IEEE JOURNAL OF SELECTED TOPICS IN SIGNAL PROCESSING, VOL. 15, NO.
4, JUNE 2021 861
Editorial:
Introduction to the Issue on Recent Advances in
Automotive Radar Signal Processing
I. INTRODUCTION the Appendix] has been published very recently. This forms
A major goal of the automotive industry is to make driving an opportunity to develop novel radar-based machine learning
safer and more comfortable. To this end, a focus has been put perception methods.
In the remaining part of this editorial, we provide an intro-
on the development of Advanced Driver Assistance Systems
(ADAS). The performance and reliability of these systems ductory discussion on the main differences between automotive
strongly depends on the capabilities of the environmental sens- radar and traditional radar, and present an overview of the articles
in this special issue.
ing, among which radar technology has become indispensable,
especially in view of Highly Automated Driving. Advantages
of radar technologies as compared to LiDAR and camera tech- II. AUTOMOTIVE RADAR VERSUS TRADITIONAL RADAR
nologies are clearly the more robust and preferred operation in In the sequel, we list notable differences between automotive
adverse weather conditions, higher range, direct measurement radar and traditional radar [item 2) in the Appendix], where
of relative velocity, and affordability. We witness considerable traditional radar includes airborne or ground surveillance ap-
efforts in research and development toward mitigating the limi- plications. It is important to appreciate the specific needs of
tations of automotive radar. Among these limitations are mainly automotive radar signal processing, as opposed to traditional
angular resolution and object classification capabilities. For in- radar signal processing. Due to space limitations, we focus on
stance, high-performance radar sensors can be developed using the following aspects only.
Multiple-Input Multiple-Output (MIMO) radar techniques and a r Modulation scheme
plurality of highly integrated radio frequency components. An- r Angle determination scheme
other trend is to exploit specific data features, e.g. micro-Doppler r Target processing
signatures, so as to improve the detection and classification of r Multipath propagation effects
vulnerable road users. In particular, machine learning and deep r Interference
learning are being continuously investigated and tested for their
use in real-life situations. A. Modulation Scheme
The goal of this special issue is to present advanced signal
Traditional radar often covers a large observation range in the
processing techniques as an enabler for novel technological ad-
order of 104 m, and is almost always of the pulsed-Doppler
vances. Both industrial and academic communities are targeted
type. To this end, important aspects are directional transmit
with this collection of papers.
power, supported by the use of high-gain antennas, and the
We wish to keep this editorial short and refer to the invited
management of range and Doppler ambiguities, e.g. by means
overview article, titled “Automotive Radar Signal Processing:
of successive coherent processing intervals using different pulse
Research Directions and Practical Challenges” by Engels et al.
repetition frequencies. For automotive radar, the observation
for a recent overview of state-of-the-art automotive signal pro-
range is limited to only several hundred meters, so that FMCW
cessing and a comprehensive list of references. In this article, the
radar techniques are employed. In particular, the chirp sequence
automotive radar signal model for the multiple-target case using
modulation is predominant. Moreover, the transmit power is
MIMO schemes is provided, along with a practical processing
limited by 10 dBm since automotive radar is operated in urban
chain to calculate the target list. Real data examples are given
environments.
to demonstrate the capabilities of a modern high-performance
radar sensor, produced in series. The article especially provides
an overview of stationary and moving target processing for B. Angle Determination Scheme
occupancy grid mapping, object tracking and classification. It State-of-the-art systems of both traditional and automotive
also addresses recent radar-based machine learning approaches, radar use phased array technology. To maximize the detec-
and points out a general lack of representative and sufficiently tion range, traditional radar often employs Tx beamforming in
large data sets with annotated radar data. In this regards, we scanning mode, and performs the angle determination using
would like to remark that the RadarScenes dataset [item 1) in Rx monopulse techniques in azimuth and elevation. MIMO
techniques are rarely used since they generally decrease the
Digital Object Identifier 10.1109/JSTSP.2021.3083445 directional transmit power. In automotive radar, usually no Tx
1932-4553 © 2021 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: INDIAN INSTITUTE OF TECHNOLOGY MADRAS. Downloaded on June 12,2023 at 13:19:51 UTC from IEEE Xplore. Restrictions apply.
862 IEEE JOURNAL OF SELECTED TOPICS IN SIGNAL PROCESSING, VOL. 15, NO. 4, JUNE 2021
beamforming in scanning mode is employed since latency is a unintentional jamming, is not an issue. However, intentional
crucial system parameter and all angles in the field of view are jamming as part of an electronic warfare tactic must be taken into
considered as equally important. Angle determination is per- account. The situation is different in automotive radar, where
formed with Rx beamforming and parametric angle estimation many modern cars use radar sensors in the same observation
in azimuth and elevation. Automotive radar also heavily relies on area. Consequently, interference is very likely in present-day
MIMO techniques to increase angular resolution. Note that due traffic scenarios and its detection and mitigation are challenges
to chirp sequence modulation and the corresponding processing, of high practical importance.
only single snapshots per range-Doppler cell are available for
angle estimation in automotive radar. III. OVERVIEW OF THE ARTICLES
Note also that whereas traditional radar systems are often
The response to the Call for Papers was overwhelming. We
“handcrafted” with costly components, state-of-the-art SoC so-
received a total of 40 papers of which we could accept only 14 in
lutions are available for automotive radar. These are generally
this special issue. Each paper was reviewed by at least three
highly cost-efficient and suited for automotive mass production.
independent experts. We wish to thank them for their careful
The invited overview article in this special issue provides more
assessment of the manuscripts and their timeliness in providing
information on recent practical automotive radar system config-
us with invaluable reviews. We believe that the final outcome
urations.
is a fine set of high-quality and interesting papers that you will
enjoy reading. Below is a short description of each paper.
C. Target Processing
One of the most notable differences between traditional and A. Model-Based High-Resolution Parameter Estimation
automotive radar is target motion. Whereas everything that The predominant solution for automotive radar is employing
does not move is considered irrelevant in traditional radar, and the chirp sequence transform for range and Doppler parameter
is consequently discarded or modelled as clutter, all targets estimation and MIMO antenna arrays for angle estimation. Four
are considered relevant for environmental sensing and scene papers consider the task of model-based parameter estimation
understanding in automotive radar. Moving targets represent with high-resolution and imaging capability.
other traffic participants, and non-moving targets represent the The paper entitled “High-Fidelity 4D Automotive Radar
stationary environment and are used e.g. to obtain free-space Sensing for Autonomous Vehicles: A Sparsity-Oriented Ap-
representations. proach” by Sun and Zhang provides a recent summary of spar-
In traditional radar surveillance applications, there are typi- sity concepts for automotive radar signal processing, including
cally only a few objects of interest and only a few detections practical adaptations. In particular, this includes range-Doppler
per object. An object classification is often performed based imaging with SSFW modulation and two-dimensional angle es-
on micro-Doppler modulations, resulting from rotating com- timation with matrix completion for sparse arrays. The benefit of
ponents such as wheels, rotors, jet engines, etc. In contrast, using sparse techniques are the economical use of the spectrum
the spatial resolution in modern automotive radar is such that and the reduction in RF transceiver hardware, and generally
modern sensors are able to see several thousands of detections come at the cost of a computational overhead. Numerical ex-
per measurement cycle in a typical traffic scenario. Since each amples in idealistic use cases with one or two point targets
object shows several detections, this point cloud information can are presented to demonstrate the achieved dynamic contrast and
be used for object classification. A modified tracking procedure sidelobe level.
for extended targets becomes necessary for automotive radar. In the paper entitled “Cognitive Antenna Selection (CASE)
Using Bobrovski-Zakai Bound” by Tabrikian et al. a new
D. Multipath Propagation cognitive receiver configuration is proposed, in which a large
Multipath propagation has been considered to some extent number of sensor array elements is connected to a small num-
in traditional radar, e.g. in low-angle target tracking when the ber of receive channels via a switching matrix. The state of
sea represents a specular surface, or in airborne ground-looking the switching matrix is sequentially updated using information
radar where interference from ground reflections can cause from previous observations and prior information. According
fading. However, for automotive radar, multipath propagation to the proposed scheme, the state of the switching matrix is
is omnipresent. There typically is vertical multipath with the obtained by the minimization of the Bobrovski-Zakai bound on
road surface, which produces distance-dependent interference the mean squared error of the angle estimate, which accounts
patterns when not resolved, and there is horizontal multipath due for the effect of ambiguity. This work shows that the proposed
to reflecting surfaces from guard rails or walls, which produces approach outperforms random and linear switching algorithms
“ghost” targets. Both phenomena are very problematic and have both asymptotically and in the threshold region.
to be taken into account in the signal processing. A SAR application to enhance the imaging resolution in
automotive radar is explored in the paper entitled “Automotive
Squint-forward-looking SAR: High Resolution and Early Warn-
E. Interference
ing” by Hu et al. By analyzing the 2D resolution of automotive
In traditional radar surveillance applications, there is usually SAR, the authors show that the key spatially-variant component
only a single radar per observation area, so interference, or is determined by the Doppler resolution vector. Based on this,
Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY MADRAS. Downloaded on June 12,2023 at 13:19:51 UTC from IEEE Xplore. Restrictions apply.
IEEE JOURNAL OF SELECTED TOPICS IN SIGNAL PROCESSING, VOL. 15, NO. 4, JUNE 2021 863
the constraints for squint-forward-looking SAR to satisfy both labeling. For demonstration, an exemplary deep neural network
the 2D resolution requirement and the safety requirement, are for radar signature classification is presented.
derived and its effectiveness under extremely high squint angle A novel CNN-based object detection method using radar only
is verified. The results show that the proposed method is a viable is proposed in the paper entitled “RODNet: A Real-Time Radar
technique in automotive radar for early obstacle detection in the Object Detection Network Cross-Supervised by Camera-Radar
squint-forward direction. Fused Object 3D Localization” by Wang et al. The method uses
In the paper entitled “Super-resolution in Automotive Pulse low-level radar data in the form of RF images and provides
Radars” by Delgado et al. the authors consider a 3D mmWave object classes and corresponding bird’s eye view coordinates.
pulse radar and analyze the super-resolution of multiple target The employed CRUW dataset has also been created for that
echoes. Upon exploiting the atomic norm to harness sparsity purpose and is now publicly available. Notably, the annotations
in the continuous parameter domain, the identification of the of this dataset are not human-labeled, but have been obtained
target parameters is re-cast in terms of rank minimization of in a cross-supervized manner from an object detection method
a properly structured multi-level Toeplitz matrix. A discussion based on radar-camera fusion.
on a low-complexity implementation of the recovery strategy
completes the paper.
D. Radar and Communications
B. Interference Detection and Mitigation
A joint consideration of radar and communication offers an
Many modern cars employ radar sensors so that the occur- efficient use of the spectrum, but is virtually unused in today’s
rence of interference in the shared unregulated frequency band practical systems. Three papers elaborate on the advantages of
is very likely in present-day traffic scenarios. The interference a joint consideration.
commonly occurs as bursts in the time domain and results in a The article “FMCW Radar Network: Multiple Access and
decreased detection sensitivity. Many signal processing methods Interference Mitigation” by Jin and Roy deals with interfer-
for de-noising and interference mitigation of radar signals are ence among vehicular radars, especially due to proliferation of
available, where recent approaches using deep neural networks vehicles with many radar sensors. The article classifies radar
show promising performance. In the paper entitled “Resource- interference into two categories: incoherent and coherent. The
efficient Deep Neural Networks for Automotive Radar Inter- coherent interference is especially challenging since it can be
ference Mitigation” by Rock et al. the authors address the detected as false targets. In this work, the authors focus on
high memory and computational requirements of CNN-based distributed medium access control protocols to avoid coherent
techniques and present a comprehensive study of quantization interference. In particular, the protocols are operated without
of network weights and activations. The data set for training and any time synchronization and cooperation.
testing uses real-world FMCW radar measurements in combi- In order to meet the high-resolution requirements for ADAS,
nation with simulated interference. The suggested solution of co-located MIMO sensor systems with various multiplexing
this work shows a competitive performance and can run in a techniques are deployed. Among various multiplexing tech-
specialized radar sensor hardware with limited memory and niques, Time Division Multiplexing (TDM) is the appropriate
computational power. choice for many automotive applications due to its low hard-
ware complexity. The drawback of the standard TDM MIMO
approach is the coupling of high velocity and angle information
C. Machine Learning and Deep Learning Applications
of targets. The coupling reduces the unambiguous estimation
Machine learning and deep learning offers a significant ad- interval of the Doppler frequencies of the targets by the number
vancement in object detection and classification capabilities of transmit antennas. The article “Joint Doppler Frequency and
and scene understanding. While recent methods have shown Direction of Arrival Estimation for TDM MIMO Automotive
that scene understanding is possible based on radar only, many Radars” by Baral and Torlak addresses this well-known Doppler
practical systems rely on a radar-camera fusion. A crucial point estimation ambiguity problem via a novel cost function to jointly
when developing supervized machine learning models is the estimate both the Doppler frequency and the angle of each target.
dataset with annotated measurements. mmWave joint communication-radar will enable high data
An experimental guide for setting up and using such a dataset rate communication and high-resolution radar sensing for appli-
is presented in the paper entitled “RaDICaL: A Synchronized cations such as autonomous driving. The article “Adaptive and
FMCW Radar, Depth, IMU and RGB Camera Data Dataset Fast Combined Waveform-Beamforming Design for mmWave
with Low-Level FMCW Radar Signals” by Lim et al. The Automotive Joint Communication-Radar” by Kumari et al. pro-
presented data set includes minimally compressed FMCW radar poses an adaptive and fast combined waveform-beamforming
data, along with synchronized odometry data and images from design with a phased-array architecture that permits a trade-off
a color and depth camera system, where all components are between communication and radar performances. The proposed
off-the-shelf with moderate cost and suitable for radar-camera design is especially effective for medium- and short-range auto-
fusion in robotics projects. In addition to the description of the motive radar channels in the range-Doppler angle domain with
data set, the authors also discuss practical techniques for spatial a wide field of view, at the cost of a small reduction in the
and temporal sensor alignment, and for automated and assisted communication rate.
Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY MADRAS. Downloaded on June 12,2023 at 13:19:51 UTC from IEEE Xplore. Restrictions apply.
864 IEEE JOURNAL OF SELECTED TOPICS IN SIGNAL PROCESSING, VOL. 15, NO. 4, JUNE 2021
E. Extended Target Tracking the significant computational cost of the resulting methods might
The paper entitled “Learning-based Extended Object Track- not be an option for today’s automotive radar systems, the pre-
ing Using Hierarchical Truncation Measurement Model with sented study serves as an excellent reference and might inspire
Automotive Radar” by Xia et al. addresses the Extended Object future automotive radar system design.
Tracking (EOT) problem, which is one of the challenging and
unique problems of automotive radar. The main idea of this PHILIPP HEIDENREICH
work is to model the spatial distribution of automotive radar Stellantis
measurements as a Hierarchical Truncated Gaussian (HTG) with Opel Automobile GmbH
structural geometry parameters that can be learned from the Germany
training data. This work shows that the proposed HTG mea-
surement model resembles the spatial distribution of real-world
automotive radar measurements and can serve as an adequate ABDELHAK ZOUBIR
data-driven model for extended object tracking. It was proposed Signal Processing Group
to learn the HTG model from the data and to incorporate it into a Technische Universitt Darmstadt
random matrix based EOT approach using a factorized Gaussian Germany
inverse Wishart and Rao-Blackwellized particle density repre-
sentations. The effectiveness of the proposed approaches was
verified using both synthetic and nuScenes datasets. IGAL BILIK
Department of Electrical and Computer Engineering
F. Prototyping and Experimentation Ben-Gurion University
Negev, Israel
In the paper entitled “Long Range, Low SWaP-C FMCW
Radar” by Lies et al. a method is developed, analyzed, and tested
to adapt low-cost, automotive-grade radar chipsets for long- MARIA GRECO
range sensing. The proposed solution employs SNR maximizing Department of Information Engineering
parameters, extended measurement intervals to maximize range. University of Pisa
Using the maximum likelihood estimator as a starting point, Italy
this paper develops more efficient techniques for target range
estimation that are robust to target motion. The results show that
MURAT TORLAK
low-cost radar can be adapted for range-extended applications.
Department of Electrical and Computer Engineering
In the paper entitled “Weighted SPICE Algorithms for Range-
University of Texas at Dallas
Doppler Imaging Using One-Bit Automotive Radar” by Shang et
USA
al. the authors extend methods in a weighted SPICE framework
to range-Doppler imaging using one-bit quantized data. A proof
REFERENCES
of concept application is shown for an equivalent linear signal
model that occurs in automotive radar LFMCW and PMCW [1] O. Schumann et al., “RadarScenes: A real-world radar point cloud data set
for automotive applications,” 2021, arXiv:2104.02493.
modulation. The general motivation for the presented work is to [2] M. Richards, J. Scheer, and W. Holm, Principles of Modern Radar: Basic
exchange costly hardware by extensive computations. Although Principles. Rijeka, Croatia: SciTech Publishing, 2010.
Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY MADRAS. Downloaded on June 12,2023 at 13:19:51 UTC from IEEE Xplore. Restrictions apply.
