowing effect produced ln(Pfa /Pd )
by the randomly distributed 16 targets on the detection perfor- SNRideal = lim SNR= . (25)
n→∞ ln(Pd )
mance will be investigated in the simulation. The numbers of
reference cells are set to 28, 28, 28, and 26 for the proposed Thus, the CFAR loss is expressed as [20]
scheme, FFT plus cross scheme, FFT plus circular scheme, CFARLoss = SNR/SNRideal
and FFT plus rectangle scheme, respectively. At the same (Pd /Pfa )1/N −1 ln(Pfa /Pd )
time, Monte Carlo simulations are repeated for 1000 times = / . (26)
1−Pd
1/N ln(Pd )
to obtain the detection probability curve under the condition
of the desired false alarm probability Pfa = 10−6 , as illustrated Fig. 11 produces the CFAR loss curves by setting the
in Fig. 9, and the receiver operating characteristic (ROC) curve target detection probability Pd to 0.9 and varying the alarm
at the SNR = 10 dB, as demonstrated in Fig. 10. false probability Pfa from 10−6 to 10−1 . According to the
In Fig. 9, the proposed scheme has a higher detection results, the CFAR loss is inversely proportional to the false
probability than the three conventional ones. The detection alarm probability for all the four schemes; moreover, the pro-
probability of proposed scheme reaches 100% at SNR = posed scheme achieves the lowest CFAR loss among the four
13 dB, while detection probabilities of the three conven- schemes and decreases the CFAR loss by 43.1% compared
tional ones reach 100% at SNR > 20 dB. Considering that with the FFT plus circular scheme, which shows the best
the detection performance is achieved for the scenario with CFAR loss among the three conventional ones, when the
16 targets, the simulation results suggest that the multitarget false alarm probability is set to 10−5 . According to (25),
shadowing effect is mitigated through applying the distributed the performance of CFAR loss is negatively correlated with
1-D CA-CFAR detection for three times in the proposed the number of reference cells. The FFT plus rectangle scheme
scheme. achieves the highest CFAR loss due to that it has the smallest
The results in Fig. 10 demonstrate that the proposed scheme number of reference cells. In contrast, the proposed scheme
obtains better detection performance over the conventional performs the distributed 1-D CA-CFAR detection for three
ones within the investigated range of false alarm probability. times and, as a result, obtains
√ an equivalent number of
The comparison results are consistent with that in Fig. 9 for reference cells equals to 28 2 3, which is the largest among the
the all four schemes. four schemes. Interestingly, the FFT plus circular scheme and

Authorized licensed use limited to: INSTITUTO MILITAR DE ENGENHARIA. Downloaded on October 11,2022 at 21:08:40 UTC from IEEE Xplore. Restrictions apply.
8001113 IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 70, 2021

Fig. 12. Simulation results of detecting multitarget distributed in a certain


Fig. 11. CFAR loss against the false alarm probability. area.

TABLE I
these valid targets. The probability of targets existence in
S UMMARY OF D ETECTION R ESULTS FOR M ULTITARGET S CENARIO
the 1-D reference cells is much lower than that in 2-D ref-
erence cells. Meanwhile, multiple distributed 1-D CA-CFAR
detections decrease the detection probability of ghost targets,
and ROI-based processing improves the detection accuracy
of actual targets. In Table I, the results also verify that the
proposed scheme improves the performance of FoM by 14%
compared with the FFT plus circular scheme.
the FFT plus cross scheme have the same number of reference
cells and thus, they achieve the completely same CFAR loss F. Influence of Realistic Factors to Detection Performance
performance. In this section, influence of antenna beam, multipath, and
ground clutter to the detection performance are investigated.
E. Figure of Merit Analysis The antenna beamwidth/directivity significantly influences
the transmitter gain and receiver gain. As a result, it has
The figure of merit (FoM) parameter describes the ability obvious influence to the energy of a radar echo signal reflected
to detect valid targets and suppress ghost targets in multitarget by a target and could, therefore, cause varying signal-to-clutter
environments. It can be denoted by ratio (SCR). The varying SCR is then used to represent the
Nd −Nd f varying antenna beamwidth/directivity here.
figure of merit = (27)
Na + Nd f The ground clutter distribution resembles to the Weibull
distribution [33], which can be modeled as
where Nd , Nd f , and Na represent the number of detected valid
targets, detected ghost targets, and actual targets, respectively. a  x a−1   x a 
p(x) = exp − (28)
The number of missing actual targets is Nma , i.e., Nma = N a − b b b
Nd + Nd f . where b is the scale parameter (b > 0), a is the shape
To simulate the performance of FoM for the four schemes in parameter (a > 0) and indicates the degree of skewness, and
multitarget environment, we design the multitarget simulation x is the random clutter variable of the Weibull distribution.
scene: 60 targets are randomly distributed in a 1024 × The SCR is defined as E/C0 , where E represents the total
1024 RDM with an average background noise power of 30 dB; average energy of radar echo reflected by a target, and the
meanwhile, the SNR of each target is randomly set from 5 to clutter power C0 under Weibull distribution is given by
25 dB and the false alarm probability Pfa is set to 10−5 .
x m2 (1 + (2/a))
The detected targets and their distribution are demonstrated C0 = (29)
in Fig. 12. Based on the results in Fig. 12, statistical analysis (ln (2))2/a
is performed and the results are summarized in Table I. where x m is the median of Weibull distribution [34].
The proposed scheme exhibits the highest value of FoM By setting the number of reference cells Nr = 10, the
among the four schemes. In multitarget environment, some number of guard cells Ng = 6, and the desired Pfa = 10−3 ,
targets might locate in the reference cells of some other valid Fig. 13 indicates how the detection probability Pd varies
targets, which then results in increased estimated noise power according to the SCR with the combination of shape and
around the valid targets and degraded detection accuracy of scale parameters. Pd degrades with the increase of a and b,

Authorized licensed use limited to: INSTITUTO MILITAR DE ENGENHARIA. Downloaded on October 11,2022 at 21:08:40 UTC from IEEE Xplore. Restrictions apply.
ZHENG et al.: TARGET DETECTION SCHEME WITH DECREASED COMPLEXITY AND ENHANCED PERFORMANCE 8001113

Fig. 13. Detection probability versus SCR under various ground clutter.

Fig. 15. Recently developed 77-GHz FMCW radar system. (a) Block diagram
of system structure. (b) Photograph of the radar system.

Fig. 14. False alarm probability versus SCR under various multipath
influence.

which represents more severe ground clutter; meanwhile, Pd


improves with the increase of SCR, which suggests better
design and setting of the antenna beamwidth/directivity.
The multipath effect could produce ghost targets [35]. The
influence of multipath on the false alarm probability of the
proposed scheme is then investigated. The results in Fig. 14 are
obtained based on the settings of a = 1 and b = 1 for the
ground clutter, Nr = 10 and Ng = 6. The results demonstrate
that the multipath effect could increase the false alarm prob-
ability, and the influence is more obvious in high SNR level.

V. E XPERIMENTAL R ESULTS
To further verify the effectiveness of the proposed scheme
in multitarget environment, experiments are carried out using
our newly developed FMCW radar system. Fig. 16. Images of various modules. (a) Transceiver front-end module.
(b) Data acquisition module. (c) GUI implemented in LabVIEW.

A. Experimental Setup of system structure and photograph of the radar system


This section briefly describes the architecture of the newly are presented in Fig. 15(a) and (b), respectively. Mean-
designed 77-GHz FMCW radar prototype. The block diagram while, Fig. 16 shows the images of various modules of the

Authorized licensed use limited to: INSTITUTO MILITAR DE ENGENHARIA. Downloaded on October 11,2022 at 21:08:40 UTC from IEEE Xplore. Restrictions apply.
8001113 IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 70, 2021

Fig. 17. Configuration of the experimental scenario.

TABLE II
C ONFIGURATION OF THE R ADAR S YSTEM PARAMETERS

Fig. 18. Experimental results with N S = 256 and NC = 32. (a) Proposed
scheme. (b) FFT plus cross scheme. (c) FFT plus circular scheme. (d) FFT
plus rectangle scheme.

radar system. The radar system is composed of a 77-GHz


transceiver front-end module with antennas, a real-time data
acquisition module, and an uplink module (Host PC).
In Fig. 16(a), four silicon-integrated millimeter-wave
chips [36], containing the complete transmitter and receiver
chains, as well as the antenna array are mounted on the trans-
ceiver front-end module. The antenna array is a uniform linear
array (ULA) and consists of nine transmitter antennas (TXs)
and 16 receiver antennas (RXs). The integrated transmitter
chain generates the required FMCW chirp signal with 4-GHz
bandwidth, ranging from 77 to 81 GHz. In receiver chain,
the received signals are mixed with the transmitted chirp signal
and then passed through the low-pass filter (LPF) to produce
the beat signals, followed by the ADC with sampling rate Fig. 19. Experimental results with N S = 512 and NC = 64. (a) Proposed
of 37.5 Million Samples per Second (MSPS) for up to 15-MHz scheme. (b) FFT plus cross scheme. (c) FFT plus circular scheme. (d) FFT
intermediate frequency (IF) bandwidth [37]. The receiver noise plus rectangle scheme.
figure is 15 dB, and the transmitter power is 12 dBm [36].
In Fig. 16(b), the data acquisition module is implemented
15.5 and 19 m away from the radar sensor, respectively;
in a field-programmable gate array (FPGA) device, which is
three metal balls are used to design the multitarget scene: one
primarily responsible for capturing the sampling data from
(object #2) is placed 5.3 m away from the radar sensor with
transceiver front-end module and transfer the logged data
the radial velocity 0.8 m/s, and the other two stationary balls
to PC. Fig. 16(c) then shows the software designed and
(objects #1 and #3) are placed 1.8 and 3.5 m away from the
implemented in LabView, which realizes the functions of
radar sensor, respectively.
receiving data packets in real time and performing high-level
algorithms on the PC.
Fig. 17 shows the designed experimental scenario for carry- B. Experimental Analysis
ing out the experiment. The configuration parameters are pre- Figs. 18 and 19 show the detection results of the all
sented in Table II. In Fig. 17, a fixed iron lamppost (object #4) four schemes, in the case of {N S = 256 and NC = 32} and
and a fixed iron box (object #5) are located approximately {N S = 512 and NC = 64}, respectively.

Authorized licensed use limited to: INSTITUTO MILITAR DE ENGENHARIA. Downloaded on October 11,2022 at 21:08:40 UTC from IEEE Xplore. Restrictions apply.
ZHENG et al.: TARGET DETECTION SCHEME WITH DECREASED COMPLEXITY AND ENHANCED PERFORMANCE 8001113

TABLE III proposed scheme suffers decreased detections of ghost targets


S TATISTICAL A NALYSIS OF F O M P ERFORMANCE than the three conventional schemes. As a result, the proposed
scheme reaches the best FoM performance among the all four
schemes and increases the FoM parameters by 33.33% over
the FFT plus cross scheme, which obtains the highest FoM
among the three conventional ones.
According to the results in Fig. 19, using the proposed
scheme, the number of range ROI cells and that of Doppler
ROI cells are Nu = 10 and N P = 30, respectively. The proposed
scheme maintains the advantage with smaller detection number
of ghost targets over the three conventional ones; and thanks
to the zero detection of ghost targets, it achieves 100% for the
FoM. By comparing the results in Fig. 19 with that in Fig. 18,
we notice that with Nu and N P being increased, the detection
number of ghost targets is decreased. Consequently, the FoM
performance in Table III is improved.
Based on the above detection results in experiment, the com-
putational complexity is evaluated in terms of number of
multiplication and that of addition defined in (21) and (22),
for the all four schemes, and the results are summarized
in Fig. 20. The comparison results between Case A and
Case B in either of Fig. 20(a) or (b) show that, with the
number of range-FFT points and Doppler-FFT points being
increased, the computational complexities arise for the all four
schemes. As compared with the three conventional schemes,
the proposed scheme decreases the number of multiplications
by 42.78% and 55.68%, respectively, in Case A and Case B,
and the number of additions by 63.26% and 74.92% in
Case A and Case B, respectively. The comparison results
verify that the proposed scheme significantly decreases the
computational complexity, and the advantage will be further
enhanced with the increase of range-FFT and Doppler-FFT
points. We also notice that in Fig. 20(b), the addition number
of FFT plus rectangle scheme is slightly lower than the other
two conventional ones. The slight gap is caused owing to that
the FFT plus rectangle scheme has slightly smaller number of
reference cells than the other two conventional ones.

VI. C ONCLUSION
In this article, a novel target detection scheme based on
distributed 1-D CA-CFAR and ROI preprocessing is proposed
for the FMCW radar to decrease computational complexity
while maintaining robust performance in multitarget detec-
tions. By conducting the 1-D CA-CFAR detection for three
times, the proposed scheme fully utilizes the characteristics
of valid targets in both range domain and Doppler domain
Fig. 20. Computational complexity based on two metrics. (a) Number of
in the target detection; in addition, by conducting range and
multiplications (normalized). (b) Number of additions (normalized). Doppler ROI detection, it performs Doppler-FFT operations
only on the range ROI cells and scans only Doppler ROI
cells to extract pretargets. As a result, the proposed scheme
As shown in Fig. 18, the all four schemes successfully detect significantly decreases the computational complexity while
the parameters of five valid targets, which, respectively, are simultaneously enhancing the detection performance in the
(1.7589 m, 0 m/s), (3.5177 m, 0 m/s), (5.2766 m, 0.7813 m/s), multitarget environment, as compared with the conventional
(15.4390 m, 0 m/s), and (19.1521 m, 0 m/s). Using the schemes. Simulations and field tests are carried out and the
proposed scheme, the results of range ROI detection and both results verify the advantages of the proposed scheme
Doppler ROI detection are Nu = 7 and N P = 12, respectively. over the conventional ones, in terms of both the computational
According to the comparison results in Fig. 18(a)–(d), the complexity and detection performance.

Authorized licensed use limited to: INSTITUTO MILITAR DE ENGENHARIA. Downloaded on October 11,2022 at 21:08:40 UTC from IEEE Xplore. Restrictions apply.
8001113 IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 70, 2021

R EFERENCES [22] D. Zhang, Y. He, X. Gong, Y. Hu, Y. Chen, and B. Zeng, “Multitarget
AOA estimation using wideband LFMCW signal and two receiver
[1] Y. Z. Wang, Q. H. Liu, and A. E. Fathy, “CW and pulse-Doppler radar antennas,” IEEE Trans. Veh. Technol., vol. 67, no. 8, pp. 7101–7112,
processing based on FPGA for human sensing applications,” IEEE Trans. Aug. 2018.
Geosci. Remote Sens., vol. 51, no. 5, pp. 3097–3107, May 2013. [23] W. Wang, D. Liang, Z. Wang, H. Yu, and Q. Liu, “Design and imple-
[2] E. Hyun, Y.-S. Jin, and J.-H. Lee, “Moving and stationary target detec- mentation of a FPGA and DSP based MIMO radar imaging system,”
tion scheme using coherent integration and subtraction for automotive Radioengineering, vol. 24, no. 2, pp. 518–526, Jun. 2015.
FMCW radar systems,” in Proc. IEEE Radar Conf. (RadarConf), Seattle, [24] B.-S. Oh, X. Guo, F. Wan, K.-A. Toh, and Z. Lin, “Micro-Doppler mini-
WA, USA, May 2017, pp. 0476–0481. UAV classification using empirical-mode decomposition features,” IEEE
[3] F. Ali and M. Vossiek, “Detection of weak moving targets based on 2D Geosci. Remote Sens. Lett., vol. 15, no. 2, pp. 227–231, Feb. 2018.
range-Doppler FMCW radar Fourier processing,” in German Microw. [25] B. S. Kim, S. Kim, Y. Jin, and J. Lee, “Low-complexity joint range and
Conf. Dig. Papers, Berlin, Germany, 2010, pp. 214–217. Doppler FMCW radar algorithm based on number of targets,” Sensors,
[4] S. Saponara, M. S. Greco, and F. Gini, “Radar-on-chip/in-package in vol. 20, no. 1, pp. 51–64, Dec. 2019.
autonomous driving vehicles and intelligent transport systems: Oppor- [26] V. Winkler, “Range Doppler detection for automotive FMCW radars,”
tunities and challenges,” IEEE Signal Process. Mag., vol. 36, no. 5, in Proc. Eur. Microw. Conf., Munich, Germany, 2007, pp. 166–169.
pp. 71–84, Sep. 2019. [27] S. He and M. Torkelson, “Computing partial DFT for comb spectrum
[5] F. Pieri, C. Zambelli, A. Nannini, P. Olivo, and S. Saponara, “Is evaluation,” IEEE Signal Process. Lett., vol. 3, no. 6, pp. 173–175,
consumer electronics redesigning our cars?: Challenges of integrated Jun. 1996.
technologies for sensing, computing, and storage,” IEEE Consum. Elec- [28] M. Kronauge and H. Rohling, “Fast two-dimensional CFAR procedure,”
tron. Mag., vol. 7, no. 5, pp. 8–17, Sep. 2018. IEEE Trans. Aerosp. Electron. Syst., vol. 49, no. 3, pp. 1817–1823,
[6] Z. Peng et al., “A portable FMCW interferometry radar with program- Jul. 2013.
mable low-IF architecture for localization, ISAR imaging, and vital [29] H. Rohling, “Radar CFAR thresholding in clutter and multiple target
sign tracking,” IEEE Trans. Microw. Theory Techn., vol. 65, no. 4, situations,” IEEE Trans. Aerosp. Electron. Syst., vol. AES-19, no. 4,
pp. 1334–1344, Apr. 2017. pp. 608–621, Jul. 1983.
[7] S. Saponara and B. Neri, “Radar sensor signal acquisition and multi- [30] M. B. El Mashade, “Performance of novel versions of CFAR detection
dimensional FFT processing for surveillance applications in transport schemes processing M-Correlated sweeps in presence of interferers,”
systems,” IEEE Trans. Instrum. Meas., vol. 66, no. 4, pp. 604–615, Radioelectron. Commun. Syst., vol. 62, no. 4, pp. 143–160, Apr. 2019.
Apr. 2017. [31] E. Grossi, M. Lops, and L. Venturino, “Detection rate optimization
[8] E. Hyun and Y. S. Jin, “Human-vehicle classification scheme using for Swerling 0, I, and III target models,” in Proc. IEEE Radar Conf.
Doppler spectrum distribution based on 2D range-Doppler FMCW (RadarConf), Seattle, WA, USA, May 2017, pp. 606–609.
radar,” J. Intell. Fuzzy Syst., vol. 35, no. 6, pp. 6035–6045, Dec. 2018. [32] M. Maynul and M. Hossam-E-Haider, “Detection capability and CFAR
[9] S. M. Patole, M. Torlak, D. Wang, and M. Ali, “Automotive radars: loss under fluctuating targets of different Swerling model for various
A review of signal processing techniques,” IEEE Signal Process. Mag., gamma parameters in RADAR,” Int. J. Adv. Comput. Sci. Appl., vol. 9,
vol. 34, no. 2, pp. 22–35, Mar. 2017. no. 2, pp. 90–93, 2018.
[10] M. Baadeche and F. Soltani, “Performance analysis of mean level [33] P. Surendran, S. J. Ko, C. U. Kang, and J. H. Lee, “Performance of
constant false alarm rate detectors with binary integration in Weibull UWB short range radar in Weibull clutter environment,” in Proc. 9th
background,” IET Radar, Sonar Navigat., vol. 9, no. 3, pp. 233–240, Eur. Radar Conf., Amsterdam, The Netherlands, 2012, pp. 194–197.
Mar. 2015. [34] G. Minkler and J. Minkler, CFAR: The Principles of Automatic Radar
[11] E. Hyun, Y.-S. Jin, and J.-H. Lee, “A pedestrian detection scheme using Detection in Clutter. Scottsdale, AZ, USA: Magellan Book Company,
a coherent phase difference method based on 2D range-Doppler FMCW 1990.
radar,” Sensors, vol. 16, no. 1, p. 124, Jan. 2016. [35] A. Kamann, P. Held, F. Perras, P. Zaumseil, T. Brandmeier, and
[12] E. Hyun, Y. S. Jin, and J. H. Lee, “Design and development of auto- U. T. Schwarz, “Automotive radar multipath propagation in uncertain
motive blind spot detection radar system based on ROI pre-processing environments,” in Proc. 21st Int. Conf. Intell. Transp. Syst. (ITSC), Maui,
scheme,” Int. J. Automot. Technol., vol. 18, no. 1, pp. 165–177, HI, USA, Nov. 2018, pp. 859–864.
Oct. 2016. [36] AWR1243 Single-Chip 77- and 79-GHz FMCW Transceiver, 3rd ed.,
[13] B.-S. Kim, Y. Jin, S. Kim, and J. Lee, “A low-complexity FMCW Texas Instrum., Dallas, TX, USA, 2017.
surveillance radar algorithm using two random beat signals,” Sensors, [37] AWR1243 Sensor: Highly Integrated 76-81-GHz Radar Front-End for
vol. 19, no. 3, pp. 608–624, Jan. 2019. Emerging ADAS Applications, 3rd ed., Texas Instrum., Dallas, TX, USA,
[14] F. D. A. Garcia, A. C. F. Rodriguez, G. Fraidenraich, and J. Filho, 2017.
“CA-CFAR detection performance in homogeneous Weibull clutter,”
IEEE Geosci. Remote Sens. Lett., vol. 16, no. 6, pp. 887–891,
Jun. 2019.
[15] A. Melebari, A. K. Mishra, and M. Y. Abdul Gaffar, “Comparison
of square law, linear and Bessel detectors for CA and OS CFAR Qiangwen Zheng was born in Anhui, China,
algorithms,” in Proc. IEEE Radar Conf., Johannesburg, South Africa, in 1994. He received the B.S. degree in microelec-
2015, pp. 383–388. tronics science and engineering from Anhui Univer-
[16] X. Yang, L. Huo, J. Su, X. Zhang, and W. Jiang, “An anti- sity, Hefei, China, in 2017. He is currently pursuing
FOD method based on CA-CM-CFAR for MMW radar in com- the M.S. degree in marine electronic information
plex clutter background,” Sensors, vol. 20, no. 6, pp. 1635–1654, engineering with Zhejiang University, Hangzhou,
Mar. 2020. China.
[17] K. El-Darymli, P. McGuire, D. Power, and C. Moloneyb, “Target His research interests include millimeter-wave
detection in synthetic aperture radar imagery: A state-of-the-art sur- radar system and its signal processing.
vey,” J. Appl. Remote Sens., vol. 7, no. 1, pp. 071598-1–071598-35,
Mar. 2013.
[18] C. Xu, Y. Li, C. Ji, Y. Huang, H. Wang, and Y. Xia, “An improved
CFAR algorithm for target detection,” in Proc. ISPACS, Xiamen, China,
2017, pp. 883–888. Lijie Yang was born in Zhejiang, China, in 1991. He
[19] C. Yardim, P. Gerstoft, and W. S. Hodgkiss, “Tracking refractivity from received the B.S. and Ph.D. degrees in electronic and
clutter using Kalman and particle filters,” IEEE Trans. Antennas Propag., information engineering from Zhejiang University,
vol. 56, no. 4, pp. 1058–1070, Apr. 2008. Hangzhou, China, in 2013 and 2019, respectively.
[20] W. Wang, R. Wang, R. Jiang, H. Yang, and X. Wang, “Modified He is currently an Assistant Researcher with the
reference window for two-dimensional CFAR in radar target detection,” Zhejiang Lab, Hangzhou. His current research inter-
J. Eng., vol. 2019, no. 21, pp. 7924–7927, Nov. 2019. ests include the design of high-resolution 4-D imag-
[21] F. Roos, M. Sadeghi, J. Bechter, N. Appenrodt, J. Dickmann, and ing radar, classification and tracking algorithm for
C. Waldschmidt, “Ghost target identification by analysis of the Doppler traffic participants, and multisensor fusion algorithm
distribution in automotive scenarios,” in Proc. 18th Int. Radar Symp. for ADAS.
(IRS), Jun. 2017, pp. 1–9.

Authorized licensed use limited to: INSTITUTO MILITAR DE ENGENHARIA. Downloaded on October 11,2022 at 21:08:40 UTC from IEEE Xplore. Restrictions apply.
ZHENG et al.: TARGET DETECTION SCHEME WITH DECREASED COMPLEXITY AND ENHANCED PERFORMANCE 8001113

Yaping Xie was born in Henan, China, in 1993. Chunyi Song (Member, IEEE) received the Ph.D.
He received the B.S. degree in electronic and infor- degree in electronic and communication engineering
mation engineering from the Henan University of from Waseda University, Tokyo, Japan, in 2009.
Technology, Zhengzhou, China, in 2012 and 2016, He was a Research Associate with Waseda Univer-
respectively. He is currently pursuing the M.S. sity, from 2007 to 2009. He was a Researcher with
degree with Zhejiang University, Hangzhou, China. the National Institute of Information and Commu-
His main interest is millimeter-wave radar systems. nications Technology (NICT), Koganei, Japan, from
2009 to 2013, and as a Senior Researcher, in 2014.
Since 2014, he has been with Zhejiang University,
Hangzhou, China, as an Associate Professor, where
he is also serving as the Vice Director of the
Engineering Research Center of Oceanic Sensing Technology and Equipment,
Junjie Li was born in Guangdong, China, in 1993. Ministry of Education.
He received the B.S. degree in naval architecture and Dr. Song was elected to the Distinguished Expert, Zhejiang, in 2016, and a
marine engineering from Dalian Maritime Univer- Core Member of the Leading Innovative Team, Zhejiang, in 2018.
sity, Dalian, China, in 2016. He is currently pursuing
the Ph.D. degree in marine information science and
engineering with Zhejiang University, Hangzhou,
China.
His research interests include passive radar and
millimeter-wave radar signal processing.

Tang Hu was born in Zhejiang, China, in 1979.


He received the B.S. degree in measurement and Zhiwei Xu (Senior Member, IEEE) received the
control technology and instruments from Yanshan Ph.D. degree in integrated circuits and systems from
University, Qinhuangdao, China, in 2002, and the the University of California at Los Angeles, Los
M.S. degree from the Nanjing University of Aero- Angeles, CA, USA, in 2003.
nautics and Astronautics, Nanjing, China, in 2005. He held senior manager positions with G-Plus Inc.,
He is currently a Senior Engineer with the Zhe- Los Angeles, SST Communications, Los Angeles,
jiang Lab, Hangzhou, China. His research interest Conexant Systems, Irvine, CA, USA, and NXP Inc.,
includes millimeter-wave radar chip and system. San Diego, CA, USA, before he joined the HRL
Laboratories, Los Angeles, where he led develop-
ment for wireless LAN and SoC solutions for propri-
etary wireless multimedia systems, CMOS cellular
Jiang Zhu (Member, IEEE) received the B.E. degree transceiver, multimedia over cable (MoCA) system, and TV tuners. He is
in electronic science and technology from Harbin currently a Professor with Zhejiang University, Hangzhou, China, with a
Engineering University, Harbin, China, in 2011, and focus on cognitive radios, high-speed analog-to-digital conversion (ADC),
the Ph.D. degree in information and communica- and mmWave ICs. He has authored or co-authored 70 publications, one
tion engineering from Tsinghua University, Beijing, contribution to the encyclopedia of wireless and mobile communications, ten
China, in 2016. granted patents, and around 20 pending ones.
Dr. Xu was elected to the Distinguished Expert, China, in 2014, and the
He was a Lecturer with the Ocean College,
Zhejiang University, Hangzhou, China. His current Team Leader of the Leading Innovative Team, Zhejiang, in 2018. He serves the
research interests include statistical signal process- technical program committee for international workshop on millimeter-wave
ing, Bayesian methods, and unlabeled sensing. communications: from circuit to network and as a Series Editor for IEEE
Communications Magazine.

Authorized licensed use limited to: INSTITUTO MILITAR DE ENGENHARIA. Downloaded on October 11,2022 at 21:08:40 UTC from IEEE Xplore. Restrictions apply.

You might also like