ADVANCES IN RADAR SYSTEMS FOR MODERN

CIVILIAN AND COMMERCIAL APPLICATIONS: PART 2

Gor Hakobyan and Bin Yang

High-Performance Automotive Radar

A review of signal processing algorithms and modulation schemes

T
he ongoing automation of driving functions in cars results in
the evolution of advanced driver assistance systems (ADAS)
into ones capable of highly automated driving, which will in
turn progress into fully autonomous, self-driving cars. To work
properly, these functions first must be able to perceive the car’s
surroundings by such means as radar, lidar, camera, and ultra-
sound sensors. As the complexity of such systems increases
along with the level of automation, the demands on environment
sensors, including radar, grow as well. For radar performance
to meet the requirements of self-driving cars, straightforward
scaling of the radar parameters is not sufficient. To refine radar
capabilities to meet more stringent requirements, fundamentally
different approaches may be required, including the use of more
sophisticated signal processing algorithms as well as alternative
radar waveforms and modulation schemes. In addition, since
radar is an active sensor (i.e., it operates by transmitting signals
and evaluating their reflections) interference becomes a crucial
issue as the number of automotive radar sensors increases. This
article gives an overview of the challenges that arise for auto-
motive radar from its development as a sensor for ADAS to a
core component of self-driving cars. It summarizes the relevant
research and discusses the following topics related to high-
performance automotive radar systems: 1) shortcomings of the
classical signal processing algorithms due to underlying fun-
©ISTOCKPHOTO.COM/TALAJ
damental assumptions and a signal processing framework that
overcomes these limitations, 2) use of digital modulations for
automotive radar, and 3) interference-mitigation methods that
enable multiple radar sensors to coexist in conditions of increas-
ing market penetration. The overview presented in this article
shows that new paradigms arise as automotive radar transitions
into a more powerful vehicular sensor, which provides a fertile
research ground for further investigation.

Introduction
When the idea of radar was first explored back in the late-19th
and early-20th centuries, it was primarily seen as a technology
for military applications. Other applications gradually emerged,
Digital Object Identifier 10.1109/MSP.2019.2911722
Date of publication: 9 September 2019 however, and in the last four decades, radar has been studied for

32 IEEE SIGNAL PROCESSING MAGAZINE | September 2019 | 1053-5888/19©2019IEEE

use in the automotive sector for such applications as predictive beat frequency predominates, and their Doppler shift, i.e., the
crash sensing, obstacle detection, and braking [1]. The term ra- velocity component, is negligible (hence the name fast chirp).
dar is short for radio detection and ranging, an indication that Figure 1(a) illustrates a sequence of identical FMCW chirps,
radar is used to detect objects (obstacles and other road users) the delayed and Doppler-shifted reflections of which after
near the vehicle and to estimate their range as well as velocity mixing with the Tx signal result in 2D complex exponentials
and angle relative to the radar. For many years, production cars in the baseband. A subsequent 2D Fourier transform yields the
have made use of these capabilities to facilitate various driver- distance–velocity radar image. To localize targets in space, tar-
assistance functions, such as emergency brake assist and adap- get angles are measured based on direction of arrival (DOA)
tive cruise control. More complex functions, such as fully au- of reflected signals with array processing techniques, most
tonomous driving, also rely heavily on radar as an environmental commonly via digital beamforming. Figure 1(b) shows the
sensor [2], as it is capable of direct range and velocity measure- DOA-induced phase differences at receive (Rx) channels and
ments, can sense long distances ahead, is robust to bad weather the principle of digital beamforming that combines Rx signals
and poor light conditions, and can be hidden behind a bumper. with phases that digitally direct the beam to a certain DOA.
A detailed overview of the status of automotive radar during In the simplest case, all three frequency-estimation tasks are
its first several years is presented in [1]. The evolution of auto- solved jointly by a 3D Fourier transform, followed by power
motive radar is discussed in [3]. Other review articles provide detection [10], parameter estimation, clustering and associa-
overviews of the signal processing architecture and of the mil- tion of reflexes [11], object classification and tracking [12],
limeter-wave technology for automotive radar [4], [5]. A more data fusion [13], and other calculations.
recent review article discusses the state-of-the-art signal pro- Typical frequency bands for automotive radar are 24
cessing algorithms for automotive radar and gives a bird’s-eye and 77 GHz, with most of the manufacturers shifting toward
view of estimation techniques, radar waveforms, and higher- 77 GHz for newer radar generations. This is due to larger avail-
level processing steps, such as tracking and classification [6]. able bandwidth (76–77 GHz for long-range and 77–81 GHz for
This article gives an overview of the signal processing and short-range applications), higher Doppler sensitivity (and thus
modulation aspects of high-end automotive radar systems and higher velocity resolution), and smaller antennas.
discusses recent advances in these fields. We address the use Automotive radar performance is measured according
of digital modulations, such as orthogonal frequency-division to the following main parameters: 1) resolution (ability to
multiplexing (OFDM) and phase modulated continuous wave separate two closely spaced targets), 2) unambiguously mea-
(PMCW) waveforms, for automotive radar and multiple-input, surable range (the range of parameter values that are unam-
multiple-output (MIMO) radar in particular; discuss their poten- biguously distinguishable), and 3) dynamic range (power
tial benefits and challenges due to increased complexity; and ratio between the strongest and the weakest of detectable tar-
survey recent research in this area. We also point out that classi- gets) in its measurement dimensions, i.e., distance, velocity,
cal automotive radar signal processing does not fully accommo- azimuth, and elevation angle.
date performance improvement through simple upscaling of the For conventional Fourier-based signal processing, the radar
radar parameters (e.g., bandwidth, measurement time, antenna resolution and unambiguous range for all of the aforemen-
aperture) due to underlying fundamental assumptions. We pro- tioned measurement parameters are directly determined by the
vide a signal processing framework based on a more advanced sampling frequency and observation length in the correspond-
signal model that surpasses these limits at a feasible compu- ing dimension. For distance, the observation length is given by
tational cost. Next we explore the reliable operation of future the bandwidth B, and its inverse determines the resolution with
automotive radar systems for which interference mitigation is which the round-trip delays x = 2d/c 0 are measured, with d
vital and complete the discussion with a survey of interference- being the target distance and c 0 being the speed of light. Thus,
mitigation methods. These include some promising paradigms, the distance resolution is given by the following bandwidth:
such as interference-aware cognitive radar [7] and centralized Dd = c 0 /(2B) . Analogously, in the velocity dimension the
coordination for interference avoidance [8]. Doppler resolution DfD is determined by the inverse of the
measurement cycle duration Tcycle, i.e., TfD = 1/Tcycle . With
Conventional automotive radar TfD = 2Tvfc /c 0, the velocity resolution is Tv = c 0 /(2fc Tcycle) .
Today, conventional automotive radar operates with a se- For DOA-induced spatial frequencies, the resolution can be
quence of frequency-modulated continuous wave (FMCW) derived analogously from the dimensions of the antenna array
signals and is a well-studied research field [6], [9]. These sys- [14]. The previous discussion makes apparent that regardless of
tems transmit a series of analog-generated chirps, which are the radar waveform, a large bandwidth and a long measurement
reflected and then mixed with the transmit (Tx) chirp at the time are required for a high distance and velocity resolution.
receiver, resulting in a frequency proportional to the target dis- To obtain a resolution higher than that of the conventional
tance and called beat frequency. The range processing is based Fourier processing, superresolution frequency estimation
on Fourier transform of the beat frequencies, and the Doppler- methods can be applied in different radar measurement dimen-
induced phase progression over the consecutive chirps is used sions. Such methods can be coarsely classified into subspace-
for velocity estimation. Chirps are commonly designed to be based, maximum-likelihood, or compressed-sensing methods.
short enough so that the distance-induced component of the A review of high-resolution methods for array processing and

IEEE SIGNAL PROCESSING MAGAZINE | September 2019 | 33

 for multidimensional automotive radar processing can be found due to analog-to-digital converters (ADCs) and significantly
in [14] and [15], respectively. An overview of compressed-sens- larger data loads—this opens up new dimensions for radar de-
ing applications for radar is given in [16]. velopment and enables advanced radar concepts.

Digital radar OFDM radar

In parallel to the described fast-chirp radar, alternative auto- The OFDM waveform is composed of a set of orthogonal
motive radar concepts based on digital modulations, such as complex exponentials [subcarriers; see the left-hand side of
OFDM [17] and PMCW [18], have been studied over the past Figure 2(a)], the complex amplitudes of which are modulated
few years. These concepts differ from FMCW radar in terms with communication data or radar modulation symbols. The
of generating waveforms digitally and performing demodula- orthogonality of subcarriers results from the constraint of all
tion in the digital domain. Broadly speaking, this is equivalent subcarriers having a whole number of periods during one evalu-
to operating with arbitrary digitally generated waveforms and ation interval, called an OFDM symbol [see the right-hand side
matched filter-based processing at the receiver. For OFDM of Figure 2(a)]. As the discrete Fourier transform (DFT) exhibits
­radar, this large degree of flexibility in the waveform choice the same characteristics, OFDM waveforms can be efficiently
enables communication and radar capabilities to be combined generated via inverse fast Fourier transform (IFFT) of the
by embedding communication information into the radar modulation symbols, i.e., complex amplitudes of OFDM sub-
waveform [17]. It further enables fully adaptive, software-de- carriers. Conversely, the communication data or radar modu-
fined behavior based on digitally generated waveforms. While lation symbols can be efficiently extracted (demodulated) at
more challenging in terms of practical realization—mainly the receiver based on FFT. From the communication standpoint,

fTx
Tcycle tf
Rx

2D-FFT
d
B fc ... Tx Baseband

1/fbeat
ts v
Tch
0 1/fD
t
(a)
TCRI
DBF Beam

Target
α

Wavefront

α
φ3
φ2
φ1 Downconversion and Sampling

α Digital Beamforming
Tx Antenna
2)
t)

1)

3)
(φ
y(

(φ

(φ
a
a

Y (α)
t)
t)

t)
y(
y(

y(

(b)

FIGURE 1. Graphs and illustrations showing the principle of distance, velocity, and DOA estimation for conventional fast-chirp automotive radar. (a) A
sequence of identical FMCW chirps. The delayed and Doppler-shifted reflections of such chirps, after mixing with the Tx signal, result in 2D complex
exponentials in the baseband. (b) The DOA-induced phase differences at Rx channels and the principle of digital beamforming that combines Rx signals
with phases that digitally direct the beam to a certain DOA. B: bandwidth; fc: carrier frequency; TCRI: chirp repetition interval; Tch: chirp duration; fbeat: beat
frequency; fD: Doppler shift; t s: slow-time; d: distance; v : velocity; a: target angle; y (z n) received signal with a phase shift z n at the nth Rx antenna.

34 IEEE SIGNAL PROCESSING MAGAZINE | September 2019 |

this achieves high spectral efficiency as well as simple ex- thus time-of-flight delays result in cyclic shifts of OFDM sym-
traction of communication data. Meanwhile, from the radar bols at the receiver. The block diagram in Figure 2(b) depicts
standpoint, it enables efficient digital demodulation of the ra- the structure of the OFDM system. The OFDM symbols gener-
dar waveform. OFDM not only enables favorable modulation ated via IFFT are shifted into the radio-frequency (RF) band
for both applications, but it also combines both functionalities via a quadrature modulation and transmitted over the channel.
via a single waveform. This initially motivated research on From the perspective of radar, the channel represents objects in
OFDM radar. Currently, OFDM is often studied as a means the vehicle’s surroundings, i.e., the driving environment. At the
for efficient implementation of digital, software-defined receiver, the CP is removed from the quadrature demodulated
­radar—independent of the communication aspect. signal, and the complex modulation symbols are obtained via
To prevent interference between consecutive OFDM sym- an FFT. For OFDM radar signal processing illustrated in Fig-
bols in a multipath channel, a cyclic prefix (CP) that contains ure 2(c), the subcarrier values of consecutive OFDM symbols
repetition of the end portion of OFDM symbol is transmitted are placed into a 2D measurement matrix. The radar waveform
before the symbol [see the right-hand side of Figure 2(a)]. This is demodulated based on spectral division, which cancels out
converts the linear convolutive channel into a cyclic one, and the transmitted complex modulation symbols by elementwise

Power
∆f
CP OFDM Symbol

X (f ) x (t )
TCP T
Frequency TOFDM

0 f1 f2 ...
(a)
cos(2πfct)

S X(0) P Re DAC
X(1) x(m) Re[x (t )]
X(n) .
. IFFT . CP
. .
. Im[x (t )]
X(Nc – 1)
P S Im DAC

sin(2πfct)
Channel/
Environment
cos(2πfct)

yf(0) ADC
P S
yf(1) y(m) Re[y (t )]
.
yf(n) . FFT . –CP
. .
. Im[y (t )]
yf(Nc – 1)
S P ADC

–j sin(2πfct)
(b)
ytf ,ts yf,ts zf,ts zf,v zd,v

FFT S.D. FFT IFFT

tf f f f d
...

Time
ts ts ts v v
(c)

FIGURE 2. Illustrations showing the OFDM radar principle. (a) On the left, the OFDM spectrum and its inverse Fourier transform resulting in a time-
domain OFDM symbol on the right. (b) The block diagram of the OFDM system. (c) The signal processing steps of OFDM radar. S.D.: spectral division;
Re, real, Im, imaginary; P/S, the parallel-to-serial blocks; S/P, the serial-to-parallel blocks.

IEEE SIGNAL PROCESSING MAGAZINE | September 2019 | 35

 ­ ultiplication with their inverse values (conjugate in case of unitary
m ing the instantaneous bandwidth, and thus the sampling rates.
subcarrier amplitudes). This operation reduces the measurement Whereas [20] covers a larger bandwidth by sweeping in multi-
matrix to a sum of 2D complex exponentials, the frequencies of ple steps (called stepped OFDM), [7] and [21] combine OFDM
which over the OFDM subcarriers and symbols correspond to waveform with a chirp to increase the effective bandwidth. By
the distances and velocities of the radar targets. Similar to fast- randomly occupying smaller portions of the full bandwidth at
chirp radar, a 2D-FFT processing (IFFT over subcarriers, FFT each time instance, [22] aims to reduce the sampling rates of
over symbols) leads to the distance–velocity radar image. OFDM radar with a compressed-sensing approach.
We make the following observations regarding the OFDM
radar signal processing. PMCW radar
■■ For unitary subcarrier amplitudes, the distance processing An alternative implementation of digital radar uses a sequence
is equivalent to matched filtering implemented efficiently of waveforms generated by phase modulation of continuous
in the frequency domain. waves [18]. The waveform generation via biphase modulation
■■ The described signal processing neglects the Doppler shift of the RF-carrier signal with 0° and 180° phase shifts is par-
of OFDM subcarriers, which might lead to intercarrier ticularly simple to implement in CMOS technology [18]. At the
interference (ICI). To limit ICI to a negligible level, the receiver, a bank of digitally implemented correlators is used for
subcarrier spacing Tf must be much larger than the maxi- range processing. The Doppler processing is done via an FFT
mum possible Doppler shift fD,max, e.g., Tf = 10fD,max over a sequence of consecutive coded waveforms, analogous
[17]. This limits, however, the parametrization freedom, to fast-chirp or OFDM radars. The block diagram of a PMCW
especially for long-range and highly dynamic applications, radar is shown in Figure 3. For favorable autocorrelation prop-
such as front long-range automotive radar. erties and thus high dynamic range in range estimation, the se-
■■ Under conditions of unitary subcarrier amplitudes and neg- lection of a proper code sequence is essential. Using orthogonal
ligible Doppler shift, the waveform has no influence on the codes, multiple Tx channels can operate simultaneously based
signal processing performance. Thus, it can carry commu- on code-domain separation, allowing MIMO processing. Fur-
nication data or be optimized with respect to peak-to-aver- thermore, a meaningful code selection can provide favorable
age power ratio for radar (e.g., [19]). properties in terms of robustness against interference. As for
■■ The distance and velocity processing is done in two inde- OFDM radar, Doppler shift has an adverse effect on PMCW
pendent dimensions and no coupling between them is con- waveforms in terms of auto- and cross-correlation properties,
sidered. Since the target velocity in practice affects both and needs to be accounted for by parametrization, code choice,
measurement dimensions, this can be interpreted as simpli- or compensation in signal processing.
fication of the 2D matched filtering into two separate
­one-dimensional matched filters, one per each measure- Discussion of modulation schemes
ment dimension. Analogous to fast-chirp radar, this ignores Fast-chirp, OFDM, and PMCW radars share the same principle
the range change for moving targets, and thus assumes all of distance–velocity measurement: time-of-flight-based coher-
OFDM symbols to have the same delay. ent distance estimation via pulse compression (fast-time) and
Since OFDM radar demodulates the radar waveform in the Doppler-based velocity estimation via FFT over a series of con-
digital domain, the entire signal bandwidth needs to be sampled, secutive waveforms (slow-time). For all three systems, the reso-
contrary to fast-chirp radar that samples only the bandwidth of lution and unambiguous range depend solely on observation
beat frequencies. This makes the practical realization of OFDM length (i.e., bandwidth in fast-time and measurement time in
radar more challenging, imposing high demands on ADCs, slow-time) and sampling rate (ADC rate in fast-time and wave-
memory, and digital signal processing. Some of the ongoing form repetition rate in slow-time). In terms of hardware effort,
research in [7] and [20]–[22] focuses on methods for limit- fast-chirp radar has an advantage due to analog mixing, i.e.,

PRN
Code
Tx
Lc Parallel Lc Parallel Lc Parallel
S

CFAR
Integrate Detection
Accumulate N-Point
ADC Lc Pulses -
M Pulses DFT
Rx DOA
Estimation
PRN Code

FIGURE 3. A schematic view of PMCW radar. The carrier signal is modulated with a pseudorandom noise (PRN) code. The distance processing is based
on L c digital correlators, followed by DFT-based Doppler processing [18]. CFAR: constant false alarm rate.

36 IEEE SIGNAL PROCESSING MAGAZINE | September 2019 |

demodulation, resulting in significantly reduced sampling rates tional TDM or FDM, e.g., more efficient use of time–fre-
for beat frequencies. In contrast, digital radar requires sampling quency resources.
of the entire bandwidth, i.e., higher ADC rates, memory, and ■■ Slope diversity multiplexing: Multiple Tx channels trans-
computational demands. The software-defined capabilities of mit chirps of the same bandwidth but different duration,
such radar allow, however, a substantially larger flexibility in i.e., slope. The signal of the nth Tx antenna is then
operation. As automotive radars become increasingly complex, x n = exp( jr [K + TK n] t 2), with TK n being the slope dif-
digital radar with software-defined modulation allows more ference to the base chirp slope K. At the receiver, the radar
features with respect to adaptive and multifunction behavior, signal is demodulated with multiple slopes, each of the sig-
advantageous MIMO concepts, and robustness against interfer- nals resulting in beat frequency for the corresponding Tx
ence based on large waveform diversity. channel and chirp for other channels. The subsequent FFT
processing focuses the signals with constant beat frequency
MIMO radar and spreads the remaining chirp signals. This achieves a
The use of MIMO radar techniques is a well-established approach separation between Tx channels, albeit in a nonorthogonal
for improved angle estimation with radar [23]. MIMO radar uses manner and thus with limited dynamic range.
multiple channels at both the Tx and the Rx sides such that the ■■ Slow-time phase modulation based multiplexing: The
number of paths between the radar and the target is efficiently phase of each chirp (or any waveform in general) in
increased. That is, with the number of paths being the product of slow-time is modulated to multiplex Tx channels. That is,
the number of Tx and Rx channels, MIMO radar obtains more for the nth Tx antenna, the phase over slow-time is modu-
paths than the number of physical channels. These paths can be lated with exp( j2rC n (t s)). When modulated with a com-
arranged into a larger virtual aperture with more elements, as de- plex exponential (linear phase progression C n (t s) = fs t s),
picted in Figure 4, and thus improve the angular resolution and this leads to a Doppler offset between the Tx channels.
estimation accuracy of the radar. The resulting virtual aperture This is advantageous in applications where the maximum
can then be processed with conventional array processing tech- possible Doppler shift is smaller than the unambiguously
niques. The main challenge for MIMO radar is thus the choice of measurable Doppler range (e.g., at lower carrier frequen-
waveforms such that the signals from different Tx antennas can cies). Alternatively, slow-time phases can be modulated
be clearly distinguished, i.e., the multiplexing of the Tx channels. with orthogonal codes to multiplex Tx channels. This
For high-performance automotive radar, efficient multiplexing of requires Doppler-robust codes, which typically exhibit lim-
a large Tx array is a key factor for achieving a high angular resolu- ited dynamic range in the velocity estimation.
tion in both azimuth and elevation. Conventionally, Tx antennas Due to its multicarrier structure, OFDM radar allows even
are multiplexed in time [24], frequency [25], or code [26]. more freedom with respect to multiplexing for MIMO radar.
Because of its simplicity, time-division multiplexing (TDM) OFDM subcarriers can be individually assigned to a Tx anten-
with equidistantly interleaved chirps is the most common multi- na, which enables the generation of various orthogonal wave-
plexing technique for fast-chirp radar. However, this approach forms for MIMO radar. With this approach, Tx antennas can
allows only one antenna to be active at a time, and thus limits con- operate simultaneously using the entire bandwidth. OFDM-
siderably the number of Tx antennas that can be multiplexed. Some specific multiplexing methods include the following:
more advanced multiplexing methods for
fast-chirp radar include the following [6]:
■■ Beat frequency multiplexing: Chirps
of multiple Tx channels run parallel MIMO Antenna Array
with an offset in time and/or fre-
quency, such that the beat frequen- φ3
φ φ2
cies of different Tx channels appear 4
φ1
as frequency division multiplexed
(FDM). Let us denote the base chirp
x 0 = exp( jrKt 2) with K being the
chirp slope. For chirps with frequency Tx Antennas Rx Antennas
offset, the signal of the nth Tx channel Virtual Antenna Array
is x n = exp( jr [Kt 2 + 2nTfTx t]), the
An Virt d
na l
ten ua
e

offset TfTx being larger than the

Wi eplac

φ3
maximum beat frequency fbeat,max . φ4 φ2
th
R

Analogously, for chirps offset in time, φ1

the signal of the nth Tx channel is
x n = exp( j2rK [t - nTx Tx] 2), with
Tx Tx being larger than the maximum Tx Antennas Virtual Rx Antennas Rx Antennas
round-trip delay max This allows
x .
tighter spacing of chirps than conven- FIGURE 4. A diagram showing the principle of MIMO radar and virtual apertures.

IEEE SIGNAL PROCESSING MAGAZINE | September 2019 | 37

 ■■ Equidistant subcarrier interleaving [17]: Subcarriers of channel from a reference frequency fc . In case one of these
OFDM radar are interleaved equidistantly over multiple phase components becomes dominant (e.g., for FDM with
Tx antennas (every N Txth subcarrier is assigned to one of very large frequency offsets), the MIMO-based DOA process-
the N Tx Tx antennas), such that all Tx channels use the ing becomes impractical. The coherency aspect of MIMO
entire bandwidth simultaneously. While maintaining the processing favors subcarrier interleaving schemes of OFDM
distance resolution, this reduces the unambiguously mea- radar, as they enable simultaneous transmission from all Tx
surable distance range, as the spacing between subcarriers channels with identical or very close carrier frequencies.
transmitted from one Tx antenna increases from Tf to
N Tx Tf (i.e., the sampling rate of distance-induced com- Limits of conventional range-Doppler processing
plex exponentials decreases). This method is thus less For both fast-chirp and digital radars, the conventional automo-
suitable for long-range applications. tive radar signal processing assumes that problems estimating the
■■ Nonequidistant subcarrier interleaving [27]: To overcome range (distance), velocity, and angle can be solved by processing
the drawback of equidistant subcarrier interleaving in the following three independent measurement dimensions (ig-
terms of reduced unambiguously measurable distance noring elevation for simplicity of discussion): 1) fast-time (single
range, the OFDM subcarriers can be interleaved nonequi- chirp, OFDM, or PMCW symbols), 2) slow-time (consecutive
distantly. This implies a nonuniform sampling of distance- waveforms), and 3) spatial domain (array elements). These mea-
induced complex exponentials that maintains unambiguous surement dimensions are, however, not entirely independent. For
distance range for each Tx channel. Since for nonuniform range-Doppler processing in particular, the range of the moving
sampling, FFT-based processing leads to increased side- target changes over consecutive waveforms, and may thus lead to
lobes, nonequidistant subcarrier interleaving requires more a migration of the target peak between range cells over slow-time,
complex distance processing, e.g., based on compressed i.e., range migration [7], [29].
sensing. The nonequidistant subcarrier interleaving can be Similarly, the Doppler processing of the conventional auto-
kept the same in slow-time (same for all OFDM symbols), motive radar is based on the narrowband assumption, as all fre-
or changed dynamically for each OFDM symbol, resulting quencies in the signal are approximated by the carrier frequency.
in 2D nonuniform sampling patterns [28]. As Doppler effect is frequency dependent, each frequency in
■■ Space–time block codes [19]: To simultaneously use all the signal undergoes a different Doppler shift for wideband sys-
subcarriers by all Tx antennas, OFDM subcarriers can be tems, and thus yields a different velocity estimate. Analogous
modulated with space–time block codes. This makes it to range migration, this leads to a Doppler frequency migration.
possible to maintain distance estimation parameters for Both effects prevent the 2D Fourier transform from collecting
each channel. It reduces, however, the unambiguous the entire signal energy into a single range–velocity cell and
­velocity range, as consecutive OFDM symbols constituting thus reduce the resolution both in range and velocity [7].
a block of code are required for distance processing. Both the range and Doppler frequency migration origi-
For PMCW radar, multiple Tx channels can be multiplexed nate from the motion of the target during the measurement.
based on orthogonal codes—in both fast- and slow-times [18]. Range migration occurs when the range change during the
To this end, the low cross correlation of codes (also under the measurement d mig = vTcycle exceeds one range cell (resolution)
condition of Doppler shift) is essential for effective separation Td = c 0 /(2B), i.e., for a target with the following velocity:
of MIMO channels.
;v; $ c 0 . (1)
Evidently, both fast-chirp and digital radars enable advanced 2BTcycle
modulation-specific multiplexing schemes. As multiplexing
implies sharing of available resources (e.g., time, frequency) Consequently, the range migration normalized to a range cell
between multiple channels, each multiplexing method leads to is the following:
some specific drawbacks compared to a single Tx channel in
d mig 2vBTcycle
terms of distance–velocity estimation. By a proper choice of the g RM = = . (2)
Td c0
multiplexing method, these drawbacks are minimized, while
obtaining improved DOA processing based on MIMO radar. From (2), the range migration is large for a large time–band-
A further important aspect of MIMO radar to consid- width product BTcycle and scales with the target velocity. The
er when choosing a multiplexing method is the coherency same equation describes the amount of Doppler frequency mi-
between the Tx channels. Maximum coherency is obtained gration [7], since both effects are inherently linked. In fact, they
when all Tx channels transmit simultaneously using the same are representations of the same phenomenon in two different
bandwidth. In case of a time offset between the measurements dimensions: range–slow-time and frequency–Doppler-esti-
of Tx channels, the target motion leads to Doppler-induced mate, respectively. Hence, for moving targets, range and Dop-
phase shifts between channels that add up to the DOA-induced pler frequency migration limit the simultaneously achievable
phase progression. Analogously, different carrier frequencies range and velocity resolution, imposing an upper bound jointly
of Tx channels imply a range-dependent, unknown phase shift on both parameters. For a typical bandwidth of 1 GHz and mea-
exp( j2rTfc,Tx 2d/c 0) adding to the DOA-induced phase differ- surement time of 20 ms, one cell migration occurs for velocities
ences, with Tfc,Tx denoting the carrier frequency offset of a Tx v 2 7.5 m/s. From the application perspective, the impact of

38 IEEE SIGNAL PROCESSING MAGAZINE | September 2019 |

migration effects is especially adverse during driving, as the 1) The waveform x (t) is delayed by x 0 due to the initial target
stationary targets appear moving relative to radar and thus are distance d 0 . This term is commonly used for range processing.
affected by migration-induced smearing in the radar image. 2) The additional delay in fast-time x v (t f ) in the argument of
A further problem especially relevant for digital modula- x denotes the delay of each time sample of the waveform,
tions, such as OFDM or PMCW, is the Doppler shift of the i.e., represents the Doppler-induced stretching/compression
signal frequencies. For OFDM radar, this leads to ICI between of the signal. For typical automotive applications, this term
subcarriers [17], [30], and for coded waveforms their cross- and is negligible [7].
autocorrelation characteristics deteriorate. Whereas the classi- 3) The third delay component x v (t s) in the argument of x is
cal approach accounts for Doppler shift by limiting the system the range change over slow-time due to the target’s motion.
parametrization such that the maximum Doppler shift is still When ignored, this term can cause range and Doppler fre-
acceptable, this becomes a critical limitation for high-perfor- quency migration.
mance automotive radar. 4) The first exponential term exp (- j2rfc x 0) in (5) represents
a constant phase shift for all samples and is irrelevant for
Signal processing framework for high- the range and velocity processing.
performance radar 5) The terms exp(- j2rfc x v (t f )) = exp( j2rfD t f ) describes the
As the discussion in the previous section indicates, a sig- Doppler shift fD = -2vfc /c 0 of the waveform in fast-time. It
nal processing framework based on a more rigorous signal has an adverse effect on the range estimation (e.g., leads to
model is needed to fully gain the benefits from upscaling of ICI in case of OFDM radar) when not taken into account.
radar parameters for increased estimation performance. This 6) The last term exp(- j2rfc x v (t s)) = exp( j2rfD t s) is the
implies that the current 2D-FFT-based processing has to be Doppler-induced phase progression over slow-time. It is
replaced with a better approximation of a 2D-matched filter. commonly used for Doppler processing.
An approach to achieve this for a single target (or multiple The conventional radar signal processing simplifies (5) to
targets with the same velocity) was proposed in [31]. Next, Ax (t f - x 0) exp( j2rfD t s) . Ignoring the third and fifth terms in
we summarize the research in [7], [29], and [30] and present (5), the range and velocity estimation problems can be decoupled
in general terms a signal processing framework capable of to fast-time and slow-time dimensions, respectively. The first term
migration-free and Doppler-robust range–velocity processing is then used for the range processing in fast-time, and the remain-
at a feasible computational cost and for an arbitrary number ing sixth term for Doppler processing is used in slow-time. For
of targets. We formulate it for arbitrary radar waveforms. high-performance automotive radar, however, neither the third nor
Consider an automotive radar transmitting a series of iden- the fifth term in (5) can be ignored, as this would lead to migration
tical waveforms (e.g., FMCW chirps, OFDM, or PMCW sym- effects and reduced performance in range processing. Later, we
bols) for distance–velocity estimation: describe a signal processing framework based on a more-precise
signal model. Table 1 gives an overview of how signal processing
x RF (t) = x (t - nTsym) exp( j2rfc [t - nTsym]), (3)
terms in (5) are treated by the conventional Fourier-based range-
where x (t) is the waveform in 0 1 t 1 Tsym that repeats pe- Doppler processing and by the reviewed framework (the terms in
riodically over slow-time t s = nTsym, n ! [0, N sym - 1], with the argument of x are discussed individually).
N sym being the number of waveforms (e.g., OFDM symbols) Ignoring the second and fourth terms that are irrelevant, the
during one measurement cycle. Let us define the fast-time signal model in (5) can be rewritten in the fast-time and slow-
t f = t - nTsym . time dimensions as follows:
Consider the radar signal in (3) reflected from a moving
y (t f , t s) = Ax ^t f - [x 0 + x v (t s)]h exp( j2rfD[t f + t s]). (6)
target at a time-dependent range
The Fourier transform of (6) in fast-time leads to the following:
d(t) = d 0 + vt = d 0 + vt f + vt s (4)
and let us denote the corresponding time-dependent round- y ( f, t s) = AX ( f - fD) exp(- j2rf [x 0 + x v (t s)]) exp( j2rfDt s), (7)
trip delay: x (t) = 2d (t) /c 0 = x 0 + x v (t f ) + x v (t s), where x 0 =
2d 0 /c 0 and x v (t) = 2vt/c 0 . where X ( f - fD) is the Doppler-shifted spectrum of the radar
After downconversion, the delayed signal at the receiver is waveform, the second exponential term is the slow-time-depen-
as follows: dent target distance, and the last term is the Doppler shift over
slow-time. The representation in (7) makes apparent that the range
y (t) = Ax RF (t - x (t)) exp(- j2rfc t f ) migration is caused by the slow-time-dependent range change

= Ax (t f - x (t)) exp(- j2rfc x (t)), (5) exp(- j2rf xv (t s)). To demonstrate that the same term is responsi-
ble for the Doppler frequency migration, we rewrite (7) as follows:
where A denotes the amplitude change of the signal through
propagation and reflection. By representing x (t) with its time- y ( f, t s) = AX ( f - fD) exp(- j2rf x0) exp( j2r[ fD + tfD ( f )]t s), (8)
independent (x 0), fast-time (x v (t f )), and slow-time (x v (t f ))
components, we can examine the following six elements of the where tfD = -2vf/c 0 is the Doppler frequency migration. The
signal model in (5): representation in (8) shows that the Doppler-induced complex

IEEE SIGNAL PROCESSING MAGAZINE | September 2019 | 39

 compensation (ACMC) for automotive
Table 1. Summary of terms in the signal model in (5).
radar in [7] as well as the Keystone trans-
Term Referred To As Conventional ­Processing Reviewed ­Framework form in [29] for synthetic aperture radar.
x (t - x 0) 1) Delay For range ­(distance) ­estimation The linear scaling of the velocity axis
x (t - x v(t f)) 2) D
 oppler scaling (fast-time) Neglected can be efficiently implemented based on
x (t - x v(t s)) 3) Migration term Neglected Compensated by ACMC chirp-Z transform, which has an order of
exp(-j2rfc x 0) 4) C  onstant phase shift Ignored (irrelevant) computational complexity O (N log N )
exp(j2rfD t f) 5) D
 oppler shift (fast-time) Neglected Compensated by ACDC that is the same as for FFT processing.
exp(j2rf D t s) 6) Doppler term For velocity ­estimation The idea of Doppler shift compensation
for all cells is known as all-cell Doppler
correction (ACDC) [30]. The correction
step is based on elementwise multipli-
exponentials over the slow-time are frequency dependent for cation, and thus its computational cost is negligible (though it
f ! [- B/2, B/2) . By compensating this frequency dependen- might need transforms between time and frequency domains,
cy, both the range and Doppler frequency migration can be depending on implementation). This makes the described signal
prevented, since a single term is causing both effects. processing framework feasible for real-time automotive radar
To accomplish this, we can start from the Doppler process- implementation as well as for other multitarget applications with
ing and perform it for each frequency in f ! [- B/2, B/2) with a sequence of identical waveforms and negligible higher-order
a kernel that scales the frequency axis to match ( fD + tfD ( f )) motion terms [see (4)]. The steps of the presented framework
as opposed to the conventional Fourier transform matching to are depicted in Figure 5 and compared to the conventional radar
fD [7], [30]. This operation scales proportionally the Doppler processing on the example of OFDM radar.
grid for each frequency, such that each frequency yields the Figure 6 shows measurement results of a car driving toward
same Doppler estimate. This results in the following migra- an OFDM-MIMO radar prototype [7]. In Figure 6(a), OFDM
tion-free Doppler spectrum in slow-time: subcarriers undergo a Doppler shift 0.34 times the subcarrier
spacing ( fD /Tf = 0.34) by reflecting from the moving car (see
N sym - 1 [30] for further details on the measurement setup). The con-
y( f, vt ) = / y ( f, t s) expc - j2r [ fc + f ] 2v nT m
t
c 0 ventional 2D-FFT processing without Doppler compensation
n=0
thus results in a considerable level of ICI, constituting itself as
= AX ( f - fD) exp(- j2rf x0) · D N sym c r[ fc + f ] T m,
2 [v - vt ]
c0 a bright trace along the distance axis in the velocity cell of the
(9) target (Figure 6(a), left; around v . -19 m/s). In contrast, by
shifting back the Doppler frequencies for each velocity cell,
where D N (x) = exp( j [N - 1] x/2) · sin (Nx/2) / sin (x/2) denotes ACDC prevents Doppler-induced performance degradation for
the Dirichlet kernel and represents the Doppler spectrum of the the entire radar image [30]. The target energy is focused into
target, with maximum at the velocity cell vt = v. This opera- its distance–velocity cell (d . 20 m; v . -19 m/s), obtaining
tion compresses the energy of each target into the correspond- full signal-to-noise ratio gain and preventing dynamic range
ing Doppler cell that is same along the frequency dimension reduction. Further performance analysis on ACDC is available
and thus migration-free. Subsequently, the Doppler shift of the in [7] and [30].
waveform can be compensated next by a frequency shift of To study the migration effects on a real-world example,
each velocity cell by its corresponding Doppler shift. This will Figure 6(b) presents a measurement with a bandwidth of
implicitly correct the Doppler shift for the entire radar signal, 625 MHz and measurement time of 39.4 ms. According to (2),
as the energy of each target is now focused in the correspond- for the car moving with v . -23 m/s, conventional Fourier
t
ing velocity cell. For the cell corresponding to the target
v processing results in a range and Doppler frequency migra-
velocity v, the Dirichlet kernel in (9) becomes D N sym(0) = 1. tion of more than three cells. Note that the scale of migration
Transforming y ( f, vt ) back to the fast-time domain, we can is equivalent to that of a system with a 1.25-GHz bandwidth
correct the Doppler shift for the cell vt = v by the following: and 19.7-ms measurement time, and thus is representative
of automotive radar. With conventional Fourier process-
y (t f , vt ) = Ax (t f - x 0) exp( j2rfD t f ) expc j2r fc t f m
2 t
v ing, range migration leads to a smearing of the target peak
c0  of more than three cells over the range axis. The Doppler
= Ax (t f - x 0) . (10) frequency migration results additionally in smearing of the
same scale over the velocity axis. In Figure 6(b), left, this
The subsequent range processing can be conventionally per- is particularly apparent for the corner reflector mounted on
formed based on a matched filter with knowledge of the wave- the roof of the vehicle to represent a distinct point target
form x (t f ) . This results in a range–velocity spectrum free of (smeared square around d . 16 m and v . -23 m/s) . Fig-
migration effects and Doppler-induced performance degradation. ure 6(b), right, shows the same radar image when processed
The described idea of migration compensation through with ACMC. For the entire image, no migration-induced
scaling of the velocity axis is the basis for the all-cell migration smearing of the peaks occurs. This is particularly clear for

40 IEEE SIGNAL PROCESSING MAGAZINE | September 2019 |

 ytf ,ts yf,ts zf,ts zf,v zd,v

FFT S.D. FFT IFFT

tf f f f d
...

Time ts ts ts v v
(a)
ytf ,ts yf,ts yf,v ytf ,v yt′f ,v zd,v
FFT ACMC IFFT ACDC FFT... S.D. IFFT
tf f f tf tf d
...

Time ts ts v v v v
(b)

FIGURE 5. An illustration comparing conventional Fourier-based processing with the reviewed migration and ICI-free signal processing framework on the
example of OFDM radar [7]. (a) 2D-FFT processing. (b) ACDC- and ACMC-based processing. S.D.: spectral division.

(dB)
30

20 20 20

10 10 10
Velocity (m/s)

Velocity (m/s)

0
0 0
−10
−10 −10
−20
−20 −20
−30

0 50 100 150 0 50 100 150

Distance (m) Distance (m)
(a)
(dB)
20 20 10

0
18 18
Distance (m)

Distance (m)

−10
16 16
−20

14 14
−30

12 12 −40
−24 −22 −20 −18 −24 −22 −20 −18
Velocity (m/s) Velocity (m/s)
(b)

FIGURE 6. Radar images of an approaching car measured with an OFDM-MIMO radar prototype. The results for 2D-FFT processing make apparent
the need for both Doppler-shift and target-motion compensation. The described migration and ICI-free processing overcomes limitations of 2D-FFT.
(a) ACDC (right) versus conventional 2D-FFT (left) [30]. The bright trace along the distance axis from the moving car is induced by ICI, which does not
occur for ACDC. (b) ACMC (right) versus conventional 2D-FFT (left) [7]. The moving car with a corner reflector mounted on top results in a range and
Doppler frequency migration of around 3.5 cells. For 2D-FFT, target reflections are smeared due to range and Doppler-frequency migration, whereas
ACMC collects the signal energy into a sharp peak.

IEEE SIGNAL PROCESSING MAGAZINE | September 2019 | 41

 the corner reflector, the energy of which is coherently formance degradation. For signal recovery, methods ranging
focused into a sharp peak. This illustrates that ACMC pre- from linear prediction [35] to sparse recovery [33] may be
vents the migration-induced degradation of the range and used. Similarly, digital beamforming can be used to focus
velocity resolution inherent to the conventional 2D-FFT interference in the angular domain toward the DOA of the
processing. For moving targets, this allows longer coherent interferer, thus reducing it in all other directions [34]. Alter-
integration times and a higher simultaneous range and veloc- native approaches are based on estimating the interfering
ity resolution. Further performance analyses of ACMC are signal and its subsequent subtraction (e.g., [36]). Such meth-
available in [7]. ods heavily rely on known characteristics of the interfering
signal. These methods for interference suppression have the
Interference mitigation drawback of being specific to a certain interference type.
As the number of automotive radar sensors on the road increas- Furthermore, since in practice perfect separation of signal
es, robustness against interference becomes a more important from interference is often impossible, such approaches dis-
challenge for reliable radar operation. Broadly speaking, high- card a portion of the radar signal. Moreover, they do not
end automotive radars are more susceptible to interference improve the overall interference situation, but only suppress
due to greater use of the time–frequency resources (e.g., large it locally at the receiver, and thus typically serve as the last
bandwidth, long measurement time). Considering the strict re- resort for interference mitigation.
quirements on reliability of operation, interference mitigation Intuitively, a more preferable approach is interference
becomes a core component of high-performance automotive avoidance instead of postprocessing. This requires active adap-
radars. Generally, to avoid interference, signals of different ra- tation of the Tx signal. An approach inspired by the interfer-
dars must be separable at least in one dimension, e.g., time, ence-avoidance mechanism of bats is to gradually steer clear
frequency, space, or code/waveform [32]. The main methods after detecting interference in the received signal by adjust-
for radar interference mitigation can be clustered into the fol- ing the carrier frequency [37]. This avoids interference instead
lowing categories. of its local suppression, and thus benefits both parties. It also
■■ Detect and suppress at the receiver: Interference is detect- does not require cooperation from the interferer. This method
ed from the measurement data and suppressed via drop- approaches its limits when the density of interferers makes it
ping the affected data and reconstructing their values (see impractical to blindly move in frequency at a risk of interfer-
[33] and [34] for FMCW radar and [35] and [36] for ing with another, initially unknown radar signal operating at a
OFDM radar). different frequency.
■■ Detect and avoid: When detecting interference in the mea- A more universal solution for adaptive interference avoid-
surement signal, the radar actively changes its signal to ance is interference-aware cognitive radar (IACR) [7], which
steer clear of interference in the subsequent cycles [37]. adopts the principles of cognitive radar [38], [39] to remedy the
■■ Interference-aware cognitive radar [7]: The radar senses automotive interference problem. IACR comprises the follow-
the entire operational spectrum and adaptively avoids inter- ing three main blocks [Figure 7(a)]:
ference via waveform adaptation [Figure 7(a)]. ■■ Spectrum sensing: This continuously analyzes the opera-
■■ Centralized coordination [8]: Self-driving cars are central- tional spectral band for interference and delivers the spec-
ly coordinated to avoid radar interference [Figure 7(b)]. tral occupancy information.
As these approaches range from local interference sup- ■■ Spectrum interpretation: This contains the cognitive intel-
pression at the receiver to coordinated interference avoid- ligence of the system. It is responsible for applying reason-
ance, they widely vary in their sovereignty and universality. ing to the information obtained via spectrum sensing by
The latter class of methods requires other radars to conform means of detection and classification of interfering signals.
to rules or cooperate, i.e., relies on actions of other radars. It also estimates various interference parameters. Based on
Methods from the first category are capable of suppressing this knowledge, prediction of the interference behavior for
interference of a specific form, and thus are effective only for the next transmission cycle is performed. This serves as a
certain interferers. basis for choosing an optimal adaptation strategy for the
The methods based on suppression of interference at the next measurement cycle.
receiver seek a representation where the interference and ■■ Waveform adaptation: This comprises possible adapta-
signal energy are maximally separable, such that the major tion methods along with the corresponding signal pro-
portion of interference can be dropped without considerable cessing algorithms. It also sets in motion the chosen
loss of the radar signal. An example of such processing is adaptation strategy. The goal is to avoid interference
mitigation of narrowband interference from OFDM radar by dynamically, while maintaining the radar estimation per-
dropping the corrupted subcarriers [35]. Equivalently, for formance. Possible adaptation space is frequency, time,
FMCW (fast-chirp) radar, interference from other FMCW and waveform.
sensors with different slopes can be filtered out from the An implementation of IACR in [7] uses a smaller por-
time signal, since only a portion of the time signal is affected tion of the operational bandwidth available for automotive
due to the antialiasing filter. The discarded portion of the radar (e.g., 0.5 GHz in the frequency range 77–81 GHz) and
radar signal needs to be recovered to prevent notable per- avoids interference by adaptation of the carrier frequency and

42 IEEE SIGNAL PROCESSING MAGAZINE | September 2019 |

starting time of the measurement. Based on spectrum sens- opment is radar operating with arbitrary digitally generated
ing and interpretation, the carrier frequency is adapted for waveforms. Multicarrier modulations, such as OFDM, not
each measurement cycle within the entire available band- only enable radar and communication to be combined in a
width to maximally mitigate interference from other radar single waveform, but also represent means for implementing
sensors. Depending on the chosen strategy, this may imply new software-defined radar concepts. In the context of MIMO
not only cycle-to-cycle hopping of the carrier frequency, but radar, multicarrier modulations permit the generation of a wide
also gradual adaptation within the measurement cycle itself. variety of orthogonal waveforms, enabling advantageous mul-
By adaptively avoiding interference from other automotive tiplexing schemes.
radars, IACR obtains robustness against interference in a uni- Based on a generalized signal model, we described the
versal, interference-agnostic manner and already in the ana- simplifications behind the conventional radar signal process-
log domain, i.e., by avoidance instead of post-treatment. This ing that no longer hold when upscaling radar parameters to
mitigates the overall interference problem through cognitive increase performance. We reviewed a signal processing
interference avoidance, benefitting the currently deployed, framework based on more rigorous modeling of radar signals
nonintelligent and nonadaptive systems through cognitive that allows migration-free and Doppler-robust performance
interference avoidance. Thus, it represents a promising path at a moderate computational cost. Finally, we surveyed auto-
for automotive radar to improve the interference situation. motive radar interference-mitigation methods. As automo-
An open issue is, however, interference avoidance between tive radar technology penetrates the market, the interference
multiple IACRs operating in the same environment, as each problem becomes more acute, which means that methods for
acting autonomously could potentially lead to mutual interfer- suppressing interference locally at the receiver will not be
ence. To this end, sets of rules or well-defined mechanisms sufficient. Promising paradigms for interference mitigation
are needed to ensure predictable behavior. in adaptive, cognitive, and/or coordinated manner arise, and
A conceptually different approach for automotive inter- their role in improving the overall interference situation is
ference mitigation is the RadarMAC architecture for central- becoming increasingly apparent. Although automotive radar
ized coordination of radar sensors on self-driving cars (Figure has been known for decades, this review indicates that there
7(b)). This approach requires vehicles to communicate their is a fertile ground for research, stimulated by the develop-
locations and routes to a control center, which represents the ment of self-driving cars.
radar operation schedule of vehicles in
the same environment as a graph color-
ing problem and determines playbooks
for each self-driving car for their radars
Environment
to operate interference-free. Especially
suited for dense interference settings,
centralized coordination is a promising
approach for reducing overall interfer-
ence. It requires, however, an additional Tx Rx
reliable communication link, availabil-
ity at all times, and participants obey-
ing a central coordinator. Thus, it does
not address the problem of ­interference Waveform Spectrum Spectrum
with existing nonadaptive, noncoopera- Adaptation Interpretation Sensing
tive radar sensors operating at the same
frequency band. Action Phase Perception Phase
(a)
Conclusions
As automated driving technology Receive and Apply Playbook Send Radar Playbooks
evolves, automotive radar is taking 4 3
major steps toward becoming a more Channel
powerful environment sensor. This (e.g., LTE) Control Center
transformation involves all aspects of 2
automotive radar, including system 1
Receive All Vehicle Updates
concept, modulation, and signal pro- Send Position/Trajectory Plan and Allocate Radar Parameters
cessing. In this article we summarized (b)
the major trends in the field of automo-
tive radar especially relevant for high- FIGURE 7. Diagram and illustration showing new paradigms for adaptive interference mitigation. (a) Closed-
performance radar systems designed for loop perception–action cycle of interference-aware cognitive radar [7]. (b) System overview of RadarMAC
self-driving cars. One primary devel- architecture for coordinated interference avoidance [8].

IEEE SIGNAL PROCESSING MAGAZINE | September 2019 | 43

 Authors [15] F. Engels, P. Heidenreich, A. M. Zoubir, F. K. Jondral, and M. Wintermantel,
“Advances in automotive radar: A framework on computationally efficient high-resolution
Gor Hakobyan ([email protected]) received his frequency estimation,” IEEE Signal Process. Mag., vol. 34, no. 2, pp. 36–46, Mar. 2017.
B.E. and M.E. degrees in radio and communication engineer- [16] D. Cohen and Y. C. Eldar, “Sub-Nyquist radar systems: Temporal, spectral,
ing from the National Polytechnic University of Armenia in and spatial compression,” IEEE Signal Process. Mag., vol. 35, no. 6, pp. 35–58,
Nov. 2018.
2011 and 2013, respectively, and his Dr.-Ing. degree in electri- [17] C. Sturm and W. Wiesbeck, “Waveform design and signal processing aspects
cal engineering from the University of Stuttgart, Germany, in for fusion of wireless communications and radar sensing,” in Proc. IEEE, July 2011,
vol. 99, pp. 1236–1259.
2018. He received the Best Paper Award in the Young Scientist
[18] A. Bourdoux, U. Ahmad, D. Guermandi, S. Brebels, A. Dewilde, and W. V.
Contest at the 2016 International Radar Symposium. He joined Thillo, “PMCW waveform and MIMO technique for a 79 GHz CMOS automotive
the Corporate Research Division of Robert Bosch GmbH and radar,” in Proc. 2016 IEEE Radar Conf. (RadarConf), May 2016. doi: 10.1109/
RADAR.2016.7485114.
is now a research scientist in the field of radar signal process-
[19] X.-g. Xia, T. Zhang, and L. Kong, “MIMO OFDM radar IRCI free range recon-
ing and system design. His current research interests include struction with sufficient cyclic prefix,” IEEE Trans. Aerosp. Electron. Syst., vol. 51,
radar signal processing, compressed sensing, radar modulation no. 3, pp. 2276–2293, July 2015.
schemes, and interference-mitigation techniques. [20] B. Schweizer, C. Knill, D. Schindler, and C. Waldschmidt, “Stepped-carrier
OFDM-radar processing scheme to retrieve high-resolution range-velocity profile at
Bin Yang ([email protected]) received his low sampling rate,” IEEE Trans. Microw. Theory Techn., vol. 66, no. 3, pp. 1610–
Dipl.-Ing. and Ph.D. degrees in electrical engineering in 1986 1618, Mar. 2018.
[21] D. Schindler, B. Schweizer, C. Knill, J. Hasch, and C. Waldschmidt, “MIMO-
and 1991, respectively, from the Ruhr University Bochum. He OFDM radar using a linear frequency modulated carrier to reduce sampling require-
is a professor and the head of the Institute for Signal ments,” IEEE Trans. Microw. Theory Techn., vol. 66, no. 7, pp. 3511–3520, July 2018.
Processing and System Theory in the Faculty of Computer [22] C. Knill, B. Schweizer, S. Sparrer, F. Roos, R. F. H. Fischer, and C.
Waldschmidt, “High range and Doppler resolution by application of compressed
Science, Electrical Engineering and Information Technology at sensing using low baseband bandwidth OFDM radar,” IEEE Trans. Microw. Theory
the University of Stuttgart, Germany. His research interests Techn., vol. 66, no. 7, pp. 3535–3546, July 2018.
include algorithms and applications of signal processing and [23] J. Li and P. Stoica, “MIMO radar with colocated antennas,” IEEE Signal
Process. Mag., vol. 24, no. 5, pp. 106–114, Sept. 2007.
machine learning, in particular radar signal processing, energy
[24] A. Zwanetski and H. Rohling, “Continuous wave MIMO radar based on time divi-
monitoring, and medical signal processing, with an increasing sion multiplexing,” in Proc. 2012 13th Int. Radar Symp., May 2012, pp. 119–121.
focus on deep-learning methods. [25] C. Pfeffer, R. Feger, C. Wagner, and A. Stelzer, “FMCW MIMO radar system
for frequency-division multiple TX-beamforming,” IEEE Trans. Microw. Theory
Techn., vol. 61, no. 12, pp. 4262–4274, Dec. 2013.
References [26] H. Haderer, R. Feger, C. Pfeffer, and A. Stelzer, “Millimeter-wave phase-cod-
[1] D. M. Grimes and T. O. Jones, “Automotive radar: A brief review,” in Proc. ed CW MIMO radar using zero- and low-correlation-zone sequence sets,” IEEE
IEEE, vol. 62, no. 6, pp. 804–822, June 1974. Trans. Microw. Theory Techn., vol. 64, no. 12, pp. 4312–4323, Dec. 2016.
[2] E. Guizzo, “How Google’s self-driving car works,” IEEE Spectrum, Oct. 11, [27] G. Hakobyan and B. Yang, “A novel OFDM-MIMO radar with non-equidistant
2011. [Online]. Available: https://spectrum.ieee.org/automaton/robotics/artificial subcarrier interleaving and compressed sensing,” in Proc. 17th Int. Radar Symp.
-intelligence/how-google-self-driving-car-works (IRS), 2016. doi: 10.1109/IRS.2016.7497312.
[3] H. H. Meinel, “Evolving automotive radar—from the very beginnings into the [28] G. Hakobyan and B. Yang, “A novel OFDM-MIMO radar with non-equidistant
future,” in Proc. 8th European Conf. Antennas and Propagation (EuCAP 2014), dynamic subcarrier interleaving,” in Proc. 2016 European Radar Conf. (EuRAD),
Apr. 2014, pp. 3107–3114. Oct. 2016, pp. 45–48.
[4] F. Foelster and H. Rohling, “Signal processing structure for automotive radar,” [29] R. P. Perry, R. C. DiPietro, and R. L. Fante, “Coherent integration with range migra-
Frequenz, vol. 60, no. 1–2, 2006. doi: 10.1515/FREQ.2006.60.1-2.20. tion using keystone formatting,” in Proc. 2007 IEEE Radar Conf., Apr. 2007, pp. 863–868.
[5] J. Hasch, E. Topak, R. Schnabel, T. Zwick, R. Weigel, and C. Waldschmidt, [30] G. Hakobyan and B. Yang, “A novel intercarrier-interference free signal pro-
“Millimeter-wave technology for automotive radar sensors in the 77 GHz frequency cessing scheme for OFDM radar,” IEEE Trans. Veh. Technol., vol. 67, no. 6, pp.
band,” IEEE Trans. Microw. Theory Techn., vol. 60, no. 3, pp. 845–860, Mar. 2012. 5158–5167, June 2018.
[6] S. M. Patole, M. Torlak, D. Wang, and M. Ali, “Automotive radars: A review of [31] R. F. Tigrek and P. V. Genderen, “Compensation of range migration for cycli-
signal processing techniques,” IEEE Signal Process. Mag., vol. 34, no. 2, pp. cally repetitive Doppler-sensitive waveform (OFDM),” IEEE Trans. Aerosp.
22–35, Mar. 2017. Electron. Syst., vol. 46, no. 4, pp. 2118–2123, Oct. 2010.
[7] G. Hakobyan, “Orthogonal frequency division multiplexing multiple-input mul- [32] M. Kunert, “The EU project MOSARIM: A general overview of project objectives
tiple-output automotive radar with novel signal processing algorithms,” Ph.D. dis- and conducted work,” in Proc. 2012 9th European Radar Conf., Oct. 2012, pp. 1–5.
sertation, Univ. of Stuttgart, Germany, 2018. [Online]. Available: http://dx.doi
.org/10.18419/opus-9830 [33] J. Bechter, F. Roos, M. Rahman, and C. Waldschmidt, “Automotive radar
interference mitigation using a sparse sampling approach,” in Proc. 2017 European
[8] J. Khoury, R. Ramanathan, D. McCloskey, R. Smith, and T. Campbell, Radar Conf. (EURAD), Oct. 2017, pp. 90–93.
“Radarmac: Mitigating radar interference in self-driving cars,” in Proc. 2016 13th
Annu. IEEE Int. Conf. Sensing, Communication, and Networking (SECON), June [34] J. Bechter, M. Rameez, and C. Waldschmidt, “Analytical and experimental
2016. doi: 10.1109/SAHCN.2016.7733011. investigations on mitigation of interference in a DBF MIMO radar,” IEEE Trans.
Microw. Theory Techn., vol. 65, no. 5, pp. 1727–1734, May 2017.
[9] M. Kronauge and H. Rohling, “New chirp sequence radar waveform,” IEEE
Trans. Aerosp. Electron. Syst., vol. 50, no. 4, pp. 2870–2877, Oct. 2014. [35] G. Hakobyan and B. Yang, “A novel narrowband interference suppression
method for OFDM radar,” in Proc. 24th European Signal Processing Conf.
[10] H. Rohling, “Radar CFAR thresholding in clutter and multiple target situations,” (EUSIPCO), 2016. doi: 10.1109/EUSIPCO.2016.7760645.
IEEE Trans. Aerosp. Electron. Syst., vol. AES-19, no. 4, pp. 608–621, July 1983.
[36] Y. L. Sit, B. Nuss, and T. Zwick, “On mutual interference cancellation in a
[11] L. Hammarstrand, L. Svensson, F. Sandblom, and J. Sorstedt, “Extended MIMO OFDM multiuser radar-communication network,” IEEE Trans. Veh.
object tracking using a radar resolution model,” IEEE Trans. Aerosp. Electron. Technol., vol. 67, no. 4, pp. 3339–3348, Apr. 2018.
Syst., vol. 48, no. 3, pp. 2371–2386, July 2012.
[37] J. Bechter, C. Sippel, and C. Waldschmidt, “Bats-inspired frequency hopping
[12] S. A. Askeland and T. Ekman, “Tracking with a high-resolution 2D spectral for mitigation of interference between automotive radars,” in Proc. 2016 IEEE
estimation based automotive radar,” IEEE Trans. Intell. Transp. Syst., vol. 16, no. 5, MTT-S Int. Conf. Microwaves for Intelligent Mobility (ICMIM), May 2016. doi:
pp. 2418–2423, Oct. 2015. 10.1109/ICMIM.2016.7533928.
[13] Z. Ji, M. Luciw, J. Weng, and S. Zeng, “Incremental online object learning in a [38] S. Haykin, “Cognitive radar: A way of the future,” IEEE Signal Process. Mag.,
vehicular radar-vision fusion framework,” IEEE Trans. Intell. Transp. Syst., vol. 12, vol. 23, no. 1, pp. 30–40, Jan. 2006.
no. 2, pp. 402–411, June 2011.
[39] M. S. Greco, F. Gini, P. Stinco, and K. Bell, “Cognitive radars: On the road to
[14] H. Krim and M. Viberg, “Two decades of array signal processing research: reality: Progress thus far and possibilities for the future,” IEEE Signal Process.
The parametric approach,” IEEE Signal Process. Mag., vol. 13, no. 4, pp. 67–94, Mag., vol. 35, no. 4, pp. 112–125, July 2018.
July 1996. SP


44 IEEE SIGNAL PROCESSING MAGAZINE | September 2019 |