Ci a ion: Gomes, T.; Ma ias, D.;
Campos, A.; Cunha, L.; Ro iz, R. A
Su ey on G ound Segmen a ion
Me hods o Au omo i e LiDAR
Senso s. Senso s 2023,23, 601.
h ps://doi.o g/10.3390/s23020601
Academic Edi o s: I-Hsi Kao,
Yi-Ho ng Lai, Jau-Woei Pe ng and
Ching-Yao Chan
Recei ed: 8 Decembe 2022
Re ised: 27 Decembe 2022
Accep ed: 29 Decembe 2022
Published: 5 Janua y 2023
Copy igh : © 2023 by he au ho s.
Licensee MDPI, Basel, Swi ze land.
This a icle is an open access a icle
dis ibu ed unde he e ms and
condi ions o he C ea i e Commons
A ibu ion (CC BY) license (h ps://
c ea i ecommons.o g/licenses/by/
4.0/).
senso s
Re iew
A Su ey on G ound Segmen a ion Me hods o Au omo i e
LiDAR Senso s
Tiago Gomes * , Diogo Ma ias , And é Campos , Luís Cunha and Rica do Ro iz
Cen o ALGORITMI/LASI, Escola de Engenha ia, Uni e sidade do Minho, 4800-058 Guima aes, Po ugal
*Co espondence: m [email p o ec ed]
Abs ac :
In he nea u u e, au onomous ehicles wi h ull sel -d i ing ea u es will popula e ou
public oads. Howe e , ully au onomous ca s will equi e obus pe cep ion sys ems o sa ely
na iga e he en i onmen , which includes came as, RADAR de ices, and Ligh De ec ion and
Ranging (LiDAR) senso s. LiDAR is cu en ly a key senso o he u u e o au onomous d i ing since
i can ead he ehicle’s icini y and p o ide a eal- ime 3D isualiza ion o he su oundings h ough
a poin cloud ep esen a ion. These ea u es can assis he au onomous ehicle in se e al asks, such
as objec iden i ica ion and obs acle a oidance, accu a e speed and dis ance measu emen s, oad
na iga ion, and mo e. Howe e , i is c ucial o de ec he g ound plane and oad limi s o sa ely
na iga e he en i onmen , which equi es ex ac ing in o ma ion om he poin cloud o accu a ely
de ec common oad bounda ies. This a icle p esen s a su ey o exis ing me hods used o de ec and
ex ac g ound poin s om LiDAR poin clouds. I summa izes he al eady ex ensi e li e a u e and
p oposes a comp ehensi e axonomy o help unde s and he cu en g ound segmen a ion me hods
ha can be used in au omo i e LiDAR senso s.
Keywo ds: au onomous d i ing; LiDAR; pe cep ion sys em; g ound segmen a ion; su ey
1. In oduc ion
In he e y nea u u e, au onomous ehicles wi h ull sel -d i ing ea u es will
ci cula e on ou public oads [
1
]. None heless, mos in elligen ehicles oday a e s ill
manually con olled, co esponding o he i s h ee le els o d i ing au oma ion de ined
by he socie y o au omo i e enginee s (SAE), i.e., he six le els o au oma ion equi ed
o each ull d i ing au oma ion ea u es. Acco ding o he SAE J3016 s anda d ( e .
202104) [
2
], in le els 0, 1, and 2, he d i e mus ac i ely moni o he d i ing ac i i ies,
while in le els 3, 4, and 5, he au oma ed ehicle should be able o moni o and na iga e
he en i onmen , au onomously. Despi e cu en ca s suppo ing only ea u es up o
SAE-le el 2, e.g., mode n ad anced d i e -assis ance sys ems (ADAS) can al eady p o ide
pa ial ehicle’s au oma ion, ca manu ac u e s a e only now ecei ing app o als o
SAE-le el 3 [
3
,
4
], which de ines explici ly ha he au oma ed d i ing unc ion can ake
o e ce ain d i ing asks. Howe e , a d i e is s ill equi ed and mus be eady o ake
con ol o he ca a all imes when p omp ed o in e ene by he ehicle. Achie ing sa e
au oma ed d i ing ea u es equi es obus and eliable pe cep ion sys ems [
5
] ha depend
on mul iple senso se ups o na iga e he en i onmen , which usually include came as,
RADAR de ices, and Ligh De ec ion and Ranging (LiDAR) senso s [
6
–
8
]. Fo ins ance, he
Ge man Fede al Mo o T anspo Au ho i y (KBA) has inally g an ed he i s SAE-le el
3 and UN Regula ion numbe 157 [
4
] app o al o Me cedes-Benz [
3
], whose pe cep ion
sys em includes a we ness senso in he wheel, LiDAR senso s, mic ophones, and a came a
in he ea window, p ima ily used o de ec ing blue ligh s and o he special signals om
eme gency ehicles. This echnical app o al was mainly possible due o he adop ion o
LiDAR o na iga e he en i onmen .
A LiDAR senso wo ks by i ing a lase pulse o a a ge and cap u ing he e lec ed
signal, whe e he dis ance o he a ge is ob ained by calcula ing he ound- ip ime o he
Senso s 2023,23, 601. h ps://doi.o g/10.3390/s23020601 h ps://www.mdpi.com/jou nal/senso s
Senso s 2023,23, 601 2 o 30
a eling ligh . I s ou pu is a 3D poin cloud ha can be used o e ain mapping and ec e-
a ing he su ounding en i onmen . Thus, his echnology has become widely popula in
se e al ai bo ne lase scanning (ALS) applica ions, such as a chaeology
[9–11]
, geology [
12
],
o es y [
13
,
14
], geog aphy and opog aphy [
14
–
18
], su eying [
14
], lase al ime y [
19
], and
many mo e [
14
]. Due o i s wide success in such domains, LiDAR senso s ecen ly s a ed
o be adop ed by he au omo i e indus y in he pe cep ion sys em o he ca . Thei 3D poin
cloud can be e y use ul in se e al au onomous d i ing applica ions
[20–22]
, such as obs a-
cles, objec s, and ehicles de ec ion [
23
–
26
]; pedes ians ecogni ion and acking [
27
,
28
];
g ound segmen a ion o oad de ec ion and
na iga ion [29]
; among o he s [
30
]. Because
his echnology wo ks wi h ac i e illumina ion, LiDAR senso s allow ound- he-clock
obse a ions, p o iding accu a e measu emen s o he ehicle’s icini y up o hund eds o
me e s. Howe e , se e al challenges may a ec he p ocessing o he ecei ed poin cloud
such as LiDAR mu ual in e e ence [
31
–
33
], and ad e se wea he
[34–38]
. Addi ionally,
because a high- esolu ion senso can p oduce a conside able amoun o da a, e.g., he
Velodyne senso VLS-128 can ou pu up o 9.6M poin s pe second, i is impo an o handle
he poin cloud be o e being deli e ed o high-le el applica ions, bo h in e ms o packe
handling [39], da a comp ession [40,41], and poin cloud denoising [36].
Rega ding au onomous na iga ion capabili ies, he ehicle mus pe cei e he su -
oundings by unde s anding he loca ion and shape o he d i ing en i onmen . Thus, oad
bounda ies, such as cu bs, asphal be ms, walls, and o he geome ic ea u es, a e ypical
oad cha ac e is ics ha he pe cep ion sys em mus au oma ically de ec o na iga e he
en i onmen sa ely [
42
]. Howe e , de ec ing such ea u es equi es he pe cep ion sys em
o pe o m g ound and objec segmen a ion echniques o e icien ly iden i y o emo e
undesi ed objec da a [
23
,
43
–
45
]. Because his opic has been ho oughly s udied in he las
ew yea s, he e is a ple ho a o g ound segmen a ion me hods in he li e a u e ha can be
used o ex ac oad bounda y in o ma ion om he poin cloud. Because such me hods
can ollow di e en app oaches, he e a e se e al ade-o s, e.g., accu acy, pe o mance,
compu ing equi emen s, e c., ha mus be conside ed be o e choosing he bes and he
mos app op ia e me hod o deploy in he pe cep ion sys em.
To help in unde s anding he exis ing g ound segmen a ion me hods in he li e a u e,
he main con ibu ions o his a icle a e as ollows: (1) a li e a u e e iew on s a e-o - he-a
app oaches o de ec and ex ac g ound poin s om LiDAR da a, applied in an au omo i e
scena io; (2) a comp ehensi e axonomy p oposal wi h i e high-le el ca ego ies, ollowed
by a discussion o he echnical aspec s and algo i hms o he app oaches ha can cu en ly
be ound in he li e a u e; (3) and a quali a i e compa ison be ween he di e en ca ego ies
ega ding impo an me ics, including eal- ime, esou ce equi emen s, and he algo-
i hm’s pe o mance in di e en en i onmen s and aspec s. The emainde o his a icle is
o ganized as ollows: Sec ion 2p esen s he concep s and applica ions behind LiDAR, while
Sec ion 3shows cu en g ound segmen a ion me hods and p oposed axonomy. Sec ion 4
discusses and gi es a quali a i e compa ison be ween he s a e-o - he-a con ibu ions.
Finally, Sec ion 5concludes his a icle.
2. Au omo i e LiDAR Senso s
Fo he las 20 yea s, au onomous d i ing has been a well-es ablished esea ch opic.
Howe e , despi e some p o o ypes being al eady a ailable and unning, he au omo i e
indus y is only now s a ing o p o ide comme cial p oduc s, which will soon u n au-
onomous d i ing ehicles mains eam. Aiming a eaching SAE-le el 5, ehicles a e being
equipped wi h se e al ADAS ea u es, which a e cons an ly being imp o ed e e y yea .
This pushes esea ch o ad ance in di e en sensing echnologies and in c ea ing mo e
obus pe cep ion sys ems. Figu e 1depic s a ypical mul i-senso pe cep ion sys em com-
posed o came as, RADAR, and LiDAR senso s, which can help in p o iding c oss- a ic
ale s, blind spo assis [
46
], Adap i e C uise Con ol (ACC) [
47
], pedes ian de ec ion and
a oidance [
48
], and Au oma ic Eme gency B aking (AEB) [
49
,
50
]. Toge he , all hese sen-
so s allow e ie ing edundan in o ma ion abou he su ounding en i onmen , ensu ing
Senso s 2023,23, 601 3 o 30
ha high-le el au onomous d i ing decisions a e made based on accu a e ep esen a ions
o he ehicle’s icini y.
Figu e 1. Mul i-senso pe cep ion sys em.
Among all senso s, LiDAR is becoming he mos impo an in au onomous appli-
ca ions as i can p o ide up o 300 m o eal- ime 3D in o ma ion abou he ehicle’s
su oundings. Compa ed wi h RADAR, which sha es he same wo king p inciple o mea-
su ing dis ances based on he ound- ip ime o an emi ed signal, LiDAR can ope a e
in much highe equencies due o he ligh p ope ies. Mo eo e , while RADAR senso s
only p o ide angula esolu ions o a mos 1 deg ee, which is insu icien o objec shape
es ima ion [
51
], LiDAR senso s can achie e esolu ions o en hs o deg ees. Thus, hei
ou pu can be used o objec segmen a ion a he han solely objec de ec ion. None heless,
because LiDAR senso s use ligh , hey s ill p esen disad an ages when compa ed wi h
adio-based senso s. Since ligh has high abso p ion in wa e , ad e se wea he condi ions,
such as hea y ain all, can a ec he o e all pe o mance o LiDAR. On he o he hand,
he highe wa eleng hs used in RADAR can p esen good pe o mance in poo wea he ,
making his senso capable o co e ing longe dis ances bu wi h less emi ed powe .
Rega ding he esolu ion o he ou pu da a, came as can achie e colo ed and high-
esolu ion in o ma ion abou he d i ing en i onmen . Howe e , hey a e subjec o se e al
ligh - ela ed issues, e.g., hey canno co ec ly wo k a nigh o in he p esence o in ense
ligh sou ces, which can u n he segmen a ion asks elying solely on came as qui e
challenging. Addi ionally, dep h in o ma ion abou he en i onmen is only possible wi h
s e eo came as and u he image p ocessing, which a ec s he dis ance measu emen s.
2.1. LiDAR Technology
The e is a ple ho a o LiDAR senso s cu en ly a ailable on he ma ke [
21
,
52
]. Despi e
all sha ing he same ligh -based ope a ion p ope ies, manu ac u e s al eady p o ide
di e en measu emen and imaging sys ems app oaches, which ansla es in o di e en
pe o mance me ics and o e all cos s. As depic ed in Figu e 2, a LiDAR senso uses a
ligh signal o measu e dis ances, which a e calcula ed based on he ound- ip delay (
τ
)
be ween a signal emi ed by a lase and e lec ed by a a ge . Since he speed o ligh (
c
) is
p e iously known, he dis ance (R) o he a ge is calcula ed using Equa ion (1).
R=
1
2cτ(1)
Senso s 2023,23, 601 4 o 30
T ansmi ed
Signal
Backsca e ed
Signal
Emi e
Recei e
LiDAR RANGE
ToF
TARGET
Figu e 2. LiDAR wo king p inciple.
2.1.1. Measu emen Techniques
Depending on he senso ’s echnology, he ound- ip delay, also known as he Time-
o -Fligh (ToF), can be measu ed h ough di e en echniques [
53
,
54
]: pulsed, Ampli ude
Modula ed Con inuous Wa e (AMCW), and F equency Modula ed Con inuous Wa e
(FMCW). Measu emen echniques based on pulsed signals a e mo e s aigh o wa d o
implemen as hey only equi e accu a e ime s o Time- o-Digi al Con e e s (TDC) o
di ec ly measu e and calcula e he ToF. Due o hei simplici y, hey o e a low-cos and
small-size implemen a ion bu a e mo e subjec o low Signal-Noise Ra io (SNR), which
limi s he accu acy o he measu emen s. Wi h AMCW and FMCW, ins ead o sending sho
high-in ensi y pulses, he senso emi s a modula ed con inuous wa e signal, consequen ly
achie ing be e SNR alues. In an AMCW sys em, he senso emi s a wa e synch onized
wi h a se ies o in eg a ing windows. Since he e lec ed wa e is expec ed o be ou o
phase, he dis ance can be calcula ed based on he ene gy a io p esen on each window:
sho dis ances esul in a e lec ed signal mo e p esen in he i s windows, while o long
dis ances, he e lec ed signal can be ound in he las windows.
Simila ly o pulsed-based senso s, AMCW can also p o ide a simple design. How-
e e , long- ange de ec ion equi es longe wa eleng hs o ensu e ha he phase di e ence
be ween signals does no exceed he signal’s pe iod, which would p oduce dis ance ambi-
gui y. The e o e, hei applica ion is limi ed mainly o mid- and sho - ange senso s due o
eye sa e y egula ions. On he o he hand, FMCW sys ems modula e he emi ing signal
equency wi h an up-chi p signal ha inc eases i s equency o e ime. Then, he senso
compu es he di e ence be ween he equency o he e u n signal and a local oscilla o ,
being his del a di ec ly p opo ional o he dis ance o he a ge . Despi e p o iding mo e
obus ness o ex e nal ligh sou ces and e en allowing o di ec ly e ie e he speed o a
mo ing a ge due o Dopple shi , he inc ease in op ical componen s makes FMCW-based
senso s mo e expensi e o build.
2.1.2. Imaging Techniques
To c ea e a 3D poin cloud ep esen a ion in eal ime, LiDAR senso s emi and collec
signals ac oss mul iple di ec ions wi hin hei suppo ed Field o View (FoV), which can
esul in a poin cloud holding millions o poin s. This is achie ed by using di e en
imaging echniques [
55
], such as solid-s a e, o o -based, and lash-based senso s. Velo-
dyne pionee ed o o -based LiDAR senso s. By ha ing a mechanical sys em ha spins he
scanning pa , hey can c ea e a 360
º
ho izon al FoV, while he numbe o exis ing emi -
e / ecep o pai s, also known as channels, de ine he e ical FoV. This echnology was so
success ul ha oday i is he mos used in LiDAR, which esul ed in se e al manu ac u e s
compe ing in his ma ke niche. Howe e , hey s ill come wi h some limi a ions, such as
p ice, bulkiness, and he educed ame a e caused by mechanical pa s.
Rega ding he exis ing solid-s a e scanning solu ions, some echnologies, e.g., mi o
MEMS [
56
] and Op ical Phased A ay (OPA) [
57
], a e being deployed o eplace he bulky
Senso s 2023,23, 601 5 o 30
o a ional pa s exis ing in o o -based senso s. Despi e achie ing as e scanning equen-
cies and o e ing a be e o e all design o mass p oduc ion (which educes he o e all
p oduc p ice), he senso ’s ange is s ill a limi a ion due o powe es ic ions in he lase
uni . Addi ionally, o applica ions equi ing la ge FoV, mul iple senso s mus be a ached
o he se up o gua an ee he ull co e age o he su ounding en i onmen . Aiming a
educing he limi a ions c ea ed by complex bulky, o iny s ee ing sys ems, a lash-based
LiDAR senso does no equi e any s ee ing sys em o emi lase signals ac oss he FoV.
Ins ead, i uses a lash o ligh o illumina e he en i e en i onmen while pho o-de ec o s
collec he back-sca e ed ligh . By simul aneously sha ing he ligh sou ce ac oss all he
FoV, cap u ing da a esul s in a as e ope a ion, making hese senso s highly immune o
ligh dis o ion [
58
]. Howe e , despi e he simplici y o he emi e , he ecei e sys em is
qui e complex since i mus be able o di e en ia e he e u ning ligh om each poin . In
o de o gua an ee a sui able senso spa ial esolu ion, hese senso s equi e high-densi y
pho o-de ec ion a ays, which makes hem mo e expensi e when compa ed wi h o he
solid-s a e solu ions.
2.2. LiDAR Applica ions
LiDAR is a key senso in a pe cep ion sys em since i s ou pu can be used o imp o ing
au onomous d i ing asks based on senso usion p ocessing [
59
], e.g., objec de ec ion
and segmen a ion o objec classi ica ion and collision a oidance [
23
–
25
], Simul aneous
Localiza ion and Mapping (SLAM) applica ions [
60
], he de ec ion and na iga ion o he
d i able a ea [
61
,
62
], and much mo e. The LiDAR ou pu is a poin cloud, which is an exac
3D image o he en i onmen cap u ed by he senso . Humans do no easily unde s and i s
in o ma ion a i s sigh , bu algo i hms can do a g ea job o in e p e ing i s con en .
2.2.1. Objec De ec ion and Classi ica ion
The e a e many me hods and algo i hms o de ec ing and classi ying objec s in
poin cloud da a, bu i s , i is necessa y o ind hem. A e con e ing aw da a in o
a poin cloud s uc u e, one o he i s s eps is he poin clus e ing o segmen a ion,
which basically consis s in g ouping poin s based on common cha ac e is ics [
23
]. A e
his s ep, edundan da a can be il e ed/ emo ed om he poin cloud, esul ing in less
da a o be ans e ed and p ocessed in he upcoming phases. In applica ions whe e he
senso keeps a s a iona y posi ion, some algo i hms s a by classi ying he poin cloud
in o backg ound and o eg ound da a [
24
,
25
]. Poin s ha sha e he same posi ion ac oss
mul iple ames a e conside ed backg ound, being disca ded as hey do no ep esen
dynamic objec s. Fo he emaining poin s ( o eg ound), he dis ance be ween poin s is
measu ed, and poin s close o each o he a e clus e ed and ma ked wi h a bounding box
as hey possibly ep esen an objec . Howe e , when he senso mo es wi h he ca , hese
app oaches a e no e ec i e, as he backg ound and objec s mo e oge he inside he poin
cloud. The e o e, au omo i e app oaches equi e obus and as e algo i hms since he
objec s in he poin cloud also change a highe equencies. Fi s app oaches applied
hand-c a ed ea u es and sliding windows algo i hms wi h Suppo Vec o Machine (SVM)
classi ie s o objec iden i ica ion, bu soon we e eplaced by o he imp o ed me hods
such as 2D ep esen a ions, olume ic-based, and aw poin -based da a, which deploy
machine lea ning echniques in he pe cep ion sys em o he ehicle [28].
2.2.2. SLAM
SLAM, a well-es ablished esea ch opic in he ield o obo ics ha s udies solu ions o
build eal- ime localiza ion maps based solely on pe cep ion da a, has also been p oposed o
be applied in au onomous ehicle applica ions, e en despi e he conside able amoun o 3D
in o ma ion he LiDAR senso gene a es. Usually, odome y uses da a om se e al senso s,
e.g., IMUs and came as, o es ima e he ehicle’s posi ion ela i e o a s a ing loca ion.
Howe e , because LiDAR senso s can gene a e high da a a es, some app oaches ollow
he end o LOAM [
60
], which p ocesses he odome y ask a a highe equency, and

Senso s 2023,23, 601 6 o 30
he mapping a a much lowe a e o keep he eal- ime equi emen s. Despi e p o iding
good esul s, hey end o su e om p oblems associa ed wi h accumula ed d i . To
minimize his issue, some me hods apply an ex a e-localiza ion s ep, ei he using o line
map in o ma ion o imp o e he posi ion es ima ion, o by combining da a om o he
senso s. Some lea ning-based app oaches a e also eme ging, which can use a pipeline
o Con olu ional Neu onal Ne wo ks (CNNs) o manage local ea u e desc ip o s, in e
he localiza ion o se , and apply empo al smoo hness due o he sequen ial na u e o he
localiza ion ask. These equi e, howe e , mo e compu a ional equi emen s.
2.2.3. D i able A ea De ec ion
In au onomous ehicle na iga ion, de ec ing he d i able a ea is one o he mos
c i ical asks. Fo he ca o sa ely mo e in he en i onmen , i is necessa y he de ec ion
o no only obs acles such as pedes ians o o he ehicles bu also se e al oad ea u es
such as oad bounda ies, i.e., cu bs, asphal be ms, walls and o he geome ic ea u es,
c osswalks, sidewalks, a ic signs, e c. [
61
,
62
]. Since LiDAR senso s can p o ide in ensi y
in o ma ion in he poin cloud, i becomes easie o iden i y high- e lec ion oad elemen s
such as a ic signs and oad ma king pain . None heless, o educe da a ans e , i is
also necessa y o dis inguish backg ound om o eg ound da a, including g ound. By
classi ying poin cloud da a in o g ound and non-g ound poin s, he d i able a ea can
be de ec ed mo e e icien ly since he e a e less poin s o p ocess, which imp o es he
na iga ion ea u es o he au onomous ca while keeping he eal- ime equi emen s.
3. G ound Segmen a ion Me hods
G ound segmen a ion me hods ha e been ho oughly s udied in he li e a u e since
hey play a c ucial ole in di e en au onomous d i ing asks. Wi h he e olu ion o LiDAR
and i s endless applica ions, se e al algo i hms and app oaches ha e eme ged h oughou
he yea s. Figu e 3depic s he adi ional da a low o a classical 3D objec de ec ion s ack,
which is ypically simple , as e , and equi es ewe dependencies [63].
UNDERSTANDING
LiDAR 1
LiDAR 1
Fusion Objec
De ec ion
Shape
Ex ac ion
LiDAR 1
LiDAR 1
LiDAR 1
LiDAR N D i e N T ans o m
Fil e
G ound
Fil e ing
D i e 1 T ans o m
Fil e
LiDAR 1
LiDAR N
D i e 1
D i e N
DATA FLOW
SENSING
Figu e 3. LiDAR p ocessing s ack.
This s ack includes he ollowing s eps in he sensing and unde s anding ope a ions:
(1) he so wa e d i e s ansla e aw da a e ie ed om he senso in o s uc u ed 3D
poin cloud da a; (2) nex , da a is s a is ically ans o med/ il e ed o emo e possible
noise o undesi ed poin s om he poin cloud; (3) da a usion can be used o c ea e a
unique ep esen a ion o he ehicle’s su oundings; (4) a g ound il e ing s ep is applied
o isola e g ound poin s om non-g ound da a; (5) he objec de ec ion ask iden i ies
objec s in he poin cloud da a; and, inally (6) he e is a shape ex ac ion s ep o enable and
pe o m he objec classi ica ion. Rega ding he g ound segmen a ion ask, i s basic concep
can be ound in Figu e 4, whe e he aw da a collec ed om he LiDAR is p ocessed o
Senso s 2023,23, 601 7 o 30
sepa a e he poin s ep esen ing he ehicle’s su ounding en i onmen om he g ound
poin s. Acco ding o he subsequen asks, il e ed da a can be used o he ehicle o sa ely
na iga e he en i onmen , o o he objec de ec ion and classi ica ion s eps. In his pape ,
we s udy mos o he ele an con ibu ions ha could be ound in he s a e-o - he-a o
he g ound il e ing s ep, esul ing in he axonomy illus a ed in Figu e 5.
(a) Raw LiDAR da a. (b) Fil e ed en i onmen and objec s da a. (c) Fil e ed g ound da a.
Figu e 4. Basic p inciple o he g ound segmen a ion ask.
Figu e 5. Classi ica ion and axonomy o exis ing g ound segmen a ion me hods.
Cu en algo i hms can be classi ied in o i e di e en ca ego ies: (1) 2.5D G id-
based algo i hms, which can be u he di ided in Occupancy G ids and Ele a ion Maps
( ha u he include Mul i-le el and Mean-based algo i hms); (2) G ound Modelling,
which can use Gaussian P ocess Reg ession (GPR), Line Ex ac ion, and Plane Fi ing
app oaches;
(3) Adjacen
poin s and Local ea u es, which include Channel-based, Range
Image, Clus e ing, and Region G owing me hods; (4) Highe O de In e ence, which deploy
me hods based on Condi ional Random Field (CRF) and Ma ko Random Field (MRF)
app oaches; (5) and Lea n-Based algo i hms, which mainly apply CNNs o iden i y and
pe o m he g ound segmen a ion asks.
Senso s 2023,23, 601 8 o 30
3.1. 2.5D G id-Based
Cu en ly, mode n LiDAR senso s may hold housands o poin s in hei 3D poin
cloud ep esen a ions, which makes i ha de o analyze he en i e poin cloud in eal ime
du ing he na iga ion asks. To help in mi iga ing hese p oblems, g id-based echniques
use a essella ed 2D ep esen a ion o he 3D space, whe e each cell con ains in o ma ion
abou he poin s inside. The u iliza ion o his echnique can d as ically educe he compu-
a ional and memo y equi emen s associa ed wi h he 3D poin cloud ep esen a ion.
Ele a ion Maps:
This echnique is he mos used in 2.5D g id ep esen a ions. As depic ed
by Figu e 6( op iew ep esen a ion o he su ounding en i onmen ), each cell con ains
ele an in o ma ion abou all poin s inside. Ele a ion maps can p o ide ad an ages in
e ms o noise educ ion when compa ed wi h o he me hods. Howe e , hey s ill ace
p oblems in e ms o e ical space ep esen a ion as hey ail o model he emp y space
be ween poin s, i.e., o e hangs and ee ops. Ne e heless, due o i s simplici y, se e al
algo i hms le e age his echnique o g ound plane segmen a ion. Douilla d e al. [
64
]
use a mean-based ele a ion map, i.e., each cell con ains he a e age heigh o all poin s,
ollowed by clus e ing echniques o de ec and emo e he g ound poin s. The algo i hm
ollows h ee simple s eps: (1) calcula es he su ace g adien s o each cell and classi ies
p o uding objec s and g ound cells; (2) clus e s adjacen g ound cells; and (3) co ec s
a i ac s w ongly gene a ed by he g adien compu a ions, e.g., i a cell ha was classi ied
as an objec has an a e age heigh close o i s neighbo ing g ound cells, i is changed o
a g ound cell. Despi e p esen ing good pe o mance esul s, he p oposed solu ion s ill
needs u he imp o emen s o achie e eal- ime cons ain s.
LiDAR
0
Figu e 6. Visual ep esen a ion o an Ele a ion Map.
The wo ks p oposed by As adi e al. [
65
] and Li e al. [
66
] include, on each cell o he
2.5D ele a ion g id, he a e age heigh and he a iance be ween all poin s. I bo h alues
a e lowe han a con igu ed h eshold, he cell is conside ed la and classi ied as pa o he
g ound plane. The i s me hod was e alua ed wi h he KITTI da ase using poin cloud
da a om a Velodyne HDL-64, and he la e wi h da a om a Velodyne HDL-64E senso on
an In el i7 dual-co e p ocesso unning a 3.4 GHz wi h 8 GB o RAM. Despi e his me hod
allowing o e y accu a e de ec ion o plana objec s, i can w ongly classi y objec s close
o he g ound as belonging o he g ound plane. To add ess his issue, Meng e al. [
67
]
apply a heigh di e ence ke nel o e each cell and i s neighbo s, allowing o de ec ing
cells wi h a sligh ly highe a e age heigh han neighbo ing cells. Despi e imp o ing he
de ec ion o low-heigh objec s, ex a s eps a e necessa y o add ess une en e ains, which
signi ican ly inc eases he o e all algo i hm’s complexi y. The e a e se e al applica ions o
hese simple echniques, e.g., Tanaka e al. e alua e e ain’s a e sabili y by calcula ing i s
Senso s 2023,23, 601 9 o 30
oughness and slope using an ele a ion map c ea ed om a HOKUYO UTM-30LX LiDAR
moun ed on a Pionee 3DX mobile obo [68].
Using ele a ion maps p esen s some disad an ages when ying o model ee e ical
space be ween poin s, i.e., in o e hanging objec s, limi ing i s u iliza ion in some ou doo
scena ios. Fo ins ance, using mean-based algo i hms in he scena io depic ed in Figu e 7a
causes he sys em o classi y he a ea below he ee as e ain unable o be d i en. To
mi iga e his, P a e al. [
69
] p oposed an algo i hm ha allows mobile obo s o model
e ical and o e hanging objec s in ele a ion maps by classi ying he su oundings in o
ou dis inc classes: (1) egions sensed om abo e; (2) e ical s uc u es; (3) e ical gaps;
and (4) a e sable cells. The g id map esul ing om his me hod is depic ed in Figu e 7b.
The algo i hm s a s by calcula ing he heigh a iance o each cell, applying a Euclidean
clus e ing echnique nex . The clus e ing o poin s whose dis ance wi hin he cell is smalle
han 10 cm allows o he de ec ion o emp y spaces, which is essen ial o iden i ying
o e hangs. Nex , when an o e hang is de ec ed, only he lowes a e age heigh clus e is
conside ed o calcula e he a e age heigh inside he cell. Despi e co ec ly iden i ying d i -
able a eas below o e hangs, his echnique does no allow o he simul aneous modeling
o mul iple su aces, i.e., modeling he e ain abo e an unde pass. To sol e his issue, he
wo k p oposed by T iebel e al. [70] addi ionally s o es he in e als be ween he de ec ed
clus e s o su aces (p e iously, only he lowes was conside ed), allowing o he co ec
ep esen a ion o all he o e hanging su aces. Bo h wo ks we e e alua ed on a Pionee II
AT obo equipped wi h a SICK LMS291 senso moun ed on an AMTEC w is PW70.
Missclassi ied cells
(a) Mean-based ele a ion map (side- iew).
Missclassi ied cells
T a e sable cells
(b) Mul i-le el ele a ion map (side iew).
Figu e 7. Ele a ion map echniques wi h o e hangs.
Occupancy G id Maps:
These algo i hms, in oduced in he 1980s by Mo a ec and El es,
use ine-g ained g ids o model he occupied and ee space in he su ounding en i on-
men [
71
]. The main goal o his echnique is o gene a e a consis en me ic map om
noisy o incomple e senso da a. This can be achie ed by measu ing a cell mul iple imes,
so ha in o ma ion can be in eg a ed using Bayes il e s [
72
]. This me hod is conside ed
highly obus and easy o implemen , which is essen ial o au onomous d i ing asks [
73
].
S anley, he obo c ea ed by he S an o d Uni e si y ha won he De ense Ad anced
Resea ch P ojec s Agency (DARPA) challenge in 2005, used occupancy g ids o obs acle
de ec ion [
74
]. Fi s , he su oundings a e modeled by assuming each cell is in an occupied,
ee, o unknown s a e. A cell is conside ed occupied i he e ical dis ance be ween he
maximum and minimum heigh o he de ec ed poin s exceeds a dis ance del a. I his e i-
ica ion ails, he a ea is conside ed d i able, meaning he g ound is success ully de ec ed.
In he ollowing edi ions o he DARPA Challenge, se e al eams also used occupancy
g ids o modeling he d i able a ea, which helped hem in inishing he challenge [
75
–
77
].
Simila ly, Himmelsbach e al. [
78
] p oposed he u iliza ion o occupancy g ids allied wi h
ehicle posi ion de ec ion mechanisms o accu a ely segmen he g ound plane o he
Senso s 2023,23, 601 16 o 30
(a) Poin cloud ame.
(b) Range image.
Figu e 14. Con e sion o a poin cloud ame o a ange image ep esen a ion.
3.4. Highe O de In e ence
One o he bigges challenges associa ed wi h analyzing LiDAR da a is ha as he
dis ance o he senso inc eases, he poin s wi hin he poin cloud become spa se a
longe anges. This spa si y can lead o alse classi ica ions in poin cloud segmen a ion
me hods, ep esen ing a signi ican d awback in au omo i e applica ions. Highe o de
in e ence me hods ha e p o en use ul in mul iple compu e ision asks, e.g., seman ic
segmen a ion [
114
], which has led o hei ecen implemen a ion in LiDAR sys ems o
o e come p oblems ela ed o g ound segmen a ion o spa se poin cloud scena ios. This
g oup o me hods includes me hods based on he MRF and CRF algo i hms.
Ma ko Random Field (MRF):
An MRF model can be seen as an undi ec ed g aph, whe e
nodes ep esen andom a iables and edges ep esen desi ed local in luences be ween
pai s o nodes. In he case o LiDAR da a, he poin s a e he nodes o he g aph and
he edges can be modeled using he heigh alues. Guo e al. [
115
] and Byun e al. [
116
]
sugges he combina ion o an MRF wi h a Belie P opaga ion (BP) algo i hm o iden i y
he d i able a ea, e en a high dis ances. None heless, hese app oaches s ill ace de ec ion
p oblems when d i ing in ough e ains. The wo k om Zhang e al. [
117
] ies o
imp o e hese app oaches o be used bo h in ough and une en e ains. I implemen s
cos unc ions o gene a e p obabilis ic g ound heigh measu emen s, c ea ing models ha
compensa e he loss o in o ma ion due o pa ial obs uc ion om close objec s. Based on
his in o ma ion, combined wi h a BP algo i hm, a mul i-label Ma ko ne wo k is used o
he g ound segmen a ion ask. The p oposed me hod shows p omising esul s in spa se
poin cloud dis ibu ions, achie ing alse posi i e a es as low as 2.12% in complex o - oad
en i onmen s. Howe e , he a e age p ocessing ime in he men ioned scena io was abo e
1 s using an In el Co e p ocesso unning a 3.2 GHz, which makes i s u iliza ion ha d
in embedded pe cep ion sys ems wi h eal- ime equi emen s. O he wo ks use simila
app oaches combined wi h heigh his og ams o es ima e he g ound heigh ange, ollowed
by he MRF model o e ine he labeling o g ound poin s [118].

Senso s 2023,23, 601 17 o 30
Huang e al. [
119
] p opose an algo i hm ha aims a sol ing he high compu a ional
equi emen s o MRF-based me hods. This algo i hm s a s by pe o ming a coa se seg-
men a ion based on a ing-based ele a ion map whe e da a poin s a e a anged in ings
whose diame e s a e p opo ional o he dis ance o he LiDAR senso . Since his ype o al-
go i hm assumes ha he g ound in he g id is la , which is no he case in mos eal-wo ld
scena ios, an op imiza ion algo i hm based on spa io empo al adjacen poin s is applied.
The ea e , o pe o m a mo e e ined segmen a ion, he poin cloud is con e ed in o a
ange image whe e each poin is p ojec ed as a g aph node o c ea e an MRF model, unlike
he p e ious app oaches ha con e g ids in o g aph nodes [
94
,
114
]. Ne e heless, MRF
me hods based on i e a i e implemen a ions, such as he BP algo i hm, a e usually compu-
a ionally expensi e and ime-consuming o pe o m he segmen a ion ask. To mi iga e
his, his app oach ini ializes he ne wo k wi h in o ma ion ega ding he high-con idence
obs acle poin s and g ound poin s cap u ed om he segmen a ion algo i hm, which helps
in educing he compu a ional complexi y and a oids con e gence p oblems ha usually
exis in o he implemen a ions. Finally, he algo i hm applies he g aph cu segmen a ion
echnique, depic ed in Figu e 15, o sol e he model, which helps in achie ing he desi ed
ine segmen a ion esul s. Due o he educ ion in he algo i hm’s complexi y, he p oposed
wo k can achie e an a e age p ocessing ime o 39.77 ms wi h a single co e om an In el
i7-3770 p ocesso while using he KITTI da ase con aining da a om a Velodyne HDL-64E.
Cu
O iginal Segmen ed
Figu e 15. G aph Cu segmen a ion me hod.
Condi ional Random Field (CRF):
CRF is a subse o MRF ha labels sequences o nodes
gi en a speci ic chain o obse a ions, which imp o es he abili y o cap u ing long- ange
dependencies be ween hem. Rummelha d e al. [
120
] p opose he addi ion o spa ial and
empo al dependencies o CRF o model he g ound su ace. Thei me hod di ides he
en i onmen in o di e en in e connec ed ele a ion cells, which is in luenced by local
obse a ions and spa io- empo al ela ionships. The empo al cons ain s a e inco po a ed
in o he segmen a ion da a using a dynamic Bayesian amewo k, allowing o a mo e
accu a e modeling o g ound poin s. The p oposed me hod was i s es ed on an In el Xeon
W3520 p ocesso unning a 2.6 GHz wi h 8 GB o RAM and wi h a Quad o 2000 g aphics
ca d wi h 2 GB o ideo memo y, achie ing a ame p ocessing equency o 6.8 Hz wi h
da a om a Velodyne HDL-64E senso . Howe e , he au ho s claim ha wi h expe imen al
pla o ms such as Teg a X1 and K1, he algo i hm achie ed eal- ime pe o mance. Simila ly,
Senso s 2023,23, 601 18 o 30
and ollowing he app oach o using CRF me hods in he segmen a ion o da a om digi al
came as [
121
], Wang e al. [
122
] model g ound da a by ep esen ing he CRF in a 2D la ice
plane. Nex , i uses he RANSAC on he c ea ed plane o ex ac he g ound poin s. Despi e
p esen ing good segmen a ion esul s, i equi es se e al i e a ions o co ec ly ex ac he
poin s in une en e ains, which hea ily comp omises eal- ime pe o mance. Addi ionally,
hese me hods a e e y compu a ionally-hung y, which makes hem unsui able o eal-
ime scena ios unless specialized ha dwa e accele a ion is used.
3.5. Lea n-Based
Lea n-based me hods ha e been widely applied o came a ision sys ems, achie ing
good esul s in a a ie y o segmen a ion asks, including g ound segmen a ion. None he-
less, due o he inhe en bene i s o deploying LiDAR senso s in he pe cep ion sys em o
he ca , se e al lea n-based me hods applied o poin cloud da a began o eme ge [
123
].
Back in 1988, Pome leau [
124
] in oduced Au onomous Land Vehicle In a Neu al Ne wo k
(ALVINN), a 3-laye back-p opaga ion ne wo k designed o he ask o oad ollowing
ha was ained om mul iple simula ed oad images. Despi e he solu ion no allowing
he ehicle o ollow he oad a high speeds, p ima ily due o compu a ional limi a ions
a he ime, i was good enough o alida e his new app oach and o demons a e ha
lea ning-based me hods could become po en ial solu ions o au onomous d i ing asks.
The success o ALVINN has igge ed he u iliza ion o machine lea ning algo i hms ap-
plied o LiDAR da a, which, based on he success o hese app oaches, se e al algo i hms
ha e eme ged and many mo e a e cu en ly being eleased. Thus, i makes i e y ha d
o compile all he exis ing in o ma ion in his sec ion. None heless, we p o ide he mos
signi ican algo i hms ha made g ea con ibu ions o his opic.
Poin Ne [
125
] is a uni ied a chi ec u e o di e en applica ions, e.g., objec classi i-
ca ion and segmen a ion, scene seman ic pa sing, e c. Despi e showing p omising esul s
esis an o inpu pe u ba ion and co up ion, by design, his a chi ec u e is only sui -
able o small-scale poin clouds. This is mainly due o he weak lea ning capabili ies o
local ea u es, a ec ing i s pe o mance in la ge-scale complex poin clouds and limi ing
i s capaci y o iden i y ine-g ained pa e ns. Aiming a imp o ing he p e ious wo k,
Poin Ne ++ [
126
] implemen s a hie a chical amewo k o disco e he mul i-scale spa ial
con ex ual elemen s, which imp o es pe o mance in la ge-scale poin clouds. None heless,
he hie a chical s uc u e inc eases he o e all compu ing complexi y. O he eme ging
app oaches, such as he Poin wise CNN [127], ocus on poin wise con olu ion ope a ions
applied o each poin in a poin cloud, which achie es good accu acy esul s in he seman ic
segmen a ion and objec ecogni ion asks. SPG aph [
128
] o ganizes he poin cloud using
a da a s uc u e known as he Supe Poin G aph (SPG), which is c ea ed om pa i ioning
he scanned image in o geome ically homogeneous elemen s. The e alua ion was con-
duc ed on a 4 GHz CPU and GTX 1080 Ti GPU o e wo di e en da ase s, Seman ic3D
and he d S an o d La ge-Scale 3D Indoo Spaces (S3DIS). Py amidPoin [
129
] employs
a dense py amid s uc u e ha p o ides a second glance a he poin cloud, allowing
he ne wo k o e isi dis inc laye s o med om o he ne wo k le els, which enables
ea u e p opaga ion. Howe e , hese me hods equi e expensi e nea es -neighbo sea ch
algo i hms in spa se space and a e ine icien when dealing wi h ou doo and au omo i e
poin clouds. None heless, based on hei success, hese p inciples ha e been used o help
sol e he g ound segmen a ion p oblem. The expe imen s we e pe o med wi h h ee
di e en ypes o da ase s, i.e., ae ial, e es ial, and mobile lase scanne s, on a single
NVIDIA TITAN RT GPU.
Paigwa e al. [
130
] de eloped he GndNe , whose s eps a e shown in Figu e 16.
The 3D poin cloud g ound plane ele a ion is es ima ed in a g id-based ep esen a ion,
and he poin s a e di ided in o g ound and non-g ound ca ego ies. The me hod s a s
by ans o ming he poin cloud in o an e enly spaced 2D g id, p oducing a se ies o
poin -popula ed pilla s. Then, o ex ac he ea u es con ained in each non-emp y pilla , a
simpli ied e sion o he Poin Ne me hod is used [
125
,
126
], which gene a es a pseudo-
Senso s 2023,23, 601 19 o 30
image o he g id. Then, a 2D con olu ional encode -decode ne wo k p ocesses his
image o p oduce a high-le el ep esen a ion. Howe e , o aining he ne wo k, i is
necessa y o use a labeled g ound ele a ion map da ase . To sol e his, he au ho s used
he Seman icKITTI [
81
,
131
] and modeled he g ound plane wi h a CRF me hod o gene a e
a g ound- u h ele a ion map sui able o he equi ed aining. This allowed GndNe o
achie e a mean in e sec ion o e union (mIoU) alue o 83.6%, while p o iding eal- ime
pe o mance wi h an a e age p ocessing ime o 17.98 ms pe ame.
Raw 3D
Poin Cloud Segmen a ion
Resul
GndNe Me hod
Poin Cloud
Disc e iza ion Fea u e Ex ac ion
Con olu ional
Encode /Decode
Ne wo k
P ocessing
Th eshold Based
Segmen a ion
2D Pilla G id 2D Pseudo Image G ound Ele a ion Es ima ion
Figu e 16. O e iew o he GndNe algo i hm.
Despi e he ma u i y o poin -wise app oaches, new app oaches s a by ep esen ing
poin clouds as s uc u es be o e applying he lea ning-based algo i hms, which helps
in achie ing mo e e iciency while equi ing less compu a ional esou ces [
132
]. Fo in-
s ance, VoxelNe [
133
], disc e izes poin clouds in o oxels be o e using a 3D CNN. The
p oposed Voxel Fea u e Encoding (VFE) laye in eg a es ea u e ex ac ion and bounding
box p edic ion in o a single s age, allowing he algo i hm o achie e an in e ence ime o
a ound 33 ms and a p ecision o up o 89.6% on a p ocessing sys em composed o a Ti anX
GPU wi h a CPU unning a 1.7 GHz. On he o he hand, SPLATNe [
134
] in e pola es
poin clouds in o pe mu ohed al spa se la ices be o e execu ing a 3D CNN, enabling a
s aigh o wa d mapping o 2D in o ma ion in o 3D and ice- e sa. Cylinde 3D [
135
]
builds a cylind ical pa i ioning and hen uses an asymme ical 3D con olu ion ne wo k
ha enables i o each an mIoU o 61.8% using he Seman icKITTI da ase , which p o ides
da a om a Velodyne HDL-64E LiDAR senso . Poin Pilla [
136
] is a no el encode me hod
ha uses Poin Ne [
125
,
126
] o c ea e pilla ep esen a ions om he inpu poin clouds.
This ep esen a ion can hen be used by a 2D con olu ional a chi ec u e o p ocess i in o
highe -le el ep esen a ions o he iden i ica ion s eps. Addi ionally, his me hod can
achie e eal- ime pe o mance, wi h in e ence aking 16.2 ms pe ame on a compu ing
sys em composed o an In el i7 p ocesso and a 1080 i GPU.
SqueezeSeg [
137
–
139
] is ano he wo k ha uses hese pa icula s uc u es, whe e
a 2D sphe ical p ojec ion poin -wise label map is used and subsequen ly e ined by a
CRF implemen ed as a ecu en laye . I s mos ecen e sion, SqueezeSegV3, suppo s
wo di e en implemen a ions ha can achie e di e en pe o mance esul s. SSGV3-53
eaches an mIoU o 52.2% wi h 11 ames pe second and SSGV3-21 achie es an mIoU
o 48.8% wi h 16 ames pe second. RangeNe ++ [
140
] is a deep-lea ning me hod ha
uses ange images and 2D con olu ions, ollowed by GPU-accele a ed pos -p ocessing
me hods, o eco e consis en seman ic in o ma ion o comple e LiDAR scans. Simila ly
o SqueezeSeg, RangeNe p esen s wo implemen a ion e sions, whe e RangeNe 21 can
achie e an mIoU o 47.4% wi h 20 ames pe second and RangeNe 53 an mIoU o 49.9%
wi h 13 ames pe second. All es s we e conduc ed on GPU-based se ups wi h he KITTI
da ase . Finally, Pola Ne [
141
] uses a pola Bi d’s-Eye-View (BEV) ep esen a ion, which
balances he poin s ac oss g id cells in a pola coo dina e sys em, allowing i o achie e an
mIoU o 54.3% wi h 16.2 ames pe second on he Seman icKITTI da ase .
Se e al wo ks use CNNs o iden i y and segmen obs acles wi hin a LiDAR poin cloud.
Howe e , pe o ming lea n-based algo i hms on LiDAR da a equi es a signi ican amoun
o p ocessing powe which o en ansla es in o e y ime-consuming asks. Consequen ly,
Senso s 2023,23, 601 20 o 30
Lyu e al. [
142
] p oposed an e icien FPGA-based solu ion ha can p ocess each LiDAR scan
wi hin 16.9 milliseconds using he KITTI da ase [
81
,
131
] as a es bed. Velas e al. [
143
] used
CNNs o pe o m g ound segmen a ion, conside ably educing he algo i hm’s p ocessing
ime when compa ed wi h p e ious lea n-based app oaches, bu ailing o achie e eal- ime
pe o mance in a CPU implemen a ion (In el i5-6500 p ocesso ). The bes pe o mance
esul s we e achie ed wi h GPU accele a ion (NVIDIA GeFo ce GTX 770), whe e di e en
ne wo k opologies we e es ed. Wi h a opology consis ing o 5 con olu ional laye s plus
a single decon olu ion, he achie ed p ocessing ime o one Velodyne HDL-64E ame is,
on a e age, less han 7 ms. Ne e heless, his me hod can only be applied o g ound poin
segmen a ion a he han o he segmen a ion asks. Zhang e al. [
144
] in oduced ShellNe ,
a me hod o applying deep lea ning o 3D poin clouds. I is buil on a con olu ion
ope a o ha ep esen s poin se s wi h locally c ea ed sphe ical shells. This me hod can
esul in as poin cloud segmen a ion and ou pe o m o he me hods in e ms o accu acy.
Howe e , he lack o labeled aining da a s ill poses a signi ican obs acle o he algo i hm’s
pe o mance in many scena ios.
Shen e al. [
145
] p opose a me hod based on he jump-con olu ion-p ocess (JCP) o
sol e he p oblems aced by segmen a ion algo i hms in handling complica ed e ains,
and he excessi e p ocessing ime and memo y equi emen s. The me hod s a s by
p ojec ing on o an RGB image he poin cloud p e iously labeled by an imp o ed local
ea u e ex ac ion algo i hm. The pixel alue is hen ini ialized wi h he poin ’s label
and con inuously upda ed using image con olu ion. Finally, i is used he con olu ion
p ocess wi h a jump ope a ion o pe o m ope a ions only on he low-con idence poin s
il e ed by he c edibili y p opaga ion mechanism, which educes he execu ion ime. This
me hod shows good pe o mance esul s in di e en en i onmen s, achie ing an a e age
p ocessing ime o 8.61 ms and 15.62 ms when dealing wi h 64-beam and 128-beam LiDAR
da a, espec i ely.
He e al. [
132
] p opose Sec o GSne , an end- o-end DNN amewo k designed o
pe o m g ound segmen a ion o ou doo LiDAR da a. The algo i hm s a s wi h a sec o
encode module in which he 3D poin cloud is di ided in o di e en sec ions based on a
BEV sec o pa i ioning me hod, as depic ed in Figu e 17. The poin cloud is ep esen ed in
a ci cula egion, ollowed by i s di ision in o equal slices (W) and adial line segmen s
(H) ha a e dynamically adjus ed along he adius so ha he poin s become mo e e enly
dis ibu ed ac oss he g id.
W
W
H
z
y
x
Spa se
Figu e 17. BEV sec o pa i ion ep esen a ion.
A e pa i ioning he poin cloud, some ea u es, e.g., poin coo dina es and in ensi y,
a e agg ega ed in each sec o using a simpli ied e sion o he Poin Ne [
125
,
126
] encode ,
esul ing in he gene a ion o a sec o ea u e map. Then, he sec o encode module
Senso s 2023,23, 601 21 o 30
ecei es he sec o ea u e map o p ocess i using he UNe CNN o p edic he sec o labels.
In he inal s age, a sec o - o-poin es o a ion is pe o med o econs uc he segmen ed
poin cloud. All models we e ained on a desk op compu e wi h an NVIDIA GTX2080Ti
g aphics ca d and an In el i7 6700k p ocesso , esul ing in in e ence esul s a 170.6 Hz.
4. Discussion
When choosing a g ound segmen a ion algo i hm, i is c ucial o unde s and hei
mos signi ican di e ences and ea u es ha bes sui he equi emen s o he inal appli-
ca ion. Table 1summa izes a quali a i e compa ison o cu en s a e-o - he-a g ound
segmen a ion me hods ega ding he ollowing me ics: eal- ime ea u es, compu a ional
equi emen s (which a e some imes dic a ed by he algo i hm’s complexi y o s eps o
achie e eal- ime capabili ies), he segmen a ion immuni y, and he abili y o deal wi h
ising obs acles and egions, une en g ound, and spa se da a in he poin cloud. This com-
pa ison is made solely based on a ailable in o ma ion e ie ed om espec i e li e a u e,
whose expe imen al se ups a e o he esponsibili y o espec i e au ho s.
Table 1. Quali a i e compa ison be ween exis ing g ound segmen a ion me hods.
Me hod
Me ic
Real-Time Compu a ional
Requi emen s Segmen a ion Immuni y
Pe o mance wi h
Rising Obs acles
and Regions
Pe o mance
wi h Une en
G ound
Pe o mance
wi h Spa se Da a
2.5D G id-based
Mean-Based [64–68]3[66,68]; 7[64] Low Unde -/o e -segmen a ion Good [67,68] Good [67,68] -
Mul i-le el [69,70]- Medium - Good Good -
Occupancy G id [74–79]3[74–78]; 7[79] High Unde -segmen a ion [74] Good [79] - -
G ound Modelling
GPR-based [90–93]3[91]; 7[90,92,93]High Unde -segmen a ion [91]Insensi i e o slowly ising
obs acles [91]Good [91,93] Good [90,92,93]
Line Ex ac ion [87,88]3[87] Medium Sligh
O e -/unde -segmen a ion
Good [88] Good [88]; Bad [87] -
Plane Fi ing [80,82–86]3[82–86]; 7[80] Medium/High P one o
o e -segmen a ion [82];
Unde -segmen a ion [83]
Good [83,86]Good [83,86];
Bad [80]Good [82]
Adjacen Poin s and
Local Fea u es
Channel-based [94–100]3[95–98,100];
7[94]Medium
Unde /o e -segmen a ion [
94
];
Unde /o e -segmen a ion [97]Good [94–98] Good [94,97] Good [94]
Range Images [109–112]3[109,110,112];
7[111]Medium P one o
o e -/unde -segmen a ion;
O e -segmen a ion [109,110]
Good [109,110] - Good [109,110]
Clus e ing [
90
,
105
,
106
]
7[90] Medium/High Unde -/o e -
segmen a ion [105]
Good [106] - Good [90]
Region G owing [101–104]7[101] Medium/High Small o e -segmen a ion
[101,102,104]; Small unde -
/o e -segmen a ion [103]
Good [104] - -
Highe O de
In e ence
MRF [115–119]3[119]; 7[117] High - Good [115,116] Good [115,117];
Bad [116]
Good [115,116]
CRF [120,122]3[120](wi h GPU);
7[122]High - - - Good [120]
Deep Lea ning
CNN [
125
–
130
,
132
–
144
]
3[130,133,136,140]
(wi h GPU), and
FPGA [142]; 7[143]
High/Ve y High - Good Good Good
Real- ime:
Wi hin an au onomous d i ing scena io, i is manda o y o he pe cep ion
sys em o p ocess and unde s and he su ounding en i onmen in eal ime, which means
ha , o a gi en senso o a se o senso s, he s eps o he LiDAR p ocessing s ack mus
be pe o med wi hin a known pe iod o ime so ha he d i ing decisions can be aken
wi hin a sa e ime ame. A high- esolu ion LiDAR senso can gene a e millions o da a
poin s. Fo ins ance, he Velodyne VLS-128 can p oduce up o 9.6 M poin s pe second in
he dual- e u n mode, wi h ame a es a ying om 5 Hz o 10 Hz. In a ypical ope a ion,
his senso can be con igu ed o p oduce, on a e age, a poin cloud o 2,403,840 poin s

Senso s 2023,23, 601 22 o 30
pe second (240,384 poin s a 10 Hz), which means ha such an amoun o da a mus
be p ocessed in unde 100 ms ac oss all so wa e s ack laye s. Rega ding he g ound
segmen a ion asks, wi h a ew excep ions, bo h 2.5D G id-based and G ound Modelling
me hods can achie e eal- ime p ocessing. Fo he algo i hms based on he mul i-le el
app oach [
69
,
70
], his in o ma ion could no be e ie ed. Due o hei complexi y, mos o
he GPR-based me hods [
90
,
92
,
93
] and one Plane- i ing me hod [
80
] a e unable o p o ide
he desi ed eal- ime ea u es.
Conce ning he algo i hms based on adjacen poin s and local ea u e ex ac ion,
i.e., Region
G owing [
101
], Clus e ing [
90
], one om Range Images [
111
], and one om
Channel-based [
94
], hey do no p o ide eal- ime (likely due o lack o op imiza ions o
ha dwa e se up used), while he emaining me hods [
94
–
98
,
100
,
109
,
110
,
112
], a e consid-
e ed eal- ime. Rega ding he Highe O de In e ence [
115
–
120
,
122
] and Deep lea ning
[
125
–
130
,
132
–
144
] me hods, hey na u ally do no p o ide eal- ime g ound segmen a ion
due o he o e all algo i hm’s complexi y. Howe e , when eso ing o ha dwa e accel-
e a ion, e.g., based on GPUs [
120
,
130
,
133
,
136
,
140
] o FPGA [
142
], o aking ad an age o
coa se segmen a ion esul s om local ea u e ex ac ion [
29
], some solu ions can mee
eal- ime equi emen s.
Compu a ional Requi emen s:
When deployed in au omo i e applica ions, he compu a-
ional equi emen s associa ed wi h g ound segmen a ion me hods a e a c ucial me ic
o conside , mainly because he pe cep ion sys em is o en composed o embedded p o-
cessing en i onmen s ha y o minimize he a ailable ha dwa e esou ces. Me hods
based on Ele a ion Maps p esen he lowes compu a ional equi emen s as hey analyze a
2D essella ed ep esen a ion o he su oundings, ins ead o he en i e 3D ep esen a ion.
Rega ding he o he 2.5D g id-based app oaches, Mul i-le el algo i hms equi e ex a
classi ica ion s eps, while Occupancy G ids equi e he in e pola ion o da a and he use
o Bayes il e s, which inhe en ly inc eases he memo y and compu a ional needs in bo h
cases. Likewise, GPR-based and Plane Fi ing me hods (based on RANSAC) ely on i e a-
i e app oaches, consequen ly inc easing he memo y equi emen s. Addi ionally, GPR
me hods equi e complex calcula ions o p ocess he poin cloud, which u he inc eases
he compu a ional needs. On he o he hand, Line Ex ac ion app oaches ea u e a lowe
esou ce consump ion han he emaining G ound Modelling me hods since hey di ide
he poin cloud in o a pola g id map, simpli ying he equi ed compu a ions. Conce n-
ing he Adjacen Poin s and Local Fea u es me hods, Channel-based and Range Image
app oaches a e no conside ed e y compu a ionally in ensi e. Howe e , hey a e based
on he analysis o geome ic condi ions be ween poin s, which can ep esen he need o
specialized ha dwa e, such as a loa ing-poin uni o igonome ic calcula ions. On he
o he hand, despi e hei simplici y, Clus e ing and Region-G owing me hods equi e e y
i e a i e ope a ions, which can ep esen high memo y equi emen s especially o la ge
poin clouds wi h poin s con aining mul iple ea u es, e.g., a ge ’s e lec i i y. On he o he
side, Highe O de In e ence me hods ea u e high compu a ional equi emen s due o he
associa ed complex ma hema ical compu a ions and espec i e i e a i e s eps, e.g., he BP
algo i hm, which ansla es in o high memo y equi emen s. Finally, among all me hods,
Lea n-based app oaches demand he highes compu a ional needs. This is mainly caused
by he ex ensi e complex compu a ions, signi ican memo y u iliza ion, and some imes
specialized ha dwa e, e.g., GPUs, gene ally associa ed wi h CNN implemen a ions.
Segmen a ion Immuni y:
Segmen a ion immuni y e e s o he algo i hm’s suscep ibili y
o unde - and/o o e -segmen a ion, which co esponds o ei he oo coa se o oo ine
segmen a ion, espec i ely. When he unde -segmen a ion occu s, he poin s belonging o
di e en objec s a e me ged in o he same g ound segmen . On he o he hand, wi h o e -
segmen a ion, a single objec can be ep esen ed by se e al clus e s. In many applica ions,
unde -segmen a ion is conside ed a mo e se e e issue han o e -segmen a ion, since he
w ong classi ica ion o he g ound plane can lead o sa e y issues o he au onomous
ehicle. 2.5D G id-based Ele a ion Map algo i hms end o su e bo h om unde - o o e -
Senso s 2023,23, 601 23 o 30
segmen a ion, which can highly a ec he o e all algo i hm’s pe o mance and accu acy.
Tha happens especially when he g ound is signi ican ly sloped o cu bed. G ound
Modelling’s Line Ex ac ion and Plane Fi ing [
82
] me hods end o sligh ly su e om
unde - o o e -segmen a ion. Howe e , GPR-based [
91
] and Plane Fi ing [
83
] a e able
o o e come unde -segmen a ion issues. Adjacen Poin s and Local Fea u es algo i hms
ypically a e suscep ible o unde - o o e -segmen a ion, excep o [
103
], which is immune
o o e -segmen a ion, while [97,105] a e immune o bo h unde - and o e -segmen a ion.
Pe o mance wi h Rising Obs acles and Regions, Une en G ound, and Spa se Da a:
In
a eal-wo ld d i ing scena io, he g ound is no la and, ising obs acles, egions, and
une en g ound usually ep esen a signi ican challenge o g ound segmen a ion algo i hms.
Addi ionally, dealing wi h spa se poin clouds can lead o pe o mance loss o algo i hm
inabili y o sol e he segmen a ion ask. The e o e, o assess he e sa ili y and sa e y o
a me hod in di e en si ua ions, i is c ucial o e alua e he segmen a ion pe o mance
wi h ising obs acles, slopped o ough e ains, and spa se da a. Rega ding 2.5 G id-
based me hods, since he g ound is modeled in o a g id whe e each cell ep esen s a small
egion o he g ound plane, almos all o hem can pe o m well wi h ising obs acles and
une en g ound su aces [
67
–
70
]. Howe e , hey can be unp edic able when dealing wi h
spa se da a. The G ound Modeling app oaches based on GPR algo i hms can su e om
insensi i i y o hese slowly ising obs acles [
91
].Since he g ound di ision in independen
angula sec o s does no gua an ee he gene al g ound ele a ion con inui y, some obs acles,
such as s ai s eps, may be classi ied as g ound. None heless, he o he app oaches, i.e., Line
Ex ac ion [
88
], and Plane Fi ing [
83
,
86
], can handle hese objec s p ope ly. Conce ning
he une en g ound egions, some G ound Modelling app oaches canno handle hem
p ope ly [
80
,
87
], e.g., he Plane Fi ing me hod [
80
] uses RANSAC in o de o es ima e
he g ound plane. The e o e, he assump ion o a single g ound plane leads o alse
classi ica ions and complex da a can deg ade he RANSAC pe o mance. None heless,
mos GPR-based algo i hms [
90
,
92
,
93
], and one Plane Fi ing me hod [
82
], can achie e
good segmen a ion esul s when handling spa se da a.
F om he Adjacen Poin s and Local Fea u es me hods, mos o hem can handle ising
obs acles and egions, e.g., Channel-based [
94
–
98
], Range Images [
109
,
110
],
Clus e ing [106]
,
and egion G owing [
104
]. Rega ding handling une en g ound, some Channel-based
app oaches [
94
,
97
] can pe o m well in sloped en i onmen s, whe e [
94
] de eloped algo-
i hms specialized in handling sloped e ains and spa se da a, and [
97
] p o ed o achie e
good esul s wi h sloped and la en i onmen s. Some me hods [
90
,
94
,
109
,
110
] can also
pe o m well in spa se da a, highligh ing he wo ks o [
94
,
109
,
110
] ha de eloped me h-
ods especially o spa se poin clouds. Some Highe O de In e ence MRF me hods can
pe o m well in ising obs acles and egions [
115
,
116
], as he e a e also o he s ha manage
o ob ain good esul s on bo h sloped and ough e ains [
115
,
117
]. Un o una ely, no o he
in o ma ion could be ound on he o he me ics o he CRF algo i hms. While handling
wi h spa se poin clouds, some MRF me hods [
115
,
117
] and one CRF me hod [
120
] can
wo k adequa ely wi h spa se poin clouds. Deep Lea ning me hods gene ally ob ain he
mos accu a e esul s in e ms o g ound segmen a ion on well- ained en i onmen s, be-
ing able o pe o m well in sloped and ough e ain, as well as in dealing wi h spa se
da a. Howe e , he lack o labeled da ase s o he aining phases, he ime-consuming
con e sions be ween 3D poin cloud da a and ne wo k inpu da a, he need o powe ul
ha dwa e suppo , and he complexi y in ol ed in de eloping algo i hms, usually hampe s
he de elopmen o CNNs o he segmen a ion g ound asks.
Fu u e ends o g ound segmen a ion me hods:
G ound and objec segmen a ion is an
impo an ask in au onomous d i ing applica ions, whe e he algo i hm’s pe o mance
and he eal- ime aspec s a e he mos c i ical equi emen s when building he pe cep ion
sys em o he ehicle. Au omo i e LiDAR senso s a e also becoming mains eam, hus,
se e al algo i hms and app oaches o p ocess poin cloud da a a e cons an ly eme ging.
Wi h he success o lea n-based app oaches in e ms o accu acy and pe o mance, i is
Senso s 2023,23, 601 24 o 30
expec ed ha u u e solu ions will adop CNNs o pe o m he g ound segmen a ion
asks [
123
]. Wi h he cons an de elopmen o echnology and esea ch a ound hese opics,
cu en ly challenges o CNNs wi h LiDAR da a applied o au omo i e, e.g., he lack o
da ase s [
146
], he high ime-consuming aining phases, he equi emen o powe ul
compu ing sys ems, and he algo i hm’s complexi y, a e slowly being mi iga ed. Despi e
pe cep ion sys ems being including mo e senso s and LiDAR de ices p o iding mo e
esolu ion da a a highe ame a es (which inhe en ly inc eases he amoun o da a o be
p ocessed), i is expec ed ha in he nea u u e lea n-based solu ions wi h lowe ha dwa e
equi emen s will pe o m he g ound segmen a ion s eps wi h educed p ocessing imes.
5. Conclusions
An au onomous ehicle equi es a good pe cep ion sys em o success ully na iga e
he en i onmen , equi ed o objec iden i ica ion and obs acle a oidance, and he g ound
plane and oad limi s de ec ion asks. De ec ing and classi ying he d i able a ea is undoub -
edly a c ucial ask ha equi es ex ac ing in o ma ion om he poin cloud o p ecisely
de ec common oad bounda ies. Howe e , his is only possible wi h as and e icien
g ound segmen a ion me hods ha can deli e eal- ime ea u es, and choosing he bes
ha sui s he inal applica ion will su ely a ec he way he ehicle mo es a ound. This
a icle p esen s a ligh weigh su ey o exis ing me hods o de ec and ex ac g ound
poin s om LiDAR poin clouds. I summa izes he ex ensi e li e a u e and p oposes a
comp ehensi e axonomy o help unde s and he cu en g ound segmen a ion me hods
ha can be used in au omo i e LiDAR senso s. The p oposed ca ego ies a e as ollows:
(1) 2.5D G id-based
me hods, (2) G ound Modeling, (3) Adjacen Poin s and Local Fea u es,
(4) Highe O de In e ence, and (5) Lea n-Based me hods, whe e mo e solu ions a e likely
o exponen ially eme ge in he e y nea u u e. Mo eo e , and o unde s and he main
di e ences be ween hem, his a icle also includes a quali a i e compa ison whe e im-
po an me ics, such as eal- ime, compu a ional equi emen s, segmen a ion immuni y,
algo i hm’s pe o mance in di e en condi ions, a e discussed.
Au ho Con ibu ions:
Concep ualiza ion, T.G. and D.M.; me hodology, T.G. and D.M.; alida-
ion, T.G., D.M., A.C., L.C. and R.R.; in es iga ion, T.G., D.M., A.C. and L.C.; esou ces, T.G.;
w i ing—o iginal
d a p epa a ion, T.G., D.M. and R.R.; w i ing— e iew and edi ing, T.G., D.M.,
A.C. and L.C.; supe ision, T.G.; p ojec adminis a ion, T.G.; unding acquisi ion, T.G.; All au ho s
ha e ead and ag eed o he published e sion o he manusc ip .
Funding:
This wo k has been suppo ed by FCT— Fundação pa a a Ciência e Tecnologia wi hin he
R&D Uni s P ojec Scope UIDB/00319/2020 and G an 2021.06782.BD.
Con lic s o In e es : The au ho s decla e no con lic o in e es .
Re e ences
1. Li man, T. Au onomous Vehicle Implemen a ion P edic ions; Vic o ia T anspo Policy Ins i u e: Vic o ia, BC, Canada, 2021.
2.
Socie y o Au omo i e Enginee s (SAE). Taxonomy and De ini ions o Te ms Rela ed o D i ing Au oma ion Sys ems o On-Road
Mo o Vehicles (Su ace Vehicle Recommended P ac ice: Supe seding J3016 Jun 2018); SAE In e na ional: Wa endale, PA, USA, 2021.
3.
Me cedes-Benz G oup. Fi s In e na ionally Valid Sys em App o al o Condi ionally Au oma ed D i ing. Me cedes
2021
.
A ailable online: h ps://g oup.me cedes-benz.com/inno a ion/p oduc -inno a ion/au onomous-d i ing/sys em-app o al-
o -condi ionally-au oma ed-d i ing.h ml (accessed on 5 Sep embe 2022).
4.
157—Au oma ed Lane Keeping Sys ems (ALKS); Na ions Economic Commission o Eu ope: Gene a, Swi ze land, 2021; pp. 75–137.
5.
Goelles, T.; Schlage , B.; Muckenhube , S. Faul De ec ion, Isola ion, Iden i ica ion and Reco e y (FDIIR) Me hods o Au omo i e
Pe cep ion Senso s Including a De ailed Li e a u e Su ey o Lida . Senso s 2020,20, 3662. [C ossRe ] [PubMed]
6.
U mson, C.; Anhal , J.; Bagnell, D.; Bake , C.; Bi ne , R.; Cla k, M.N.; Dolan, J.; Duggins, D.; Gala ali, T.; Geye , C.; e al.
Au onomous d i ing in u ban en i onmen s: Boss and he U ban Challenge. J. Field Robo . 2008,25, 425–466. [C ossRe ]
7.
Ma i, E.; de Miguel, M.A.; Ga cia, F.; Pe ez, J. A Re iew o Senso Technologies o Pe cep ion in Au oma ed D i ing. IEEE In ell.
T ansp. Sys . Mag. 2019,11, 94–108. [C ossRe ]
8.
Shahian Jah omi, B.; Tulabandhula, T.; Ce in, S. Real-Time Hyb id Mul i-Senso Fusion F amewo k o Pe cep ion in Au onomous
Vehicles. Senso s 2019,19, 4537. [C ossRe ] [PubMed]
Senso s 2023,23, 601 25 o 30
9.
Chase, A.F.; Chase, D.Z.; Weishampel, J.F.; D ake, J.B.; Sh es ha, R.L.; Sla on, K.C.; Awe, J.J.; Ca e , W.E. Ai bo ne LiDAR,
a chaeology, and he ancien Maya landscape a Ca acol, Belize. J. A chaeol. Sci. 2011,38, 387–398. [C ossRe ]
10.
Chase, A.S.Z.; Chase, D.Z.; Chase, A.F., LiDAR o A chaeological Resea ch and he S udy o His o ical Landscapes. In Sensing
he Pas : F om A i ac o His o ical Si e; Masini, N., Soldo ie i, F., Eds.; Sp inge In e na ional Publishing: Cham, Swi ze land,
2017; pp. 89–100. [C ossRe ]
11.
Š ula , B.; Lozi´c, E.; Eiche , S. Ai bo ne LiDAR-De i ed Digi al Ele a ion Model o A chaeology. Remo e Sens.
2021
,13, 1855.
[C ossRe ]
12.
Jones, L.; Hobbs, P. The Applica ion o Te es ial LiDAR o Geohaza d Mapping, Moni o ing and Modelling in he B i ish
Geological Su ey. Remo e Sens. 2021,13, 395. [C ossRe ]
13.
Asne , G.P.; Masca o, J.; Mulle -Landau, H.C.; Vieilleden , G.; Vaud y, R.; Rasamoelina, M.; Hall, J.S.; an B eugel, M. A uni e sal
ai bo ne LiDAR app oach o opical o es ca bon mapping. Oecologia 2012,168, 1147–1160. [C ossRe ]
14.
Li, X.; Liu, C.; Wang, Z.; Xie, X.; Li, D.; Xu, L. Ai bo ne LiDAR: S a e-o - he-a o sys em design, echnology and applica ion.
Meas. Sci. Technol. 2020,32, 032002. [C ossRe ]
15. Liu, X. Ai bo ne LiDAR o DEM gene a ion: Some c i ical issues. P og. Phys. Geog . Ea h En i on. 2008,32, 31–49. [C ossRe ]
16.
Meng, X.; Cu i , N.; Zhao, K. G ound Fil e ing Algo i hms o Ai bo ne LiDAR Da a: A Re iew o C i ical Issues. Remo e Sens.
2010,2, 833–860. [C ossRe ]
17.
Yan, W.Y.; Shake , A.; El-Ashmawy, N. U ban land co e classi ica ion using ai bo ne LiDAR da a: A e iew. Remo e Sens.
En i on. 2015,158, 295–310. [C ossRe ]
18.
Chen, Z.; Gao, B.; De e eux, B. S a e-o - he-A : DTM Gene a ion Using Ai bo ne LIDAR Da a. Senso s
2017
,17, 150. [C ossRe ]
[PubMed]
19.
He z eld, U.C.; McDonald, B.W.; Wallin, B.F.; Neumann, T.A.; Ma kus, T.; B enne , A.; Field, C. Algo i hm o De ec ion o
G ound and Canopy Co e in Mic opulse Pho on-Coun ing Lida Al ime e Da a in P epa a ion o he ICESa -2 Mission. IEEE
T ans. Geosci. Remo e Sens. 2014,52, 2109–2125. [C ossRe ]
20.
Li, Y.; Ibanez-Guzman, J. Lida o Au onomous D i ing: The P inciples, Challenges, and T ends o Au omo i e Lida and
Pe cep ion Sys ems. IEEE Signal P ocess. Mag. 2020,37, 50–61. [C ossRe ]
21.
Ro iz, R.; Cab al, J.; Gomes, T. Au omo i e LiDAR Technology: A Su ey. IEEE T ans. In ell. T ansp. Sys .
2021
,23, 6282–6297.
[C ossRe ]
22.
Lopac, N.; Ju dana, I.; B neli´c, A.; K ljan, T. Applica ion o Lase Sys ems o De ec ion and Ranging in he Mode n Road
T anspo a ion and Ma i ime Sec o . Senso s 2022,22, 5946. [C ossRe ]
23.
A nold, E.; Al-Ja ah, O.Y.; Diana i, M.; Fallah, S.; Ox oby, D.; Mouzaki is, A. A Su ey on 3D Objec De ec ion Me hods o
Au onomous D i ing Applica ions. IEEE T ans. In ell. T ansp. Sys . 2019,20, 3782–3795. [C ossRe ]
24.
Shi, S.; Wang, X.; Li, H. Poin RCNN: 3D Objec P oposal Gene a ion and De ec ion F om Poin Cloud. In P oceedings o he 2019
IEEE/CVF Con e ence on Compu e Vision and Pa e n Recogni ion (CVPR), Long Beach, CA, USA, 16–20 June 2019; pp. 770–779.
25.
Wu, J.; Xu, H.; Tian, Y.; Pi, R.; Yue, R. Vehicle De ec ion unde Ad e se Wea he om Roadside LiDAR Da a. Senso s
2020
,
20, 3433. [C ossRe ]
26.
Li, Y.; Ma, L.; Zhong, Z.; Liu, F.; Chapman, M.A.; Cao, D.; Li, J. Deep Lea ning o LiDAR Poin Clouds in Au onomous D i ing:
A Re iew. IEEE T ans. Neu al Ne w. Lea n. Sys . 2021,32, 3412–3432. [C ossRe ]
27.
Wang, H.; Wang, B.; Liu, B.; Meng, X.; Yang, G. Pedes ian ecogni ion and acking using 3D LiDAR o au onomous ehicle.
Robo . Au on. Sys . 2017,88, 71–78. [C ossRe ]
28. Peng, X.; Shan, J. De ec ion and T acking o Pedes ians Using Dopple LiDAR. Remo e Sens. 2021,13, 2952. [C ossRe ]
29.
Chen, T.; Dai, B.; Liu, D.; Zhang, B.; Liu, Q. 3D LIDAR-based g ound segmen a ion. In P oceedings o he The Fi s Asian
Con e ence on Pa e n Recogni ion, Beijing, China, 28 No embe 2011; pp. 446–450. [C ossRe ]
30.
Ka lsson, R.; Wong, D.R.; Kawaba a, K.; Thompson, S.; Sakai, N. P obabilis ic Rain all Es ima ion om Au omo i e Lida . In
P oceedings o he 2022 IEEE In elligen Vehicles Symposium (IV), Aachen, Ge many, 4–9 June 2022; pp. 37–44. [C ossRe ]
31.
Kim, G.; Eom, J.; Pa k, Y. An Expe imen o Mu ual In e e ence be ween Au omo i e LIDAR Scanne s. In P oceedings o
he 2015 12 h In e na ional Con e ence on In o ma ion Technology—New Gene a ions, Las Vegas, NV, USA, 13–15 Ap il 2015;
pp. 680–685.
32.
Hwang, I.P.; Yun, S.J.; Lee, C.H. Mu ual in e e ences in equency-modula ed con inuous-wa e (FMCW) LiDARs. Op ik
2020
,
220, 165109. [C ossRe ]
33.
Hwang, I.P.; Yun, S.j.; Lee, C.H. S udy on he F equency-Modula ed Con inuous-Wa e LiDAR Mu ual In e e ence.
In P oceedings
o he 2019 IEEE 19 h In e na ional Con e ence Communica ion Technology (ICCT), Xi’an, China,
16–19 Oc obe 2019;
pp. 1053–1056.
34.
Wallace, A.M.; Halimi, A.; Bulle , G.S. Full Wa e o m LiDAR o Ad e se Wea he Condi ions. IEEE T ans. Veh. Technol.
2020
,
69, 7064–7077. [C ossRe ]
35.
Goodin, C.; Ca u h, D.; Doude, M.; Hudson, C. P edic ing he In luence o Rain on LIDAR in ADAS. Elec onics
2019
,8, 89.
[C ossRe ]
36.
Heinzle , R.; Schindle , P.; Seeki che , J.; Ri e , W.; S o k, W. Wea he In luence and Classi ica ion wi h Au omo i e Lida Senso s.
In P oceedings o he 2019 IEEE In elligen Vehicles Symposium (IV), Pa is, F ance, 9–12 June 2019; pp. 1527–1534. [C ossRe ]