Pe cep ion O loading o Au onomous Mobili y
in a Beyond-5G Edge-enabled En i onmen
Tiago Mos a dinha∗†, João Gamei o∗†, Ped o Valen e∗†, Ped o Ri o∗†, Dua e Raposo∗, Susana Sa gen o∗†,
Ca los Ma ques‡, Miguel Mesqui a‡, Filipe Pin o‡
∗Ins i u o de Telecomunicações, 3810-193 A ei o, Po ugal
†Depa amen o de Ele ónica, Telecomunicações e In o má ica, Uni e sidade de A ei o, 3810-193 A ei o, Po ugal
‡Al ice Labs, 3810-106 A ei o, Po ugal
Abs ac —In his pape , we p opose an a chi ec u e in eg a ing
Mul i-access Edge Compu ing (MEC) nodes in o 5G base s a ions
and oadside uni s o enable dynamic se ice o ches a ion o
on-demand pe cep ion o loading. In his app oach, compu e-
in ensi e pe cep ion asks, such as objec de ec ion, a e ins an-
ia ed as se ices on nea by edge nodes, wi h esou ces being
alloca ed adap i ely based on he loca ions and ajec o ies o
he ehicles. The sys em was designed o suppo use mobili y
h ough se ice eplica ion and seamless hando e s be ween
pe cep ion se ice ins ances in di e en edge nodes as ehicles
mo e ac oss co e age a eas, ensu ing con inuous se ice deli e y
and Quali y o Se ice (QoS) gua an ees. We e alua e he
p oposed a chi ec u e in a eal-wo ld 5G edge-enabled es bed
using an au onomous ehicle unning an objec de ec ion se ice.
The esul s demons a e ha o loading pe cep ion ia 5G and
MEC yields subs an ially lowe p ocessing la ency, and ha
sma o ches a ion mechanisms a e able o eac o use mobili y
and pla o m s ess si ua ions, a oiding se ice deg ada ion,
hus alida ing he app oach o Coope a i e, Connec ed, and
Au oma ed Mobili y (CCAM) use cases.
Index Te ms—Edge O loading, 5G Ne wo ks, Vehicula Se -
ices, Au onomous Mobili y
I. INTRODUCTION
The g owing numbe o sensing de ices in sma ci ies,
anging om oadside came as o ehicle senso s, has gen-
e a ed a wide a ie y o mobili y and sensing da a ha , when
p ocessed in eal- ime, can be le e aged o op imize u ban
a ic managemen , enhance oad sa e y, and suppo da a-
d i en decision-making o ci y planning.
Howe e , adi ional cloud solu ions a e no well-sui ed o
hese applica ions. They s uggle o mee he s ic ul a-low
la ency and eliabili y equi emen s o ehicula sys ems, since
accessing emo e p ocessing acili ies ha hos cloud se ices
inhe en ly in ol es highe la ency [1]. This makes hem unsui -
able o c i ical ehicula applica ions ha demand eal- ime
o nea - eal- ime p ocessing [1]. To add ess his limi a ion,
5G sys ems in eg a ed wi h MEC ha e eme ged as powe ul
enable s o mobili y-awa e se ices. MEC educes la ency and
backhaul a ic by b inging he compu a ion close o he
ne wo k edge [2]. 5G p o ides seamless mobili y suppo and
QoS gua an ees h ough ne wo k slicing and p og ammable
in e aces o esou ce con ol. These concep s a e c ucial
o deli e sma ci y se ices capable o adap ing o use
mobili y and he condi ions o he ci y’s in as uc u e, while
main aining app op ia e QoS le els and ensu ing e icien
esou ce managemen [2].
Examples o sma ci y se ices include asks in connec ed
and au onomous ehicles (CAVs) ha ely on objec de ec ion,
acking, and o he pe cep ion- ela ed compu a ions. O en,
hese asks equi e subs an ial p ocessing powe ha exceeds
he capabili ies o indi idual ehicles which a e c ucial in he
Coope a i e, Connec ed, and Au oma ed Mobili y (CCAM)
scena ios ha in ol e connec ed Onboa d Uni s (OBUs) o e -
ing c i ical se ices in he a eas o oad sa e y and au onomous
ope a ion [3]. By o loading hese compu ing-in ensi e and
la ency-sensi i e asks o mo e capable edge de ices, CAVs
can access iche , mo e accu a e en i onmen al in o ma ion
while main aining eal- ime pe o mance [4]. Howe e , one
o he challenges o o loading hese CCAM applica ions o
5G-Edge enabled in as uc u es is add essing he mobili y
challenge o ensu ing se ice con inui y [2].
Se e al s udies ha e explo ed 5G and edge-enabled a -
chi ec u es o suppo CCAM applica ions, wi h a ocus on
ask o loading and la ency educ ion [5]–[9]. Speci ically,
o loading objec de ec ion asks o he ne wo k edge has
been shown o educe la ency and alle ia e onboa d esou ce
cons ain s by le e aging compu a ional esou ces close o
ehicles. Howe e , many exis ing solu ions o objec de ec ion
and ideo p ocessing ha e limi a ions. These solu ions o en
employ monoli hic designs ha lack modula i y, hinde ing
scalabili y and adap abili y [10]. Fo ins ance, he o loading
algo i hm p oposed in [5] conside s cell esou ces and ehicle
dis ance bu is e alua ed only h ough simula ions, lacking
eal-wo ld deploymen . Simila ly, he wo k in [6] highligh s
he bene i s o o loading compu e ision applica ions o
he edge, imp o ing la ency, esponse ime, and h oughpu ;
howe e , i does no add ess dynamic pla o m s a us, sys em
o e load, o mobili y managemen . In [7], a digi al win and
ein o cemen lea ning app oach was designed o op imize
o loading decisions based on mobili y and bandwid h awa e-
ness, bu se ice con inui y and hando e managemen emain
unadd essed. Las ly, [9] p oposed a mobili y and esou ce-
awa e o loading s a egy o ehicula edge compu ing ne -
wo ks, also e alua ed h ough simula ions. Collec i ely, p e-
ious wo ks highligh he po en ial o MEC and 5G slicing
o CCAM applica ions. Howe e , hey e eal a lack o
comp ehensi e solu ions es ed in eal-wo ld deploymen s wi h
au onomous ehicles, in eg a ing edge coo dina ion and u ban
pla o ms.
In con as , his pape p oposes a CCAM-based a chi ec u e
enabled by 5G and MEC pla o ms ha suppo s eal- ime
ideo analy ics and in elligen se ice o ches a ion capable
o adap ing o use mobili y and dynamic pla o m s a us.
The sys em was alida ed wi h a use case based on an objec
de ec ion pipeline deployed on he edge. The sys em demon-
s a es how li e ideo s eams can be cap u ed, p e-p ocessed,
and ansmi ed ia Real-Time T anspo P o ocol (RTP)/Use
Da ag am P o ocol (UDP) o be p ocessed wi h a YOLO-based
se ice a he edge,while sma o ches a ion decisions ake
in o accoun eal- ime mobili y da a, adio ne wo k s a us, and
pla o m me ics o pe o m se ice eplica ion and scaling
ac oss edge nodes in esponse o use mo emen o esou ce
o e load, p o iding an impo an suppo s uc u e capable
o main aining se ice con inui y o c i ical CCAM se ices
coupled wi h dynamic ne wo k managemen o p o ide end-
o-end se ice o ches a ion spanning edge p ocessing and
ne wo k in as uc u es.
The emainde o his wo k is o ganized as ollows. Sec-
ion II de ails he 5G-MEC a chi ec u e, emphasizing on
i s mobili y-based a chi ec u e, me ics acquisi ion, and how
i can accommoda e sma ci y se ices. Then, sec ion III
p esen s he pe cep ion se ice and he suppo o edge
o loading. Sec ion IV depic s he alida ion esul s o he
use case and edge o loading. Finally, sec ion V concludes
he wo k and p esen s u u e di ec ions.
II. EDGE-ENABLED ARCHITECTURE FOR AUTONOMOUS
MOBILITY
In ou p e ious wo k [11], we ex ended 5G MEC o ches-
a ion o include mobili y, bu wi h some cons ain s. Use s
could eques he ins an ia ion o in o ainmen o eme gency
se ices and he alloca ion o ne wo k esou ces by speci ying
hei ou e in ad ance. This eliance on p ede ined pa hs,
howe e , p e en ed he sys em om p oac i ely adap ing o
sudden ajec o y changes and limi ed i s abili y o exploi
ad anced 5G capabili ies beyond ne wo k slicing. As such,
in his pape , a new a chi ec u e is p oposed ha ex ends he
p e ious wo k o suppo unplanned ou e changes, wi h he
de elopmen o new o ches a ion mechanisms ha le e age
eal- ime ehicle posi ion da a, pe o mance me ics om he
Radio Access Ne wo k (RAN) and edge pla o ms.
A. 5G-Edge enabled A chi ec u e
The co e idea o he p oposed sys em ep esen ed in Fig-
u e 1 is o educe end- o-end la ency by op imizing he
o loading o asks o he mos app op ia e MEC node o
he mo ing use . To accomplish his, he sys em comp ises
dis ibu ed componen s ha collabo a e o deploy and eloca e
he se ice o he op imal edge zone as he use mo es.
The MEC a chi ec u e, p esen ed in Figu e 1, comp ises a
5G-edge enabled deploymen . This means ha he 5G co e
ne wo k ollows a Con ol and Use Plane Sepa a ion (CUPS)
Fig. 1: Beyond-5G and Edge O loading A chi ec u e.
a chi ec u e, wi h Use Plane Func ion (UPF) disagg ega ed
om he co e ne wo k and placed a he edge. The p esen ed
concep is al eady a es ed a chi ec u e in ou sma ci y
deploymen , as de ailed in [12]. Along he edge a chi ec u e,
he e is also a me ics’ pla o m ha moni o s and collec s
da a om he ne wo k. The da a, in o m o Key Pe o mance
Indica o s (KPIs), can be collec ed om he co e, he an-
ennas o he use equipmen s (whe he being sma phones
o modems). The Expe imen Li ecycle Manage (ELCM), a
common module used ac oss he 6G-PATH p ojec , collec s
hese KPIs o agg ega e me ics om da a p ocesso s and
collec o s, such as Ka ka and MQTT B oke s, MongoDB
and o he sou ces. This enables he use o me ics o KPIs
isualiza ion. The sys em p oposed in his wo k u ilizes he
da a om he pla o m o make decisions, based on se e al
KPI amilies. Las ly, besides p o iding in o ma ion, he 5G
sys em also p o ides an API o QoS and RAN con ol. The
con ol o e he RAN allows o hando e managemen , which
p o ides a hando e , no only awa e o he ne wo k esou ces,
bu also based on he s a e o he se ice and edge nodes.
B. O ches a ion and In elligence Componen s
An o ches a ion h ough Kube ne es k3s1clus e , wi h
nodes dis ibu ed on di e en si es, se es as he ounda-
ional laye o ins an ia ing se ices, which is subsequen ly
ex ended wi h componen s esponsible o coo dina ing edge
in elligence acco ding o he mobili y. These componen s
collec i ely ac as he o ches a o ha de e mines he mos
app op ia e ac ions o he sys em. I is comp ised o ou main
componen s: Se ice Managemen Moni o (SMM), Ac ions
P oxy (AP), Edge Applica ion Manage (EAM), and Edge
Coo dina o (EC).
1h ps://k3s.io/
Algo i hm 1 O ches a ion Wo k low
Requi e: Use Equipmen (UE), EC, MEC, SMM, A ei o Tech Ci y
Li ing Lab (ATCLL), AP, EAM, 5G API, RAN
1: UE →EC: REQSERV
2: EC →MEC: INSTSERV(Zones)
3: MEC →EC: READY
4: EC →UE: MECREADY
5: while MECREADY = alse do
6: wai
7: end while
8: Ini : wm, we, w , θ ▷ Weigh s, h eshold
9: while onT ip do
10: (m, e, )←SMM ↔ATCLL: GETDATA
11: (sm, se, s )←SMM.SCORES(m, e, )
12: s←wmsm+wese+w s
13: i s > θ hen
14: SMM →AP: MIGRSERV(Zone)
15: AP →EAM: EXECMIGR(Zone)
16: SMM →AP: TRIGHAND
17: AP →5G API: HANDOVER(gNB)
18: end i
19: end while
The SMM is a se ice-speci ic moni o ing componen . I
u ilizes he loca ions o 5G gNodeBs (gNBs) a ached o
he MEC in as uc u e and acks Coope a i e Awa eness
Messages (CAMs) o igina ing om CAVs o de e mine hei
posi ions. Using his da a, he SMM iden i ies when a ehicle
is app oaching a di e en co e age a ea and decides whe he
o ini ia e a se ice mig a ion o ha ehicle’s se ice in-
s ance. Thus, he SMM se es as he igge o p o isioning
mobili y-awa e se ices.
The AP is a se ice ha o wa ds ac ion messages om
mul iple applica ions ha moni o he eplicas associa ed wi h
he se ices. I was designed wi h in e ope abili y in mind,
enabling seamless inco po a ion o SMM om o he se ices.
I se es as a elay o alida ing he ac ion messages o
he moni o s. Fo example, when a SMM decides o enable
a se ice in MEC, i sends a message o he AP, which
o wa ds i o he EAM. This componen also ensu es ha
only legi ima e ac ions a e passed h ough he ac ua o s while
decoupling he moni o ing logic om he MEC and 5G sys em
o ches a ion logic.
The EAM is esponsible o in e ac ing di ec ly wi h he
MEC in as uc u e. This componen in e aces wi h he un-
de lying k3s clus e o manage he ac ions eques ed by he
pods. In his sys em, i can enable o disable pod ins ances
and pe o m he a o emen ioned ac ions on pods belonging o
speci ic zones. In addi ion, i can be eques ed o show which
pods a e enabled, disabled, o heal hy.
The EC manages he esou ce alloca ion o he MEC
in as uc u e o se ices eques ed by use s. When needed
a REST API pos is used o ins an ia e se ices o p ocess
hei da a. I he se ice is al eady a ailable on he MEC,
i will espond ha he in as uc u e is eady. I i is no
eady, i will selec he op imal edge zones o ins an ia ion
on he k3s clus e . When he pods a e eady, EC e u ns ha
he in as uc u e is eady.
Algo i hm 1 ou lines he o ches a ion wo k low o MEC
in elligence o ches a ion. I illus a es how he o ches a-
o ’s componen s in e ac o implemen mobili y-awa e se ice
logic. In his wo k low, he EC ecei es a se ice ins an ia ion
eques om he UE o deploy a se ice on he MEC k3s
clus e . The SMM con inuously moni o s ehicle mobili y,
RAN condi ions, and edge node s a us o de e mine whe he
he ehicle should be eassigned o a new edge node and
gNB. The decision o ins an ia e o mig a e a se ice is based
on he ehicle’s loca ion and p oximi y o he nea es gNB
along i s ajec o y. When he SMM de e mines ha a ehicle
should ansi ion o a new zone, i ini ia es se ice mig a ion
and eques s a hando e o he ehicle’s modem o ano he
gNB. The AP o wa ds hese eques s o he ele an ac ua o s,
including he EAM and he 5G API, o execu e he necessa y
ac ions seamlessly.
In conclusion, he p oposed mobili y-awa e edge o loading
a chi ec u e o e s a signi ican ly mo e adap i e and obus
solu ion han he p e ious app oach [11]. I o e comes he
limi a ion o elying on ixed use ou es by le e aging eal-
ime mobili y da a and igh ly in eg a ing wi h he 5G con ol
plane, allowing he sys em o eloca e se ices o he nea es
edge node as a ehicle’s ajec o y changes. This deep coupling
o 5G ne wo k in elligence wi h MEC ope a ions ensu es
con inuous low-la ency se ice deli e y and op imal esou ce
usage e en unde dynamic condi ions.
III. MOBILITY-AWARE OBJECT DETECTION SERVICE
In CCAM scena ios, au onomous ehicles mus accu a ely
and p omp ly de ec objec s in o de o pe cei e hei su -
oundings and make sa e decisions in eal- ime. Howe e ,
pe o ming his ask solely on he ehicle demands signi i-
can esou ces, whe eas o loading o cloud in as uc u es e-
sul s in la ency incompa ible wi h sa e y-c i ical equi emen s.
The p oposed solu ion in oduces a scalable, mobili y-awa e
pipeline o objec de ec ion explici ly designed o CCAM
con ex s. This modula pipeline o mic ose ices manages
he objec de ec ion ask, encompassing acquisi ion, de ec ion,
and isualiza ion o objec s. The acquisi ion and p ocessing
mic ose ices a e designed o be s a eless, acili a ing seamless
hando e be ween eplicas o suppo use mobili y. The ame
acquisi ion ask is deployed in ehicles, while he objec de-
ec ion and isualiza ion p ocesses a e handled on edge nodes.
This s ack o mic ose ices is scalable, allowing mul iple
s eams o be ed o one o mo e objec de ec ion se ices.
The componen s o objec de ec ion asks a e F ame Acqui e
and F ame Inges o . The a chi ec u e is depic ed in Figu e 2.
Each F ame Acqui e ins ance, unning on a ehicle, han-
dles one came a s eam. Upon s a up, i opens a ne wo k
socke o announce i s p esence o he F ame Inges o se -
ice by b oadcas ing, pe iodically, a unique iden i ie . Then,
i uses a GS eame 2 amewo k o cap u e he li e ideo
eed om he came a. The F ame Acqui e also pe o ms
ame scaling by encoding each ame in he H.264 o ma
2h ps://gs eame . eedesk op.o g/
Fig. 2: Objec de ec ion se ice a chi ec u e.
o imp o e e iciency. The ou going ames a e ansmi ed
h ough RTP, wi h unique iden i ie s using RTP heade . In
mobili y scena ios, RTP s a eless na u e, which uses UDP
as he anspo laye p o ocol, suppo s as anspo ac oss
dynamic pa hs. E en wi h packe loss, he sys em bene i s om
he mos ecen a ailable da a, which aligns wi h eal- ime
de ec ion use cases o which edge compu ing was designed. I
enables seamless hando e s wi hou session in e up ion, and
consequen ly ensu es unin e up ed se ice consump ion.
The F ame Inges o se ice, deployed on edge nodes and
used on a ailable edge compu ing esou ces, comp ises wo
igh ly coupled componen s: he Recei e and he De ec ion
Se ice. The ecei e handles he ecei ed s eams and b oad-
cas ed announcemen s, iden i ying each F ame Acqui e and
egis e ing hei unique iden i ie and s eam de ails as hey
become a ailable. Once egis e ed, he ecei e con inuously
moni o s a designa ed UDP po o incoming RTP s eams.
All came as ansmi hei H.264-encoded ideo o his sha ed
po . The ecei e hen demul iplexes he ames and co ela es
hem wi h he announcemen da a by examining he embedded
F ame Acqui e iden i ie in he RTP heade s. As packe s
a i e, he ecei e emo es he RTP heade s, associa es each
packe wi h he co ec came a, and econs uc s comple e
ames om he s eam. Since a single ideo ame may span
mul iple RTP packe s, he ecei e bu e s he da a un il a ull
ame is assembled and sen o he objec de ec o .
The objec de ec o p ocesses each ame using he YOLO
model3, a s a e-o - he-a algo i hm widely adop ed o eal-
ime objec de ec ion. YOLO iden i ies objec s such as ehi-
cles, pedes ians, and bicycles, anno a ing hem wi h bounding
boxes di ec ly on he image. The gene a ed me ada a p o ides
aluable in o ma ion abou he ehicle’s su oundings, sup-
po ing he c ea ion o he Wo ld Model—a undamen al com-
ponen o enhancing oad sa e y and enabling au onomous
mobili y.
The modula design o he se ice simpli ies deploymen
and main enance. Each componen is con aine ized, enabling
apid deploymen on edge nodes and acili a ing easy scaling
and upda es compa ed o monoli hic sys ems. In summa y, his
design allows each componen o be dynamically ins an ia ed
o eloca ed ac oss mo ing ehicles and edge nodes wi hou
e aining p e ious s a e. These ea u es yield a scalable, e-
silien objec de ec ion se ice ha is p epa ed o mo ing
ehicle en i onmen s.
3h ps://docs.ul aly ics.com/models/yolo11/
IV. USE CASE VALIDATION
To alida e he sys em, a CCAM-based use case o an
au onomous ehicle equipped wi h a came a was de ised
o he objec de ec ion se ice desc ibed in he p e ious
sec ion. I was inco po a ed in he objec i es o he Eu opean
p ojec 6G-PATH4. Besides i s de elopmen and alida ion,
he e we e open demons a ions o edge o loading wi h a eal
au onomous ehicle o he public. This sec ion p esen s he
use case, i s esul s, and hei impac on he se ice.
A. Desc ip ion
In his use case, he au onomous ehicle is equipped wi h
a came a ha s eamed ideo o e he cellula ne wo k o
he edge in as uc u e, whe e he p e iously p esen ed objec
de ec ion se ice p ocessed he oo age. The esul s o he
objec de ec ion ask, p ocessed by he o loaded se ice a
he edge, can p o ide aluable insigh s o he au onomous
ehicle and o he cloud se ices. Fo example, de ec ing oad
signs and a ic ligh s inc eases he ehicle’s awa eness o i s
su oundings, while iden i ying o he ehicles and pedes ians
can help p e en acciden s and imp o e a ic e iciency. O he
cloud se ices can use hese da a o op imize a ic low and
upda e maps wi h conges ion in o ma ion, con ibu ing o an
imp o ed pe cep ion awa eness o o he en i ies in he CCAM
en i onmen . The s eam p ocessing se ice was eplica ed on
demand in sui able edge nodes along he ehicle’s pa h. Gi en
he ul a-low la ency and high eliabili y equi emen s o his
use case, a key goal o he demons a ion was o showcase
sma o ches a ion capabili ies and dynamic ne wo k man-
agemen in a CCAM scena io, and as such, one o he node’s
memo y and CPU we e s essed o showcase he ac ion o
hose o ches a ion and ne wo k managemen mechanisms.
The demons a ion low s a ed when he ehicle began he
ip accessing he se ice om an o e loaded edge node. I s
mobili y and pla o m s a us in o ma ion e lec ed he si ua-
ion, and sma o ches a ion mechanisms igge ed a eplica
ac i a ion on a heal hy node alongside a ne wo k hando e
o he nea es an enna (gNB) o o load he se ice and he
ne wo k connec ion, ensu ing ha he QoS would no be
deg aded. The esul s e idence he impo ance o a sys em
and applica ion like his in a CCAM en i onmen wi h an
au onomous ca .
(a) PIXKIT Au onomous ehicle (b) Edge Node wi h gNB
Fig. 3: Equipmen used in public ials o he use case.
4h ps://6gpa h.eu
B. Expe imen al Se up and KPIs
The se up o he use case combined an au onomous ehicle,
a non-public 5G ne wo k, and an edge compu ing clus e .
The au onomous ehicle is a PIXKIT5(Figu e 3a) which had
onboa d a p ocessing de ice, a came a, and a SIMCom modem
o p o ide a 5G ups eam link. In he p ocessing de ice,
PIXKIT he ollowing ins an ia ed se ices: he GPS loca ion
se ice, he ehicula ne wo k s ack (Vane za-NAP [13]), and
he F ame Acqui e se ice desc ibed in he p e ious sec ion.
The 5G in as uc u e includes wo CableF ee gNBs loca ed
along he p ede e mined demons a ion pa h, and Open5GS as
he co e ne wo k. The edge compu ing in as uc u e equi ed
he c ea ion o a k3s clus e , whe e one Vi ual Machine (VM)
se ed as he mas e node and wo N idia Je son boa ds se ed
as edge nodes, each wi h a co esponding gNB, ep esen ing
sepa a e edge zones (Figu e 3b).
F ame Inges o ins ances we e dis ibu ed be ween he edge
nodes, whe e hey pe o med he objec de ec ion ask and
wo ked in coope a ion wi h he ehicle. In his se up, one o
he eplicas was s essed, while he o he was heal hy. When
he ne wo k hando e occu ed, he se ice was o loaded o
he heal hie and close edge node, hus imp o ing no only
Quali y o Expe ience (QoE), specially in he p ocessed ideo
s eam, bu also he sa e y o e e yone on he oad.
This use case p o ided KPIs om bo h in as uc u e (RAN)
and se ice, ga he ed in he clien isualiza ion applica ion.
F ames pe second and he numbe o ames d opped we e
also ep esen a i e o he quali y o he s eam, hus indica ing
he pe cep ion o he quali y by use s. The KPIs a e ep e-
sen ed in able I, and hei de ini ion ollows he speci ica ion
by he 6G Sma Ne wo ks and Se ices Join Unde aking
(SNS JU) [14].
Fig. 4: Bi a e measu ed inside he se ice eplicas.
i-0: ame-inges o 0, i-1: ame-inges o 1.
C. Resul s dissemina ion
Figu e 4 illus a es one o he p e iously men ioned KPIs,
o e ing insigh s in o how he o ches a ion mechanisms e-
sponded o he node o e load scena io. The bi a e was ini ially
high, bu once he o e load p ocess s a ed, i d opped o
alues below 10 Mbps. The o ches a ion mechanisms de ec ed
he si ua ion and ac i a ed eplica ame-inges o -0 ( i-0
in Figu e 4) in he heal hy node, which caused a sligh inc ease
in bi a e, ha ose o alues highe han 12 Mbps. I is clea
ha , al hough pe o mance is sligh ly be e on he heal hy
node, i does no each he le el obse ed on he i s node
be o e and a e he o e load pe iod. This is due o he ac ha
he au onomous ehicle was mo ing away om he gNB, hus
deg ading he ne wo k connec ion wi h i s mo emen , despi e
he edge in as uc u e co ec ly iden i ying he o e loaded
node.
Figu e 5 (lowe cha ) shows a g ea e imp o emen in
he numbe o p ocessed ames pe minu e on he heal hy
node compa ed o he o e loaded one. On he s essed node,
whe e CPU and RAM we e unde hea y load, he objec
de ec ion se ice expe ienced delays and pe o mance issues.
Howe e , he si ua ion imp o ed signi ican ly once he eplica
on he heal hy node was ac i a ed and ame p ocessing was
edi ec ed o i . The heal hy node demons a ed highe pe o -
mance han he o he one, e en be o e he o e load p ocess
s a ed, because i was mo e powe ul. This u he highligh s
he impo ance o in elligen o ches a ion mechanisms o e -
ec i ely managing se ice loads ac oss he e ogeneous nodes.
The inal se ice-speci ic me ic unde analysis is he num-
be o ames d opped by each one o he eplicas ( ep esen ed
in Figu e 5, in he uppe cha ). I also demons a es he
impo ance o he sma o ches a ion mechanisms, because
as soon as he o e load p ocesses s a ed, he numbe o
ames d opped, and hen s a ed o ise, causing pe o mance
deg ada ion. Howe e , once he second eplica is ac i a ed
and i s a s ecei ing ames, i is clea ha he numbe o
ames d opped su e s an imp o emen , wi h only a esidual
numbe o ames being d opped in he eplica swi ch momen .
A e wa d, he applica ion’s pe o mance s abilized, wi h no
addi ional ame loss un il he ame p ocessing was edi ec ed
back o he i s eplica.
The esul s o his demons a ion alida e he co ec ope a-
ion o he sys em and i s impo ance in a CCAM en i onmen .
Se ice in e up ions we e a oided h ough he applica ion’s
s a eless design, which enabled seamless handling o use
mobili y by edi ec ing a ic o he nea es an enna ha
p e en ed se ice deg ada ion. Fu he mo e, in elligen o ches-
a ion mechanisms de ec ed a c i ical o e load si ua ion and
ook ac ion o p e en comple e se ice in e up ion ha would
comp omise he ope a ion o his c i ical se ice. O loading
pe cep ion applica ions o au onomous ehicles o he edge
o e s signi ican bene i s bu demands ul a-high eliabili y
and low la ency. The p o ision o p oac i e esou ces based
on mobili y was a c ucial addi ion o he in as uc u e enabled
5h ps://www.pixmo ing.com/pixki 3
Family KPI Desc ip ion Ta ge Value
Compu e
con aine _memo y_usage_by es To al memo y usage o a con aine (P ome heus
de aul ).
≥minimum equi ed o ame p ocessing.
node_load1 1-minu e load a e age on a node (P ome heus
de aul ).
≥minimum equi ed o ame p ocessing.
cpu_u iliza ion Pe cen age o CPU used by con aine s. ≤80% o a oid bo lenecks.
scale_ou _la ency Time o deploy addi ional con aine s. ≤500 ms.
Co e age
cqi Channel quali y indica o epo ed by UE o
gNB.
Ta ge 15, ≥13 op imal.
i Rank Indica o o MIMO (2x2). Ta ge 2 (no an enna co ela ion).
pusch_sn Las ecei ed PUSCH SNR. ≥15 (good quali y).
s p Re e ence-Signal Recei ed Powe (dBm). ≥ −85 dBm.
Da a Ra e dl_bi a e, ul_bi a e Downlink, Uplink bi a e in bi s/s. ≥ equi ed by se ice (e.g., 10 Mbps).
peak_dl_bi a e Max achie able downlink bi a e unde ideal
condi ions.
≥100 Mbps.
Reliabili y dl_ e x Numbe o downlink e ansmi ed blocks. Ideally 0, some accep able.
ul_ e x Numbe o uplink blocks wi h CRC e o s. Ideally 0, some accep able.
dl_ x, ul_ x Numbe o downlink, uplink blocks ansmi ed
wi hou e ansmissions.
>dl_ e x.
Posi ioning cam_longi ude, cam_la i ude Longi ude, la i ude o he ehicle. Accu a e posi ion acking.
Mobili y hando e _success_ a e Pe cen age o success ul hando e s. ≥99%.
hando e _la ency Time o comple e a hando e (ms). ≤50 ms.
La ency o ches a ion_la ency Time o mig a e a se ice o ano he node (ms). ≤100 ms.
Se ice Quali y num_p ocessed_ ames Numbe o ames p ocessed by se ice eplicas. ≥minimum equi ed FPS, else mig a e se ice.
TABLE I: KPI De ini ion o he Use Case.
Fig. 5: F ames d opped and ames pe minu e analysis
i-0: ame-inges o 0, i-1: ame-inges o 1.
by he 5G-edge, and he in eg a ion o all o hese pla o ms
was shown o be a i s esea ch e o o build a coope a i e
awa eness pla o m capable o adap ing o he dynamic na u e
o he 5G-edge pla o ms and he use mobili y inhe en o his
en i onmen . The esul s o his demons a ion yielded wo
da ase s, which include he ex ac ed KPIs and KVIs and a e
a ailable o public consul a ion [15] [16]. A demons a ion
o he use case is also a ailable in his ideo6.
V. CONCLUSIONS
This pape p oposed and alida ed an edge o loading ap-
p oach in 5G Edge-enabled en i onmen s, wi h an au onomous
ehicle and objec de ec ion se ice. In a comple e in eg a ion
o 5G ne wo ks, MEC se ices, and ehicula se ices, his
s udy deli e s: an a chi ec u e, a de ini ion o e ie able KPIs,
an o ches a ion wo k low, sys em alida ion, esul s and a
public demons a ion. The esul s alida e he co ec ope a ion
o he sys em and i s impo ance in a CCAM en i onmen .
Se ice in e up ions we e a oided h ough he applica ion’s
s a eless design, which enabled seamless handling o use
mobili y by edi ec ing a ic o he nea es an enna ha
p e en ed se ice deg ada ion. Fu he mo e, in elligen o ches-
a ion mechanisms de ec ed a c i ical o e load si ua ion and
ook ac ion o p e en comple e se ice in e up ion ha would
comp omise he ope a ion o his c i ical se ice. As u u e
lines o wo k, we aim o u he imp o e he in eg a ion wi h
he 5G ne wo k, mainly h ough he use o 3GPP de ined
APIs in he co e ne wo k and RAN. In he o ches a ion side,
desc ip o s o an easie deploymen o se ices can allow
o scalabili y. While he wo k p esen ed is mainly ocused
on an objec de ec ion se ice, a u u e esea ch di ec ion
mo e ocused on he edge in as uc u e would include he
6h ps://you u.be/-4GHgNh4D6I, accessed 10 Sep embe 2025
ex ension o his a chi ec u e and i s alida ion wi h new use
cases in mo e complex scena ios in ol ing a la ge numbe o
use s consuming di e en ypes o se ices hos ed on he edge
pla o m.
VI. ACKNOWLEDGMENT
This wo k was suppo ed by he EU’s HE esea ch and in-
no a ion p og amme HORIZON-JU-SNS-2023 unde he 6G-
PATH p ojec (G an No. 101139172), and by he Eu opean
Union/Nex Gene a ion EU, h ough P og ama de Recupe -
ação e Resiliência (PRR) [P ojec N . 29: Rou e 25 (02/C05-
i01.01/2022.PC645463824-00000063)].
REFERENCES
[1] B. R. Che uku i, “Edge compu ing s. cloud compu ing: A compa a i e
analysis o eal- ime ai applica ions,” In e na ional Jou nal Fo Mul i-
disciplina y Resea ch, ol. 6, pp. 1–17, 10 2024.
[2] S. M. H. Hosseini, M. J. Joo iah, D. Rocha, J. A. Almeida, P. Ba -
olomeu, J. Fe ei a, . Rosales, and M. M. Mi anda, “Coope a i e,
connec ed and au oma ed mobili y se ice con inui y in a c oss-bo de
mul i-access edge compu ing ede a ion scena io,” on ie s in u u e
anspo a ion, ol. 3, no. 911923, pp. 1–16, July 2022.
[3] J. O iz, P. J. Fe nández, R. Sanchez-Ibo a, J. B. Be nabe, J. San a,
and A. Ska me a, “En o cing gdp egula ion o ehicula 5g commu-
nica ions using edge i ual coun e pa s,” in 2020 IEEE 3 d 5G Wo ld
Fo um (5GWF), 2020, pp. 121–126.
[4] E. Co onado, G. Ceb ián-Má quez, and R. Riggio, “Enabling au-
onomous and connec ed ehicles a he 5g ne wo k edge,” in 2020
6 h IEEE Con e ence on Ne wo k So wa iza ion (Ne So ), 2020, pp.
350–352.
[5] S. B. De agi i and M. Vijayalakshmi, “Enhancing 5g mobili y and
pe o mance h ough mul i-access edge compu ing in eg a ion and coop-
e a i e compu ing o loading,” in 2024 Second In e na ional Con e ence
on Ne wo ks, Mul imedia and In o ma ion Technology (NMITCON),
2024, pp. 1–7.
[6] M. V. B. Da Sil a, M. Ba bosa, A. Quei oz, and K. L. Dias, “A 5g-
edge a chi ec u e o compu a ional o loading o compu e ision appli-
ca ions,” in 2025 In e na ional Con e ence on In o ma ion Ne wo king
(ICOIN), 2025, pp. 93–98.
[7] X. Chen, J. Cao, Y. Sahni, M. Zhang, Z. Liang, and L. Yang, “Mobili y-
awa e dependen ask o loading in edge compu ing: A digi al win-
assis ed ein o cemen lea ning app oach,” IEEE T ansac ions on Mobile
Compu ing, ol. 24, no. 4, pp. 2979–2994, 2025.
[8] P. F. Moshi i, M. Simsek, and B. Kan a ci, “On he in e play be ween
ne wo k me ics and pe o mance o mobile edge o loading,” in ICC
2024 - IEEE In e na ional Con e ence on Communica ions, 2024, pp.
4018–4023.
[9] Y. Li, C. Yang, X. Chen, and Y. Liu, “Mobili y
and dependency-awa e ask o loading o in elligen assis ed
d i ing in ehicula edge compu ing ne wo ks,” Vehicula
Communica ions, ol. 45, p. 100720, 2024. [Online]. A ailable:
h ps://www.sciencedi ec .com/science/a icle/pii/S221420962300150X
[10] R. Xu e al., “Edge ideo analy ics: A su ey on applica ions, sys ems
and challenges,” IEEE Communica ions Su eys & Tu o ials, 2023.
[11] J. Gamei o, R. Rosmaninho, G. Pe na, P. Ri o, S. Sa gen o, C. Ma ques,
and F. Pin o, “Edge compu ing and 5g ne wo k in eg a ion o mobili y-
awa e se ice deploymen s,” Pe asi e and Mobile Compu ing, 2025,
Submi ed.
[12] P. Valen e, D. Raposo, P. Ri o, and S. Sa gen o, “Disagg ega ed mobile
co e o edge ci y se ices,” in 2023 IEEE 24 h In e na ional Symposium
on a Wo ld o Wi eless, Mobile and Mul imedia Ne wo ks (WoWMoM),
2023, pp. 157–166.
[13] R. Rosmaninho e al., “Vane za-nap: Vehicula communica ions and
se ices in mic ose ices a chi ec u es,” in 2024 IEEE Vehicula Ne -
wo king Con e ence (VNC), 2024, pp. 297–304.
[14] I. Mesogi i, A. Ga as, K. T ichias, A. Kaloxylos, F. Fon es,
A. Cha emis, M. Gho aishi, M. Mollà Roselló, P. Yad anjee Aghdam,
G. Agapiou, A. Sgambellu i, M. Al-Rawi, C. Ma ques, and W. Moh ,
“6g kpis - de ini ions and a ge alues,” Ma . 2025. [Online]. A ailable:
h ps://doi.o g/10.5281/zenodo.15011631
[15] I. de Telecomunicações and A. Labs, “6g-pa h_wp1_da ase _uc- ci ies-
1_expe imen al_edge_pe o mance_i ,” Sep. 2025. [Online]. A ailable:
h ps://doi.o g/10.5281/zenodo.17092093
[16] ——, “6g-pa h_wp1_da ase _uc- ci ies-1_expe imen al_ an_ a ic_i ,”
Sep. 2025. [Online]. A ailable: h ps://doi.o g/10.5281/zenodo.17092369
