Making Radar Detections Safe for Autonomous Driving: A Review

Tim Br

uhl

1,2

, Lukas Ewecker

, Robin Schwager

1,2

, Tin Stribor Sohn

and S

oren Hohmann

Dept. of Highly Automated Driving Systems, Dr. Ing. h.c. F. Porsche AG, Weissach, Germany

Institute of Control Systems, Karlsruhe Institute of Technology, Germany

Keywords:

SOTIF, Radar Perception, Radar Signal Processing, Radar-Camera Fusion, Functional Safety.

Abstract:

Radar sensors rank among the most common sensors used for highly automated driving functions due to their

solid distance and velocity measurement capabilities and their robustness against adversarial environmental

conditions. However, radar point clouds are noisy and must therefore be ﬁltered. This work reviews current

research with the aim to make radar detections usable for safe perception functions which require a guarantee

for correctness of the measured environmental representation. The impact on radar errors on the distinct

downstream tasks is explained. Besides, the term of safety for automated driving functions is illuminated under

consideration of the current standards and state-of-the-art research interpreting these standards is presented.

Furthermore, this work discusses safe radar signal processing and ﬁltering, approaches to enrich radar data

points by information fusion, e.g. from cameras and other radars, and development tools for safe radar-based

perception functions. Finally, next steps on the way towards safety guarantees for radar sensors are identiﬁed.

1 INTRODUCTION

Autonomous driving and parking are two of the major

emerging ﬁelds in the current development of the au-

tomotive industry. In contrast to assisting functions,

the driver is not involved during autonomous func-

tions which exceed SAE level 2 (SAE International.,

2021). With this, the safety demands on elements of

the function rise signiﬁcantly. This applies for both

planning and acting, but these two items will only

work properly if the sensing element performs accu-

rately. Critical situations must not be missed as falling

back on the driver is not an option anymore. The

system needs to handle a variety of tasks in the op-

erational design domain, which includes unforeseen,

probably even unimaginable situations. During devel-

opment, it needs to be ensured that these situations are

covered. Typical sensors which are used for environ-

ment perception are radar sensors, lidar sensors and

cameras (Yeong et al., 2021), which are combined to

sensor sets in case of autonomous vehicles.

Radar is an acronym for Radio Detection and

Ranging, which provides an indication on its work-

ing principle. Main parts of a radar sensor are the

Voltage-Controlled Oscillator (VCO) which gener-

ates an electromagnetic radio wave, which is trans-

mitted by a set of antennas. These radio waves reﬂect

in the world. The reﬂected waves are received by a

second set of antennas and processed to determine at-

tributes like position, signal power and relative veloc-

ity of the reﬂected obstacle. Initially, radar sensors

were primarily applied in military, aviation and nau-

tical ﬁelds. Meanwhile, millimeter-wave radar sen-

sors also gained a considerable proportion in automo-

tive sensor systems, starting as a sensor for assistance

systems like adaptive cruise control and safety func-

tions such as autonomous emergency braking, and

now being an enabler for autonomous driving func-

tions (Waldschmidt et al., 2021). Advantages of radar

sensors are the compact form factor, due to high inte-

gration density of components, as well as the strong

performance when it comes to measuring even high

distances as well as velocities in a single scan (Stein-

baeck et al., 2017). Common modern sensor systems

achieve a range of over 200 m with an accuracy and

resolution of 0.1m, and a velocity accuracy of 0.05

(Aptiv PLC, 2023) (Robert Bosch GmbH, 2023). The

capability of radar waves to pervade plastics as well

as even thin coating layers provides the opportunity

of integrating the sensor invisible behind the vehicle’s

bumpers. Additionally, the radar is the only sensor

showing exceptional environmental and atmospheric

robustness (Marti et al., 2019) Considering the com-

plementary of sensors, a combination of radar sen-

sors and camera sensors seems particularly promising

(Zhou et al., 2022). The major drawbacks of radar

Brühl, T., Ewecker, L., Schwager, R., Sohn, T. and Hohmann, S.

Making Radar Detections Safe for Autonomous Driving: A Review.

DOI: 10.5220/0012630400003702

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 10th International Conference on Vehicle Technology and Intelligent Transport Systems (VEHITS 2024), pages 299-310

ISBN: 978-989-758-703-0; ISSN: 2184-495X

299

sensor systems like moderate orientation measuring

capability or weak classiﬁcation performance are ad-

dressed by the camera.

As lidar is an acronym for Light Detection and

Ranging, these sensors share their working princi-

ple of transmission and reception of electromagnetic

wave with the radar sensors. However, the wave-

length of lidar sensors is close to the visible part of the

spectrum and hence shorter compared to radar sen-

sors. Elementary disparities in the physical properties

of the measures come with this difference in wave-

length. Lidar sensors exceed radar sensors in spatio-

temporal consistency, meaning that data points are

very congruent in two consecutive measures of the

same scene. This is not given for radar sensors, as

the points spread and scatter along an object (Bilik,

2023). Another asset is the high lateral and, thanks

to multiple lidar scan layers, even elevational reso-

lution of the lidar point cloud. Hence, lidar sensors

are superior to radars in object detection and classi-

ﬁcation tasks. However, except for new approaches

which try to integrate the Frequency-Modulated Con-

tinuous Wave (FMCW) technique to lidars (Sayyah

et al., 2022), common lidar sensors are not capable of

directly measuring the relative velocity per point. The

achievable range of lidar sensors is slightly less com-

pared to radars. In addition, lidar sensors are heavier

and more cumbersome, which is another aspect for

automotive use cases.

Most important for the considerations of this work

is that, due to its working principle, lidar suffers from

impairments in adversarial weather conditions like

rain, snow and especially fog (Zang et al., 2019).

Looking from the perspective of safe environment

perception, that is why lidar does not qualify that well

for this task in compared to the radar sensor. Contrary

to radar sensors, lidars need a cleaning solution which

often includes washing ﬂuid. This work aims to show

that radars have the potential to serve as sensors for

safe environmental perception even in bad lighting

and weather, if attention is paid to certain singulari-

ties during the signal processing.

2 STRUCTURE

The paper is structured as follows. At ﬁrst, an

overview of current radar tasks in the ﬁeld of auto-

mated driving functions is given. Next, we show the

current radar processing approaches and point out that

safety considerations are a special recess in the bunch

of radar processing methods. The term of safe radar

detections is elucidated and current literature in this

area is introduced. In this context, we highlight the

role of the two most important safety standards for au-

tomotive applications. We look at fusion techniques

with the potential to add information and reduce un-

certainty of radar points, and we describe data sets

for advancement of radar point treatment. Eventually,

we outline the results of our review, draw conclusions

and demonstrate further potential research activities.

3 LITERATURE REVIEW

The literature section review comprises the radar ca-

pabilities, a survey of radar signal processing, an anal-

ysis of error sources in radar sensors, an introduction

into the concept of safe detection, a deﬁnition of crit-

ical radar points and development approaches to alle-

viate the criticality of radar points.

3.1 Radar Capabilities

This section gives an introduction into the different

tasks in automated driving functions which a radar

can handle by today. The four presented ﬁelds of ap-

plication are depicted in ﬁgure 1.

3.1.1 Radar Localization

Currently, measurement data acquired by radar sen-

sors is mainly used for object detection (Schumann

et al., 2020)(Scheiner et al., 2021). However, new

application ﬁelds for radar offer to the research com-

munity. The property that radar measurements have

a point cloud as an output makes them suitable for

localization in already seen environments via point

cloud registration. In the past, several works dealt

with the challenge to apply Simultaneous Localiza-

tion and Mapping (SLAM) algorithms to radar data.

(Hong et al., 2021) describe a SLAM algorithm to

deal with bad weather and illumination conditions.

A blob detector ﬁnds key points in a cartesian radar

image and detections on moving vehicles are deleted

by a graph-based outlier detection algorithm. The

key points are tracked, and motion is compensated

using a pose graph. After every measure, the min-

imally required number of features is completed by

adding new features in case that some were lost dur-

ing the last tracking cycle. The loop closure is done

by comparing candidates in a Principal Component

Analysis (PCA) and rejecting the least likely ones.

Comparable performance to vision-based and lidar-

based approaches was achieved. Similar work was

done by (Schuster et al., 2016), who used an occu-

pancy grid based key point extractor and constructed

a graph, adding poses as well as odometry measures

VEHITS 2024 - 10th International Conference on Vehicle Technology and Intelligent Transport Systems

300

(a) Matching radar point clouds (red dots) to map

data (black structure) for localization purposes

(b) Radar free space detection (red line) based on

radar points (black dots)

points at constant vehicle speed of v = 2.0

(d) Radar object detection and tracking (Schu-

mann et al., 2021)

Figure 1: Different application ﬁelds for radar sensors.

of the vehicle. The outliers are detected with the

Random Sample Consensus (RANSAC) algorithm in

this work. The graph is optimized using Levenberg-

Marquardt-optimization.

3.1.2 Radar Free Space Detection

Another ﬁeld of frequent research activity is radar free

space detection. This can also be considered as the

inverse problem of the detection of static obstacles in

the surrounding of the sensor. The Radar Cross Sec-

tion (RCS) is proportional to the backscatter signal

power density (Knott, 2012), meaning that a higher

RCS increases the peak of the detection in the noise

ﬂoor. Hence, given that the backscatter power is not

constant, but normally distributed, the probability that

the object is detected increases as well. The proba-

bility to receive a backscatter signal and the average

signal power depends on the RCS of the object which

is measured (Knott, 2012). A typical approach to ﬁnd

drivable space is to discretize the environment to oc-

cupancy grids, ﬁlling cells with reﬂections which fea-

ture reﬂections, and leaving out the ones without re-

ﬂections (Xu et al., 2020)(Li et al., 2018). Due to

the noisiness of the radar point cloud, ﬁltering algo-

rithms are applied to reduce the false positive rate of

the boundary detection. (Popov et al., 2023) tackle

the problem of free space detection by a deep neural

network, whose input is a radar point cloud which is

accumulated over a time of 0.5 s, ego-motion compen-

sated, and projected in a 2D bird’s-eye-view format.

Human-annotated bounding boxes based on lidar de-

tections are used as the training data. The network

features not only a free space segmentation head, but

also a bounding box detection and a classiﬁcation.

Making Radar Detections Safe for Autonomous Driving: A Review

301

3.1.3 Radar Odometry

Recent approaches showed that radar sensors can also

be used for motion estimation due to their ability to

precisely determine the Doppler velocity. Assuming

that most of the perceived targets are stationary, the

reﬂections from these points can be used, together

with the angular information, to estimate the longi-

tudinal and lateral velocity as well as the yaw rate

of a vehicle. (Kellner et al., 2014) showed the fea-

sibility of this approach and investigated the preci-

sion improvement when using multiple radars. The

geometric coherence between the measurement is de-

rived and based on that, the vehicle’s motion state is

estimated, using the Ackermann conditions. Radar

odometry can be tackled on a second way, which is

closer to the SLAM approach described before. Point

clouds can be matched and the vehicle’s motion is the

ﬁrst derivative of the pose transform, the transforma-

tion between two successive point clouds. The work

of Adolfsson et al. proposes a conservative ﬁltering

approach which keeps only a set of the strongest radar

detections with constant size. Additionally, it is as-

sumed that true positives represent objects with a sur-

face. Thus, a surface vector is estimated for every

radar point. The points with a surface vector corre-

lating only weakly with the other points are deleted.

The thinned-out point cloud is registered to the previ-

ous one using the Broyden-Fletcher-Goldfarb-Shanno

(BFGS) line search method (Adolfsson et al., 2021).

3.2 Radar Signal Processing

Radar errors may also happen in the digital signal pro-

cessing steps and should be prevented here as best as

possible. Modern radars send FMCW waveforms and

determine the range of a target point and its relative

velocity by the frequency shift of the received sig-

nal (Patole et al., 2017). The whole radar signal pro-

cessing is a many-layered process, starting with in-

terference mitigation, fast-fourier transformation and

beamforming in the preprocessing step, followed by

the creation of the target list (Engels et al., 2021).

In this step, a Constant False Alarm Rate (CFAR)

mechanism is often applied to determine the thresh-

old for a signal peak to be considered as a target

point. The peak’s position allows for calculation of

the other point attributes, the so-called parameter es-

timation (Engels et al., 2021). The resulting point

cloud is the basis for all the applications which were

described before, such as object detection or local-

ization. However, the techniques described in sec-

tion 3.1 all try to achieve a regularly working result,

ﬁltering the point cloud and keeping only strong re-

ﬂections while accepting to ignore weak reﬂections

of objects in favor of a low false positive rate. These

ﬁltering methods are not sufﬁcient when undercutting

low failure in time rates must be guaranteed. In the

following, an overview of techniques in the state-of-

the-art for guaranteeing safe perception is given. For

this section, we state the term of the Object of Inter-

est (OoI), which can be both an object which should

be detected during an object detection or free space

estimation task or an arbitrary object with backscat-

tering, generating at least one true positive point, that

is considered for measures of radar odometry or lo-

calization.

3.3 Radar Hardware Development

Starting in front of the chain, it needs to be ensured

that the OoI generates at least one point in the point

cloud. This requires a sufﬁciently strong backscat-

ter signal which exceeds the required threshold. The

radar sensor itself is accountable for parts of the over-

all perception performance. (Gerstmair et al., 2019)

show that phase noise in the transmitted signal caused

by the voltage controlled oscillator plays a signiﬁcant

role for the signal-to-noise ratio. If phase noise can be

kept on a minor level, Vulnerable Road Users (VRUs)

such as pedestrians elicit reﬂections which exceed the

noise ﬂoor level, while a poor phase noise can lead

to a pedestrian masked by the noise ﬂoor. Further-

more, they outline methods to estimate the power

spectrum density of the phase noise and monitoring

approaches using a cascaded Monolithic Microwave

Integrated Circuit (MMIC) system. Since the inter-

ference phenomenon may also lead to a decreased

signal-to-noise level, other works consider its mitiga-

tion. (Aydogdu et al., 2020) discuss the effects of in-

terference and propose proactive strategies, proactive

meaning to avoid or reduce interference. These works

are undoubtedly relevant to reduce the occurrence of

false negative points, existing objects of the environ-

ment which are not represented by a radar target point.

However, they do also not focus on the safety-aware

selection of radar points. Hence, in the following, the

methods to treat the radar points which are returned

by the sensor for consecutive purposes shall be dis-

cussed.

3.4 Origin and Classiﬁcation of False

Radar Points

Radar point clouds are noisy, meaning that they con-

tain points which do not represent an OoI. (Bilik et al.,

2019) names clutter as a reason for noisiness and a

major challenge for radar signal processing. Clutter

VEHITS 2024 - 10th International Conference on Vehicle Technology and Intelligent Transport Systems

302

points are road echoes in front of the vehicle which

are barely distinct from signal returns of OoIs. Mul-

tipath clutter is described as another effect which

can obscure true positive targets due to the doppler

spread effect (Yu and Krolik, 2012). On the other

hand, this may cause false positive radar points as

well. The impact of these false points has been de-

scribed in various works in the past. (Barnes and Pos-

ner, 2020) explains a method to identify those radar

points which are robust enough to serve as key points

for tasks such as motion estimation or localization.

A U-Net like architecture is used and trained based

on ground-truth transformation between consecutive

point clouds. The output of this network is the loca-

tion and quality of the key point as well as a unique-

ness attribute. The drawback of this approach is that

the selection of key points might be suitable for the

scan matching task, whereas the selection decisions

are not explainable.

In general, to make radar points safe and discover

potential unsafe points, it needs to be expounded what

safety is and which mechanisms in perception impair

the safety of the function.

3.5 What Is Safety?

One established safety concept is Functional Safety

(FuSa), which is described by the ISO 26262 stan-

dard (International Organization for Standardization,

2018). FuSa means that electrical systems need to

fulﬁll safety guarantees depending on their potential

to injure their user. The safety goals to avoid mal-

function in a situation are harder, the more likely the

user is exposed to a situation, the worse the situation

is controllable and the more severe resulting injuries

at a malfunction are. In automated driving functions,

it turned out that FuSa, originally thought as a con-

cept for hardware ﬂawlessness, cannot be applied well

for malfunctions which are caused by poor perception

capabilities or misbehavior of an algorithm. Hence,

the Safety Of The Intended Functionality (SOTIF)

standard was deﬁned, which ampliﬁes safety deﬁni-

tions and targets to provide a description for these

cases. According to (International Organization for

Standardization, 2022), SOTIF is deﬁned as the “ab-

sence of unreasonable risk due to hazards resulting

from insufﬁciencies of the intended functionality or

by reasonably foreseeable misuse by persons”, which

must be guaranteed for autonomous driving functions.

(Peng et al., 2023) name two classes for risks

of autonomous functions, external risk and internal

risk. Safe radar perception focuses mainly on exter-

nal risks, meaning that e.g., velocities and positions of

adversarial road participants are not determined cor-

rectly. In contrast to this, internal risk describes the

probability that an algorithm performs erroneously,

and external measurements are appropriate. Further-

more, they establish the term “SOTIF entropy” based

on the Shannon entropy, which describes the uncer-

tainty of a label prediction in an object classiﬁcation

task. This entropy can be determined for perception

as well as for prediction and planning.

Ensuring that autonomous vehicles make safe de-

cisions, the term “safety” can be interpreted in dif-

ferent ways. (Bila et al., 2017) (Muhammad et al.,

2021) name different tasks such as collision avoid-

ance, detection and tracking of vehicles or pedestrian

detection. Mastering each of these tasks will lead to a

safe autonomous driving function. Additionally, they

highlight the existence of the three stages measure-

ment, analysis, and execution. Each stage of this so-

called cognitive control cycle can impair the safety of

the function.

(Khatun et al., 2020) combines the two concepts

of SOTIF and FuSa by a scenario-based hazard analy-

sis and risk assessment (HARA). Hazards can emerge

as vehicle-based or functional malfunction, cover-

ing typical FuSa parts as well as SOTIF parts of the

safety contemplation. Scenarios can result from both

branches and are preselected. To rate the risk, the

three criteria of FuSa are employed. Hazard and Op-

erability (HAZOP) keywords are collected and trans-

ferred to SOTIF functional imperfections. Various

works (Chu et al., 2023)(Peng et al., 2021) empha-

size that dealing with uncertainties is an inevitable

part when it comes to making automated functions

safer. (Dietmayer, 2016) categorizes uncertainties in

determining the state (i.e., position, orientation, mo-

tion state etc.), the existence (presence detection) and

the classiﬁcation of an OoI.

3.6 Safety-Aware Function Evaluation

The review of (Hoss et al., 2022) discovers test-

ing methods to ensure methodically that functions

meet the safety requirements. They describe metrics

to measure relevance of certain scenarios, introduce

ways to specify scenarios as well as the Operational

Design Domain (ODD) and propose to create a test

catalog on either a knowledge-driven, a data-driven

or a hybrid way. Furthermore, it is outlined how to

acquire data for safety evaluation through different

ways, either by the vehicle-under-test itself or by us-

ing data which is received e.g., via vehicle-to-X.

Solving tasks such as object detection with ma-

chine learning methods is problematic due to their

black-box characteristics and lacking measure of un-

certainty of the decision made. However, one work

Making Radar Detections Safe for Autonomous Driving: A Review

303

shows how to apply SOTIF on these methods using

the example of a lane keeping function (Abdulazim

et al., 2021). They use a baseline model for the func-

tion which consists of a ﬁlter block, a memory unit,

and a predictor. Trigger events for a safety-critical

scenario are deﬁned and tested during the veriﬁcation

process. The veriﬁcation and validation process hap-

pened on both a synthesized data set and a data set

recorded by a real vehicle during a road test. Sum-

marizing, the work offers a guideline on how to apply

SOTIF concepts on machine learning methods, and

describes the procedure in a very simpliﬁed example.

As mentioned before, evaluation of the automated

function should already take place on the perception

stage, since failures in the perception affect all down-

stream algorithms. One way to identify the threats of

perception functions is to dissect the event chain dur-

ing perception, consisting of the information recep-

tion and information processing (Philipp et al., 2020).

Errors, leading to failure, can occur in the raw scan as

well as in the feature level. The work focuses on fea-

ture level errors and thus it lacks a detailed analysis of

the radar sensor disturbance and failure mechanisms.

Sch

onemann outlines a procedure to ensure safe

behavior of an autonomous function, in this case the

Automated Valet Parking (AVP) function, already in

the function design. He splits his observations in three

aspects: the minimum safety requirements deﬁne the

situations which an automated function should be able

to handle, the minimum required perception zone de-

ﬁnes the free space needed to monitor to set the vehi-

cle in a safe state, i.e., from driving to standstill, and

the minimum functional requirements, deviating ele-

ments such as perception or planner for the automated

function to work (Sch

onemann, 2019).

Chu et al. made valuable work in the ﬁeld of con-

necting SOTIF with perception challenges. Same to

Sch

onemann, they also use the safety distance con-

cept to model a minimum required perception area.

They derive the velocity of objects near to the vehi-

cle to become a potential collision candidate. Addi-

tionally, they connect these ﬁndings to a sensor model

with uncertainties to identify the required capabili-

ties of the perception system for collision avoidance.

Weaknesses in the coverage of the considered sen-

sor set were identiﬁed during a case study, and they

also suggest including degrading conditions regard-

ing lighting and weather or to apply the ideas to more

complex tasks (Chu et al., 2023).

3.7 Deﬁnition of Critical Radar Points

The previously introduced works deﬁne the safety

term and enable applying it to the perception of radar

systems. We emphasize that the ignorance of a radar

point can lead to a safety hazard as well as misinter-

pretation or wrong assessment of a situation. How-

ever, as false positive radar points occur abundantly,

ﬁltering needs to be done while being aware of the

hazards this brings along. Commonly known and

widespread are the terms to deﬁne the fraction of

points which are relevant from the given data set,

precision, and the fraction of points which are cov-

ered from the total relevant object set, recall (Powers,

2011)(Sun et al., 2023). Precision and recall are de-

ﬁned as

Precision =

T P

T P + FP

(1)

Recall =

T P

T P + FN

(2)

See ﬁgure 2 as an explanation of how the terms

are used in this context. The grey, transparent area

shows the ﬁeld of view of a radar sensor. The

backscatter of the car on the left hand side gener-

ate strong detection points, same for the posterior

walker. These two objects would be considered as

true positives. The anterior walker (”On-edge ob-

ject”) generates only one weak detection. Assuming

that weak detections are ﬁltered, this object would

be a false negative. Whereas some clutter from the

ground generates strong detections and is likely to

be interpreted as an obstacle, illustrated by the red-

dashed box. Lowering the threshold for detections to

count as ”strong” would include the on-edge object,

but also clutter from the ground. In succession, recall

increases and precision declines. The reverse effect

happens if the threshold is elevated. The on-edge ob-

ject is not included in the object list anymore and also

the false positive clutter is sparser.

False Positive Point

True Positive Point

On-edge Object

Weak detection

False Positive Object

True Positive Object

Strong detection

Figure 2: Tradeoff between recall and precision.

This tradeoff has been described elaborately in

the literature (Buckland and Gey, 1994)(Yang et al.,

VEHITS 2024 - 10th International Conference on Vehicle Technology and Intelligent Transport Systems

304

2019). With reference to radar points, in urban sce-

narios with increased multi-path propagation, much

interference by other vehicles and many objects, the

noise level is high compared to free ﬁeld scenarios.

Additionally, pedestrians may have a RCS of 0.01m

(Chen et al., 2014), which is a thousand or less of the

RCS of the truck next to them (Waldschmidt et al.,

2021). As VRUs with low RCS do hardly backscat-

ter and may vanish in the noise ﬂoor, a recall of 1.0

will not be reached, even if every point delivered by

the radar sensor is considered. Hence, increasing the

maximum recall rate is a main working ﬁeld for radar

specialists. Still, optimizing the recall rate will in-

volve more false positives, which is to the disadvan-

tage for the precision rate. We assume that a certain

precision rate needs to be reached to ensure that the

function is robust enough to master most of the situ-

ation without undergoing a deadlock, i.e., too many

false positives are detected in crowded situations and

the vehicle does not move anymore. Enriching radar

points which may be a threat for SOTIF conformity

with information may be one of the keys. This infor-

mation can descend from other sensors such as cam-

eras or radar sensors.

3.8 Camera-Radar Fusion

Fusion is the combination of information from more

than one source, utilizing the redundancy and the

congruence of data. According to (Khaleghi et al.,

2013), it can be distinguished between four different

fusion problems: imperfection, correlation, inconsis-

tency, and disparateness. Speaking about adding in-

formation to radar point clouds via fusion, especially

two aspects are relevant. The radar data may be de-

ﬁcient (imperfection) or miss information (inconsis-

tency). Correlation could become problematic if the

data from two sensors of a kind or the data of one

sensor at different time stamps shall be incorporated

in the fusion. Attributes that are usually fused, such

as position and velocity, don’t exhibit a disparity be-

tween different perception sensors.

(Tang et al., 2022) highlight that the use of camera

information is beneﬁcial to use for fusion with radar

sensors due to its complementarity. Moreover, they

propose different fusion schemes, e.g., object fusion

before and after tracking and region-of-interest based

fusion. However, their work does not concern about

early fusion stages such as fusion at the level of radar

detection points or camera pixels, but aim to solve the

task of object detection for radar and camera data sep-

arately before fusing the object lists.

The fusion in these levels is explained more pre-

cisely in (Zhou et al., 2022). Data-level fusion ex-

tracts the features and generates the object list from

the fusion of radar and camera data. Target-level

fusion extracts the features separately and combines

these features together. The structure called decision-

level fusion in their work is an object fusion which is

similar to a fusion before tracking. Furthermore, the

transformation of radar points to camera pixels, the

camera calibration required to that end, and the syn-

chronization process is discussed. Feature level fu-

sion is recognized as commonly used due to the erro-

neousness of radar point clouds. Thinking about safe

radar detection points, this may pertain for most non-

critical areas, while a fusion level at an earlier stage

might be beneﬁcial to keep the precision rate high.

(Velasco-Hernandez et al., 2020) mention that

deep learning plays a major role for multi sensor ob-

ject detection and fusion. According to (Zhou et al.,

2020), deep learning architectures can be split into

a two-stage detection network, where the radar as-

sumes the task of a region proposal network in a

deep learning image-based object detection algorithm

such as (Girshick, 2015)(Brazil and Liu, 2019), and

a one-stage detection network which solves the ob-

ject detection task inside of the image separately

and projects radar points in the 2D front camera im-

age, generating a sparse radar image (John and Mita,

2019). (Zhou et al., 2020) also name special attention

fusion (SAF) which can be combined with the already

known Fully Convolutional One-Stage Object Detec-

tion (FCOS) framework in a one-stage detection net-

work manner for the purpose of radar-camera-fusion.

The SAF block is a modiﬁed ResNet-50 block which

is applied on a radar image and results are multiplied

with those of the vision branch. By this means, the

training can be performed end-to-end (Chang et al.,

2020).

While the attention mechanism is applied to con-

trol how deep learning image recognition works, the

attention focuses purely on the presence of radar

points, not evaluating if these radar points can be haz-

ardous or relevant for the downstream automated driv-

ing function. Prior aspiration for the method is to

reach a high average precision on data sets, leaving

out safety aspects for the fusion.

3.9 Multi-Radar Fusion

The work on the fusion of multiple radar point clouds

which are generated by different sensors of a radar

belt conﬁguration at a vehicle is not as widespread,

compared to camera fusion. The reason might be that

the fusion can be trivial, aggregating points of all sen-

sors to a common point cloud, as it was also done in

data sets like (Schumann et al., 2021). However, the

Making Radar Detections Safe for Autonomous Driving: A Review

305

distribution of detections can be interesting for safety

aspects. Speaking about the identiﬁcation of false

positive points, the probability of a point being false

positive reduces signiﬁcantly if a second radar raises

a detection one cycle later in the same area. (Diehl

et al., 2020) handle multi-radar fusion by transferring

all measurements to a grid representation, modeling

the sensor uncertainties by 2D gaussian distributions.

The tracking can also be done in this grid represen-

tation using multiple particle ﬁlters. (Li et al., 2022)

shed light on the temporal relations between consec-

utive radar frames. Two frames, called current frame

and previous frame, are both passed through a back-

bone neural network to extract features. A ﬁlter layer

selects object features, and the feature matrices of

multiple object candidates are passed to the tempo-

ral relational layer, which encodes object spatial posi-

tions with the features of the two consecutive frames.

Eventually, a decoding step leads to the object list rep-

resentation. The achieved results exceeding the state-

of-the-art show the importance of temporal informa-

tion in radar data.

Overlapping ﬁelds-of-view in a radar belt offer the

opportunity to attain more information about a spe-

ciﬁc region of interest. Imagine a point that has a low

signal strength and is hence considered to be deleted.

If it is identiﬁed as being a critical point, multiple oc-

currences of backscatter in the same region at previ-

ous time steps can help to amplify evidence on the

existence of the point.

3.10 Data Sets

(Yurtsever et al., 2020) review a couple of data sets,

listing lighting conditions, weather conditions and

recorded data, but do not show which of the data sets

include radar data. For this paper, their work is taken

as a basis for an own small review about current data

sets which promote the development of radar data

processing algorithms. The results can be found in ta-

ble 1, and the most important data sets are explained

in the following passage.

The Oxford Radar RobotCar data set provides

radar point clouds in urban environment together with

a lidar point cloud and six cameras, indeed they used

a Navtech CTS350-X which is a 360-degree scanning

radar without doppler information and holds proper-

ties different from a belt consisting of multiple radars,

as it is conventionally used in today’s automotive in-

dustry (Barnes et al., 2020).

(Schumann et al., 2021) encounter this issue and

serve RadarScenes, which features a point cloud cap-

tured by four radar sensors and doppler velocities.

Additionally, all points are manually labeled, which is

to our knowledge unique for a radar data set. Unfortu-

nately, the radar sensors cover only the front and sides

of the vehicle, but no rear area. RadarScenes works

well when multi radar fusion shall be investigated, but

is not useful for radar-camera fusion. NuScenes is one

of the most popular radar data sets, featuring drives in

Boston and Singapore, six camera perspectives and

point clouds from ﬁve radars. Additionally, the 3D

ground truth bounding boxes are available (Caesar

et al., 2020). Although this data set has a compara-

tively large size, it does not feature a lot of hazardous

situations. RADIATE is a radar data set which puts

special effort in including bad weather and lighting

conditions such as night, snow, and rain, but unfortu-

nately lacking 3D annotations (Sheeny et al., 2021).

We ascertained that there are various data sets

which are made for the development of autonomous

driving functions. However, we could not ﬁnd a data

set which addresses safety critical scenes only. This is

a substantial lack in research, since improving fusion

algorithms on the existing data sets will not necessar-

ily improve the performance of the system handling

safety-critical situations. We suggest evaluating the

methods on data of the own system and putting effort

in a general description of how to make sensor setups

safe by identifying their individual weaknesses.

Using own data requires to record ground truth

data in parallel as a reference. Generally, annotated

data must be highly precise, and labels need to be

complete (Xiao et al., 2021). The ground truth data

must hold signiﬁcantly more details than the system

which is under development. For tasks like object

or occupancy detection, 3D bounding boxes represent

the ground truth. As 3D bounding box labeling by

hand is tedious, Lee et al. propose to use 3D object

detectors applied on a lidar point cloud to generate

labeling proposals while the annotator’s only task is

to select the individual instances (Lee et al., 2018).

The lidar setup for own ground truth generation de-

pends on the task, but the setups of the vehicles used

to generate data sets (Caesar et al., 2020) (Geyer et al.,

2020) can serve as a source of inspiration. Looking

to other radar tasks such as odometry or SLAM, the

ground truth is a motion measurement respective a

pose. (Maddern et al., 2020) describe the procedure

to post-process raw Global Positioning System (GPS)

data and data from an inertial measurement unit to re-

ceive a centimeter-accurate representation of the vehi-

cle’s movements and positions during the data record-

ing. Techniques like these allow for comparison be-

tween distinct algorithm modiﬁcations.

VEHITS 2024 - 10th International Conference on Vehicle Technology and Intelligent Transport Systems

306

Table 1: Recent data sets of autonomous driving including radar data.

Data Set Author and Year Radar Sensors Advantages Drawbacks

NuScenes (Caesar et al., 2020) 5x Continental • calibrated camera data provided • radar points sparse in near range

ARS 408-21 • lidar data provided • vacant spots at side of the vehicle

(2Tx/6Rx) ea. near/far • variance of situations

Oxford Radar RobotCar (Barnes et al., 2020) 1x Navtech CTS350-X • lidar data included • no conventional multi-radar setup

• odometry data included • no radar point cloud provided

• stereo-camera images provided • no annotations provided

RadarScenes (Schumann et al., 2021) 4x 77 GHz corner radar • semantic class annotation per radar point • no 360° radar

• documentary camera only

RADIATE (Sheeny et al., 2021) 1x Navtech CTS350-X • focus on adversarial weather conditions • no conventional multi-radar setup

• lidar data provided • no radar point cloud provided

• no 360° camera coverage

• no annotations provided

CARRADA (Ouaknine et al., 2021) 1x 77 GHz radar • good camera images provided • no 360° radar/camera view provided

(2Tx/4Rx) • range/azimuth annotations provided • only static scenes

• numerous VRUs in object selection • 4 GHz sweep used

RADDet (Zhang et al., 2021) 1x AWR1843-BOOST • annotations per radar point available • no adversarial weather conditions

(3Tx/4Rx) • stereo-camera images provided • only static scenes

• only one radar

CRUW (Wang et al., 2021) 1x TI AWR1843 • camera-only annotations provided • no 360° radar/camera view provided

(3Tx/4Rx) • camera-radar-fused annotations provided • no radar point annotations provided

• good distribution of various road types • no adversarial weather conditions

K-Radar (Paek et al., 2022) 1x 4D radar 77GHz • 4D tensor with elevation information provided • only one radar

• lidar data provided • no radar point annotations provided

• 360° camera data provided

• good road type and weather distribution

4 DISCUSSION

As this work points out, the potential of radar sen-

sors for automated driving and parking functions is

notable. When it comes to environmental perception

in poor weather and illumination conditions or mea-

suring the relative velocity of an object in a direct

manner, there is no way around this technique.

However, while a large volume of works addresses

the overall performance and accuracy of e.g., SLAM

or object detection, there is minor research focused

on how radar sensors could provide error guarantees.

It can be stated that rethinking needs to happen in

the signal processing, where the works comprised in

this survey rely on simple probabilistic ﬁltering meth-

ods. such as CFAR. At this, sensors stay behind their

potential in radar target detection, since a criticality-

based threshold adaption would outperform targeting

constant false alarm rates. Failure can already occur

by mistakes in the hardware design. Thus, the possi-

ble error sources must be investigated during the de-

sign process, as this paves the way for a general eval-

uation of failure rates.

The safety term got extended recently by the SO-

TIF deﬁnition, and various work is done to describe

hazard analyses. However, the perception elements

are hardly included in these considerations. Concepts

like the safety distance are a meaningful foundation,

but propagation through the whole perception task,

respecting the radar speciﬁcs, still needs to be done.

It is essential to process radar points differently and

to make the difference not with the probability of the

situation, but with its effect, especially regarding the

harm of VRUs. A raw point in front of the vehicle in

its passage route needs to be treated completely dif-

ferent than a raw point behind the vehicle, which will

not be hit in the near future. Not many works respond

to this new point of view, as the survey showed.

Handling the still enormous amount of potentially

critical raw points will only be possible by consider-

ing data of more than one sensor. The fact that re-

search is still at its beginning is reﬂected in the small

number of works which were presented for such sen-

sor fusion algorithms which aim at providing better

safety. While a lot of works focus on improving the

overall detection performance and stability, none of

the works try to conﬁrm or disprove the object exis-

tence at locations of radar points, which could advan-

tageously be approached by radar-camera fusion or

multi-camera fusion.

The sparseness of work on this area continues

when it comes to suitable data sets for the develop-

ment of such algorithms. We did not ﬁnd one radar

data set containing critical road situations only. It is

desirable to have a data set containing different corner

case scenarios including critical scenarios with VRUs

for radar sensors, in diverse weather and illumination

conditions, with the most important perturbing effects

included. Data should contain radar data as a point

cloud, calibrated camera data, annotated 3D bound-

ing boxes, and vehicle odometry to infer the position

from the starting point.

The new standards which are resumed in this pa-

per show that level 3 or level 4 automated functions

will not ﬁnd their way to the road without an approval

with a pervasive failure-in-time guarantee as the most

important component. This guarantee requires a new

way to process data of the perception sensors, such as

Making Radar Detections Safe for Autonomous Driving: A Review

307

radar, lidar or cameras. If no special action for such

systems is taken at the beginning of the design pro-

cess, long time before veriﬁcation and validation of

these systems become imminent, the approval will be

almost impossible.

5 CONCLUSION AND OUTLOOK

In this paper, we pointed out the notable potential

which radar sensors show to promote automated driv-

ing functions. The diversity of tasks which can be

tackled by radar sensors is shown, making their usage

attractive. Next, the current progress in the area of

radar hardware development and signal processing is

shown. These works serve as a basis for the consec-

utive inspection of safety-aware radar data process-

ing. To explain the line of reasoning, we scrutinized

interpretations of the safety term in context of auto-

mated driving functions. The role of critical false pos-

itive radar points was highlighted. Selected, safety-

focused works in the ﬁeld of function evaluation were

investigated to work out how the perception data can

be connected to the criticality of scenarios. As en-

richment of data points with information is indispens-

able to declare their signiﬁcance, the area of sensor

data fusion is illuminated. Radar-camera fusion and

multi-radar-fusion were especially emphasized in this

survey. We ended with giving an overview about cur-

rent radar data sets and rate them with respect to the

usability for safety-aware function development.

It is recognized that future work should be done

investigating the impact of erroneously ﬁltered posi-

tive radar points on the individual tasks described in

this work. A method to identify critical situations is

important to control enriching the right regions with

additional information, originating e.g. from camera

setups. Pursuing these aspects will bring the research

community one step closer to perception which is safe

under a guarantee.

REFERENCES

Abdulazim, A., Elbahaey, M., and Mohamed, A. (2021).

Putting Safety of Intended Functionality SOTIF into

Practice. Technical report, Warrendale, PA.

Adolfsson, D., Magnusson, M., Alhashimi, A., Lilienthal,

A. J., and Andreasson, H. (2021). CFEAR Radaro-

dometry - Conservative Filtering for Efﬁcient and Ac-

curate Radar Odometry. In 2021 IEEE/RSJ Interna-

tional Conference on Intelligent Robots and Systems

(IROS), pages 5462–5469. ISSN: 2153-0866.

Aptiv PLC (2023). Gen7 radar family.

Aydogdu, C., Keskin, M. F., Carvajal, G. K., Eriksson, O.,

Hellsten, H., Herbertsson, H., Nilsson, E., Rydstrom,

M., Vanas, K., and Wymeersch, H. (2020). Radar

Interference Mitigation for Automated Driving: Ex-

ploring Proactive Strategies. IEEE Signal Processing

Magazine, 37(4):72–84.

Barnes, D., Gadd, M., Murcutt, P., Newman, P., and Posner,

I. (2020). The Oxford Radar RobotCar Dataset: A

Radar Extension to the Oxford RobotCar Dataset. In

2020 IEEE International Conference on Robotics and

Automation (ICRA), pages 6433–6438. ISSN: 2577-

087X.

Barnes, D. and Posner, I. (2020). Under the Radar: Learn-

ing to Predict Robust Keypoints for Odometry Estima-

tion and Metric Localisation in Radar. In 2020 IEEE

International Conference on Robotics and Automation

(ICRA), pages 9484–9490. ISSN: 2577-087X.

Bila, C., Sivrikaya, F., Khan, M. A., and Albayrak, S.

(2017). Vehicles of the Future: A Survey of Research

on Safety Issues. IEEE Transactions on Intelligent

Transportation Systems, 18(5):1046–1065.

Bilik, I. (2023). Comparative Analysis of Radar and

Lidar Technologies for Automotive Applications.

IEEE Intelligent Transportation Systems Magazine,

15(1):244–269.

Bilik, I., Longman, O., Villeval, S., and Tabrikian, J. (2019).

The Rise of Radar for Autonomous Vehicles: Signal

Processing Solutions and Future Research Directions.

IEEE Signal Processing Magazine, 36(5):20–31.

Brazil, G. and Liu, X. (2019). M3d-rpn: Monocular 3d

region proposal network for object detection. In Pro-

ceedings of the IEEE/CVF International Conference

on Computer Vision (ICCV).

Buckland, M. and Gey, F. (1994). The relationship between

Recall and Precision. Journal of the American Society

for Information Science, 45(1):12–19.

Caesar, H., Bankiti, V., Lang, A. H., Vora, S., Liong, V. E.,

Xu, Q., Krishnan, A., Pan, Y., Baldan, G., and Bei-

jbom, O. (2020). nuscenes: A multimodal dataset for

autonomous driving. In Proceedings of the IEEE/CVF

Conference on Computer Vision and Pattern Recogni-

tion (CVPR).

Chang, S., Zhang, Y., Zhang, F., Zhao, X., Huang, S., Feng,

Z., and Wei, Z. (2020). Spatial Attention Fusion for

Obstacle Detection Using MmWave Radar and Vision

Sensor. Sensors, 20(4):956.

Chen, M., Belgiovane, D., and Chen, C.-C. (2014). Radar

characteristics of pedestrians at 77 GHz. In 2014 IEEE

Antennas and Propagation Society International Sym-

posium (APSURSI), pages 2232–2233. ISSN: 1947-

1491.

Chu, J., Zhao, T., Jiao, J., Yuan, Y., and Jing, Y. (2023).

SOTIF-Oriented Perception Evaluation Method for

Forward Obstacle Detection of Autonomous Vehicles.

IEEE Systems Journal, 17(2):2319–2330.

Diehl, C., Feicho, E., Schwambach, A., Dammeier, T.,

Mares, E., and Bertram, T. (2020). Radar-based Dy-

namic Occupancy Grid Mapping and Object Detec-

tion. In 2020 IEEE 23rd International Conference on

Intelligent Transportation Systems (ITSC), pages 1–6.

VEHITS 2024 - 10th International Conference on Vehicle Technology and Intelligent Transport Systems

308

Dietmayer, K. (2016). Predicting of Machine Perception for

Automated Driving. pages 407–424. Springer, Berlin,

Heidelberg.

Engels, F., Heidenreich, P., Wintermantel, M., St

acker, L.,

Al Kadi, M., and Zoubir, A. M. (2021). Automotive

Radar Signal Processing: Research Directions and

Practical Challenges. IEEE Journal of Selected Topics

in Signal Processing, 15(4):865–878.

Gerstmair, M., Melzer, A., Onic, A., and Huemer, M.

(2019). On the Safe Road Toward Autonomous Driv-

ing: Phase Noise Monitoring in Radar Sensors for

Functional Safety Compliance. IEEE Signal Process-

ing Magazine, 36(5):60–70.

Geyer, J., Kassahun, Y., Mahmudi, M., Ricou, X., Durgesh,

R., Chung, A. S., Hauswald, L., Hoang Pham, V.,

uhlegg, M., Dorn, S., Fernandez, T., J

anicke, M.,

Mirashi, S., Savani, C., Sturm, M., Vorobiov, O.,

Oelker, M., Garreis, S., and Schuberth, P. (2020).

A2D2: Audi Autonomous Driving Dataset. Technical

report. ADS Bibcode: 2020arXiv200406320G Type:

article.

Girshick, R. (2015). Fast r-cnn. In Proceedings of the IEEE

International Conference on Computer Vision (ICCV).

Hong, Z., Petillot, Y., Wallace, A., and Wang, S.

(2021). Radar SLAM: A Robust SLAM Sys-

tem for All Weather Conditions. Technical report.

arXiv:2104.05347 [cs] type: article.

Hoss, M., Scholtes, M., and Eckstein, L. (2022). A Re-

view of Testing Object-Based Environment Percep-

tion for Safe Automated Driving. Automotive Inno-

vation, 5(3):223–250.

International Organization for Standardization (2018). ISO

26262:2018 - Road vehicles - Functional safety. Tech-

nical report.

International Organization for Standardization (2022). ISO

21448:2022 - Road vehicles - Safety of the intended

functionality. Technical report.

John, V. and Mita, S. (2019). RVNet: Deep Sensor Fusion

of Monocular Camera and Radar for Image-Based Ob-

stacle Detection in Challenging Environments. In Lee,

C., Su, Z., and Sugimoto, A., editors, Image and

Video Technology, Lecture Notes in Computer Sci-

ence, pages 351–364, Cham. Springer International

Publishing.

Kellner, D., Barjenbruch, M., Klappstein, J., Dickmann,

J., and Dietmayer, K. (2014). Instantaneous ego-

motion estimation using multiple Doppler radars. In

2014 IEEE International Conference on Robotics and

Automation (ICRA), pages 1592–1597. ISSN: 1050-

4729.

Khaleghi, B., Khamis, A., Karray, F. O., and Razavi, S. N.

(2013). Multisensor data fusion: A review of the state-

of-the-art. Information Fusion, 14(1):28–44.

Khatun, M., Glaß, M., and Jung, R. (2020). Scenario-based

extended hara incorporating functional safety & so-

tif for autonomous driving. In Baraldi, P., editor, E-

proceedings of the 30th European Safety and Reli-

ability Conference and 15th Probabilistic Safety As-

sessment and Management Conference, volume 2020

of (ESREL2020 PSAM15) ; 01 - 05 November 2020

Venice, Italy, pages 53 – 59.

Knott, E. F. (2012). Radar cross section measurements.

Springer Science & Business Media.

Lee, J., Walsh, S., Harakeh, A., and Waslander, S. L. (2018).

Leveraging Pre-Trained 3D Object Detection Models

for Fast Ground Truth Generation. In 2018 21st In-

ternational Conference on Intelligent Transportation

Systems (ITSC), pages 2504–2510. ISSN: 2153-0017.

Li, M., Feng, Z., Stolz, M., Kunert, M., Henze, R., and

uc¸

ukay, F. (2018). High resolution radar-based oc-

cupancy grid mapping and free space detection.

Li, P., Wang, P., Berntorp, K., and Liu, H. (2022). Ex-

ploiting temporal relations on radar perception for au-

tonomous driving. In Proceedings of the IEEE/CVF

Conference on Computer Vision and Pattern Recogni-

tion (CVPR), pages 17071–17080.

Maddern, W., Pascoe, G., Gadd, M., Barnes, D., Yeomans,

B., and Newman, P. (2020). Real-time Kinematic

Ground Truth for the Oxford RobotCar Dataset. Tech-

nical report. ADS Bibcode: 2020arXiv200210152M

Type: article.

Marti, E., de Miguel, M. A., Garcia, F., and Perez, J. (2019).

A Review of Sensor Technologies for Perception in

Automated Driving. IEEE Intelligent Transportation

Systems Magazine, 11(4):94–108.

Muhammad, K., Ullah, A., Lloret, J., Ser, J. D., and de Al-

buquerque, V. H. C. (2021). Deep Learning for Safe

Autonomous Driving: Current Challenges and Future

Directions. IEEE Transactions on Intelligent Trans-

portation Systems, 22(7):4316–4336.

Ouaknine, A., Newson, A., Rebut, J., Tupin, F., and P

erez,

P. (2021). CARRADA Dataset: Camera and Auto-

motive Radar with Range- Angle- Doppler Annota-

tions. In 2020 25th International Conference on Pat-

tern Recognition (ICPR), pages 5068–5075. ISSN:

1051-4651.

Paek, D.-H., KONG, S.-H., and Wijaya, K. T. (2022). K-

radar: 4d radar object detection for autonomous driv-

ing in various weather conditions. In Koyejo, S., Mo-

hamed, S., Agarwal, A., Belgrave, D., Cho, K., and

Oh, A., editors, Advances in Neural Information Pro-

cessing Systems, volume 35, pages 3819–3829. Cur-

ran Associates, Inc.

Patole, S. M., Torlak, M., Wang, D., and Ali, M. (2017).

Automotive radars: A review of signal process-

ing techniques. IEEE Signal Processing Magazine,

34(2):22–35.

Peng, L., Li, B., Yu, W., Yang, K., Shao, W., and Wang, H.

(2023). SOTIF Entropy: Online SOTIF Risk Quantiﬁ-

cation and Mitigation for Autonomous Driving. IEEE

Transactions on Intelligent Transportation Systems,

pages 1–17.

Peng, L., Wang, H., and Li, J. (2021). Uncertainty Evalua-

tion of Object Detection Algorithms for Autonomous

Vehicles. Automotive Innovation, 4(3):241–252.

Philipp, R., Schuldt, F., and Howar, F. (2020). Functional

decomposition of automated driving systems for the

classiﬁcation and evaluation of perceptual threats. In

13. Uni-DAS eV Workshop Fahrerassistenz und au-

tomatisiertes Fahren 2020. Walting.

Making Radar Detections Safe for Autonomous Driving: A Review

309

Popov, A., Gebhardt, P., Chen, K., and Oldja, R. (2023).

NVRadarNet: Real-Time Radar Obstacle and Free

Space Detection for Autonomous Driving. In 2023

IEEE International Conference on Robotics and Au-

tomation (ICRA), pages 6958–6964.

Powers, D. (2011). Evaluation: From precision, recall and

f-measure to roc, informedness, markedness & cor-

relation. Journal of Machine Learning Technologies,

2(1):37–63.

Robert Bosch GmbH (2023). Front-radar premium.

SAE International. (2021). J3016 - Taxonomy and Deﬁni-

tions for Terms Related to Driving Automation Sys-

tems for On-Road Motor Vehicles. Technical report.

Sayyah, K., Sarkissian, R., Patterson, P., Huang, B., Eﬁmov,

O., Kim, D., Elliott, K., Yang, L., and Hammon, D.

(2022). Fully Integrated FMCW LiDAR Optical En-

gine on a Single Silicon Chip. Journal of Lightwave

Technology, 40(9):2763–2772.

Scheiner, N., Kraus, F., Appenrodt, N., Dickmann, J., and

Sick, B. (2021). Object detection for automotive radar

point clouds – a comparison. AI Perspectives, 3(1):6.

Sch

onemann, V. (2019). Safety requirements and distribu-

tion of functions for automated valet parking.

Schumann, O., Hahn, M., Scheiner, N., Weishaupt, F.,

Tilly, J. F., Dickmann, J., and W

ohler, C. (2021).

RadarScenes: A Real-World Radar Point Cloud Data

Set for Automotive Applications. In 2021 IEEE 24th

International Conference on Information Fusion (FU-

SION), pages 1–8.

Schumann, O., Lombacher, J., Hahn, M., W

ohler, C., and

Dickmann, J. (2020). Scene Understanding With Au-

tomotive Radar. IEEE Transactions on Intelligent Ve-

hicles, 5(2):188–203.

Schuster, F., Keller, C. G., Rapp, M., Haueis, M., and Curio,

C. (2016). Landmark based radar SLAM using graph

optimization. In 2016 IEEE 19th International Con-

ference on Intelligent Transportation Systems (ITSC),

pages 2559–2564. ISSN: 2153-0017.

Sheeny, M., De Pellegrin, E., Mukherjee, S., Ahrabian,

A., Wang, S., and Wallace, A. (2021). RADIATE:

A Radar Dataset for Automotive Perception in Bad

Weather. In 2021 IEEE International Conference on

Robotics and Automation (ICRA), pages 1–7. ISSN:

2577-087X.

Steinbaeck, J., Steger, C., Holweg, G., and Druml, N.

(2017). Next generation radar sensors in automotive

sensor fusion systems. In 2017 Sensor Data Fusion:

Trends, Solutions, Applications (SDF), pages 1–6.

Sun, C., Zhang, R., Lu, Y., Cui, Y., Deng, Z., Cao, D.,

and Khajepour, A. (2023). Toward Ensuring Safety

for Autonomous Driving Perception: Standardization

Progress, Research Advances, and Perspectives. IEEE

Transactions on Intelligent Transportation Systems,

pages 1–19.

Tang, X., Zhang, Z., and Qin, Y. (2022). On-Road Object

Detection and Tracking Based on Radar and Vision

Fusion: A Review. IEEE Intelligent Transportation

Systems Magazine, 14(5):103–128.

Velasco-Hernandez, G., Yeong, D. J., Barry, J., and Walsh,

J. (2020). Autonomous Driving Architectures, Per-

ception and Data Fusion: A Review. In 2020 IEEE

16th International Conference on Intelligent Com-

puter Communication and Processing (ICCP), pages

315–321.

Waldschmidt, C., Hasch, J., and Menzel, W. (2021). Auto-

motive Radar — From First Efforts to Future Systems.

IEEE Journal of Microwaves, 1(1):135–148.

Wang, Y., Jiang, Z., Gao, X., Hwang, J.-N., Xing, G.,

and Liu, H. (2021). Rodnet: Radar object detec-

tion using cross-modal supervision. In Proceedings

of the IEEE/CVF Winter Conference on Applications

of Computer Vision (WACV), pages 504–513.

Xiao, P., Shao, Z., Hao, S., Zhang, Z., Chai, X., Jiao, J., Li,

Z., Wu, J., Sun, K., Jiang, K., Wang, Y., and Yang, D.

(2021). PandaSet: Advanced Sensor Suite Dataset for

Autonomous Driving. In 2021 IEEE International In-

telligent Transportation Systems Conference (ITSC),

pages 3095–3101.

Xu, F., Wang, H., Hu, B., and Ren, M. (2020). Road Bound-

aries Detection based on Modiﬁed Occupancy Grid

Map Using Millimeter-wave Radar. Mobile Networks

and Applications, 25(4):1496–1503.

Yang, M., Wang, S., Bakita, J., Vu, T., Smith, F. D., Ander-

son, J. H., and Frahm, J.-M. (2019). Re-thinking cnn

frameworks for time-sensitive autonomous-driving

applications: Addressing an industrial challenge. In

2019 IEEE Real-Time and Embedded Technology

and Applications Symposium (RTAS), pages 305–317.

IEEE.

Yeong, D. J., Velasco-Hernandez, G., Barry, J., and

Walsh, J. (2021). Sensor and Sensor Fusion Tech-

nology in Autonomous Vehicles: A Review. Sensors,

21(6):2140.

Yu, J. and Krolik, J. (2012). MIMO multipath clutter mit-

igation for GMTI automotive radar in urban environ-

ments. In IET International Conference on Radar Sys-

tems (Radar 2012), pages 1–5.

Yurtsever, E., Lambert, J., Carballo, A., and Takeda, K.

(2020). A Survey of Autonomous Driving: Common

Practices and Emerging Technologies. IEEE Access,

8:58443–58469.

Zang, S., Ding, M., Smith, D., Tyler, P., Rakotoarivelo, T.,

and Kaafar, M. A. (2019). The Impact of Adverse

Weather Conditions on Autonomous Vehicles: How

Rain, Snow, Fog, and Hail Affect the Performance of

a Self-Driving Car. IEEE Vehicular Technology Mag-

azine, 14(2):103–111.

Zhang, A., Nowruzi, F. E., and Laganiere, R. (2021). RAD-

Det: Range-Azimuth-Doppler based Radar Object

Detection for Dynamic Road Users. In 2021 18th

Conference on Robots and Vision (CRV), pages 95–

102.

Zhou, T., Yang, M., Jiang, K., Wong, H., and Yang, D.

(2020). MMW Radar-Based Technologies in Au-

tonomous Driving: A Review. Sensors, 20(24):7283.

Zhou, Y., Dong, Y., Hou, F., and Wu, J. (2022). Review on

Millimeter-Wave Radar and Camera Fusion Technol-

ogy. Sustainability, 14(9):5114.

VEHITS 2024 - 10th International Conference on Vehicle Technology and Intelligent Transport Systems

310