Recei ed 26 Ma ch 2025, accep ed 13 Ap il 2025. Da e o publica ion 00 xxxx 0000, da e o cu en e sion 00 xxxx 0000.
Digi al Objec Iden i ie 10.1109/ACCESS.2025.3561238
Mobili App: A Deep Lea ning-Based Tool o
T anspo Mode De ec ion o Suppo
Sus ainable U ban Mobili y
GERARD CARAVACA IBAÑEZ, LUIS J. DE LA CRUZ LLOPIS , (Membe , IEEE),
ADRIAN CATALÍN DIACONEASA, ALBERTO BAZÁN GUILLÉN ,
AND MÓNICA AGUILAR IGARTUA , (Membe , IEEE)
Depa men o Ne wo k Enginee ing, Uni e si a Poli ècnica de Ca alunya (UPC), 08034 Ba celona, Spain
Co esponding au ho : Mónica Aguila Iga ua ([email p o ec ed])
This wo k was suppo ed in pa by Spanish Go e nmen unded by MCIN/AEI/10.13039/501100011033 unde Resea ch P ojec s
’’DIs ibu ed Sma Communica ions wi h Ve i iable EneRgy-op imal Yields (DISCOVERY)’’ unde G an PID2023-148716OB-C32 and
’’Enhancing Communica ion P o ocols wi h Machine Lea ning while P o ec ing Sensi i e Da a (COMPROMISE)’’ unde G an
PID2020-113795RB-C31; pa ially unded by MCIN/AEI/10.13039/501100011033 and by he Eu opean Union (EU)
Nex Gene a ionEU/PRTR (Plan de Recupe ación, T ans o mación y Resiliencia) unde Resea ch P ojec ’’Anonymiza ion Technology o
AI-Based Analy ics o Mobili y Da a (MOBILYTICS)’’ unde G an TED2021-129782B-I00; in pa by he P edoc o al Schola ship
‘‘Gene ación de Conocimien o - P ojec s Call 2022’’ unde G an PRE2021-099830; and in pa by he Gene ali a de Ca alunya unde
AGAUR (Agència de Ges ió d’Aju s Uni e si a is i de Rece ca) G an 2021-SGR-01413.
ABSTRACT The isible e ec s o clima e change in u ban a eas a e d i ing a pa adigm shi in socie al and
poli ical p io i ies. U ban planne s, public anspo p o ide s, and a ic manage s a e inc easingly ocused
on edesigning ci ies o p omo e sus ainable mobili y and c ea e g een spaces o pedes ians, cyclis s, and
scoo e use s. In alignmen wi h hese objec i es, he Eu opean Clima e Law manda es a minimum 55%
educ ion in g eenhouse gas emissions by 2030 and clima e neu ali y by 2050. Achie ing hese a ge s
equi es obus ools o collec and analyze mobili y da a, enabling he e alua ion o ci izens’ a el habi s
and he planning o sus ainable u ban in as uc u e. This s udy p esen s Mobili App, a ool de eloped
based on a deep lea ning (DL) model o eal- ime de ec ion o anspo a ion modes using sma phone
senso da a. Ou app oach le e ages a hie a chical model combining con olu ional neu al ne wo ks (CNNs)
o ea u e ex ac ion and long sho - e m memo y (LSTM) laye s o empo al p ocessing, enhanced by
skip connec ions. To ensu e compu a ional e iciency on mobile de ices, he sys em in eg a es s a is ical
echniques o ea ly mo ion de ec ion, minimizing eliance on DL models. The model was ained on a
da ase o mul imodal ips in Ba celona, achie ing o e 80% accu acy o mos anspo modes and a
weigh ed a e age accu acy o 88%. These esul s highligh he e ec i eness o ou app oach o accu a ely
p edic ing use s’ anspo modes du ing hei ips. The Mobili App ool p o ides an in ui i e pla o m
o collec ing and analyzing u ban mobili y da a. By analyzing a el pa e ns, anspo modes, and mode-
swi ching beha io s, i deli e s ac ionable insigh s o ci y planne s, aiding in he enhancemen o u ban
mobili y, p omo ion o sus ainable de elopmen , and ansi ion o g eene ci ies.
INDEX TERMS T anspo mode de ec ion, ac i i y ecogni ion, deep lea ning, mobili y senso s, sus ainable
u ban mobili y, Mobili App.
I. INTRODUCTION
The isible impac o clima e change in u ban a eas is d i ing
a signi ican ans o ma ion in socie al and poli ical p io i ies.
The associa e edi o coo dina ing he e iew o his manusc ip and
app o ing i o publica ion was Mouquan Shen .
This shi has p omp ed u ban planne s, public anspo
p o ide s, and a ic manage s o u gen ly edesign ci ies
wi h a ocus on sus ainable mobili y and he c ea ion o g een
spaces dedica ed o pedes ians, cyclis s, and scoo e use s.
In esponse o he clima e c isis, he Eu opean Pa liamen
enac ed he Eu opean Clima e Law, which manda es a
VOLUME 13, 2025
2025 The Au ho s. This wo k is licensed unde a C ea i e Commons A ibu ion-NonComme cial-NoDe i a i es 4.0 License.
Fo mo e in o ma ion, see h ps://c ea i ecommons.o g/licenses/by-nc-nd/4.0/ 1
G. C. Ibañez e al.: Mobili App: A DL-Based Tool o T anspo Mode De ec ion
minimum 55% educ ion in g eenhouse gas emissions
by 2030 and es ablishes clima e neu ali y as a legally binding
objec i e by 2050. P i a e ehicles, which con ibu e 15% o
he EU’s CO2emissions, emain a c i ical a ge o hese
e o s. To achie e hese goals, u ban planne s equi e obus
ools o collec and analyze eal-wo ld mobili y da a. Such
ools enable he e alua ion o ci izens’ a el habi s and
suppo he planning o essen ial in as uc u e changes, such
as iden i ying op imal loca ions o mobili y hubs, con e ing
s ee s in o pedes ian zones, and enhancing public anspo
sys ems o os e mo e sus ainable u ban en i onmen s.
To suppo u ban mobili y planne s in achie ing hei
goals, his pape p o ides an in-dep h explo a ion o
sma phone-based sys ems o anspo a ion mode de ec ion.
The widesp ead adop ion o sma phones has e olu ionized
da a collec ion, enabling dynamic and eal- ime moni o -
ing o anspo a ion pa e ns. These de ices, endowed
wi h di e se senso s and compu ing capabili ies, enable
con inuous da a collec ion on indi idual mo emen s and
anspo a ion choices. To suppo his goal, we de eloped
Mobili App [4], a ool designed o analyze ci izens’ mobili y
pa e ns, including a el o igins and des ina ions, modes o
anspo a ion used, and ans e poin s be ween di e en
modes. Ou ool accu a ely iden i ies 12 dis inc anspo
modes and human ac i i ies wi h an imp essi e accu acy o
app oxima ely 88%. Designed o seamless in eg a ion, he
code can unc ion as a lib a y wi hin applica ions managed
by public anspo en i ies, ully complian wi h s ingen
p i acy egula ions. This unc ionali y empowe s u ban
mobili y planne s wi h p ecise insigh s in o ci izens’ mobili y
pa e ns, suppo ing in o med decisions o enhance u ban
mobili y and os e sus ainable anspo a ion solu ions.
Ou goal is o suppo municipal en i ies, such as he
Au o i a del T anspo Me opoli à (ATM) o Ba celona [1],
in enhancing u ban mobili y and ad ancing owa ds mo e
sus ainable ci ies. This s udy aims o ully le e age sma -
phone echnology o anspo mode de ec ion, au oma ically
iden i ying a pe son’s mode o anspo . We add ess key
aspec s such as algo i hm de elopmen , da a analysis, p e-
p ocessing echniques, p i acy conside a ions, and p ac ical
implemen a ion. By ans o ming he unde s anding and
managemen o anspo a ion, we aim o con ibu e o he
c ea ion o mo e sus ainable, heal hie , and use - iendly
u ban en i onmen s, wi h a speci ic ocus on he ci y o
Ba celona as a case s udy.
To achie e he desi ed ou come, we ha e enginee ed a
sophis ica ed hie a chical deep lea ning model, i.e. a machine
lea ning model de eloped h ough he sys ema ic cons uc-
ion o deep neu al ne wo ks o ganized in a hie a chical s uc-
u e. This amewo k ma ies he s eng hs o con olu ional
neu al ne wo ks (CNNs) [32] o high-le el ea u e ex ac-
ion, wi h he capabili ies o long sho - e m memo y (LSTM)
[24] ne wo ks o p ocessing sequen ial da a. To enhance
i s e iciency and obus ness, his in eg a ed app oach is
u he augmen ed wi h equency analysis echniques and
GPS loca ion me hods. This combina ion gua an ees a
comp ehensi e and e icien sys em o add essing he ask
a hand.
The main key con ibu ions o his wo k a e lis ed as
ollows:
•In his wo k, we in oduce Mobili App, a ool capable o
p edic ing ci izens’ ac i i ies and anspo a ion modes
du ing hei ips ac oss a me opoli an a ea. Mobili-
App ea u es an ad anced hie a chical deep lea ning
model designed wi h a s ong emphasis on minimizing
compu a ional cos s, which is pa icula ly impo an as
he ool is in ended o un on sma phones wi h limi ed
p ocessing powe . This p edic i e amewo k p o ides
u ban mobili y planne s and public anspo p o ide s
wi h p ecise insigh s in o ci izens’ mobili y pa -
e ns, acili a ing in o med decision-making o enhance
u ban mobili y and p omo e sus ainable anspo a ion
solu ions.
•We ha e de eloped a hie a chical deep lea ning model
o p edic ci izens’ ac i i ies and anspo a ion modes,
suppo ing u ban planne s in os e ing sus ainable u ban
mobili y. To ain he model, we c ea ed a da ase using
h ee sma phone senso s (accele ome e , magne ome-
e , and gy oscope), achie ing an a e age accu acy
o 88%. This model is used o cha ac e ize ci izens’
mul imodal ips, o e ing aluable insigh s in o u ban
mobili y pa e ns and aiding e o s o p omo e mo e
sus ainable anspo a ion p ac ices.
•We ha e de eloped an And oid applica ion, named
Mobili App [4], able o pe o m hese asks: (i) eal-
ime da a acquisi ion o sma phone mobili y senso s;
(ii) da a p e-p ocessing; (iii) anspo mode p edic ion
using ou hie a chical deep lea ning model; and
(i ) use eedback showing he p edic ed mode o e a
map. The sys em is s uc u ed hie a chically consis ing
o : (i) a kinema ic mo ion classi ie o as de ec Walk
and S a iona y s a es ha de ine consecu i e segmen s
o he mul imodal ip (commu ing poin s); (ii) he
p oposed hie a chical deep lea ning model; and
(iii) a s op-de ec ion algo i hm ha au oma ically
iden i ies he end o a ip and hal s senso da a collec ion
on he sma phone, ensu ing ha no addi ional da a is
ga he ed once he ip has concluded.
•We ha e es ed he Mobili App ool in a la ge-scale
mul imodal a el da a collec ion campaign wi hin he
Uni e si a Poli ècnica de Ca alunya (UPC) communi y
in Ba celona. This campaign p oduced an ex ensi e
da ase o mul imodal ips, consis ing o consecu i e
segmen s p edic ed by he Mobili App ool. The col-
lec ed da a shows p omising po en ial o u u e analysis
o he mobili y habi s o a communi y (e.g., a Campus,
a ci y, a company).
The es o his wo k is s uc u ed as ollows. Sec-
ion II p o ides backg ound and e iews ela ed s udies,
es ablishing he con ex o ou esea ch. Sec ion III-D
ou lines he me hods used o da a p e-p ocessing and he
da ase collec ion p ocess. Then, Sec ion IV discusses he
2VOLUME 13, 2025
G. C. Ibañez e al.: Mobili App: A DL-Based Tool o T anspo Mode De ec ion
model a chi ec u e and i s pe o mance ou comes. Sec ion V
p esen s a comp ehensi e o e iew o he And oid appli-
ca ion Mobili App [4], which encapsula es ou comple e
sys em. A e wa ds, Sec ion VI includes a pe o mance
e alua ion o he Mobili App p oposal, showcasing esul s
om a pilo campaign on mul imodal ip collec ion. Finally,
Sec ion VII o e s concluding ema ks and sugges s po en ial
u u e esea ch di ec ions.
II. RELATED WORK
This s udy is inspi ed by a di e se ange o me hodologies,
encompassing bo h classical and mode n app oaches. I in e-
g a es adi ional machine lea ning algo i hms and senso
equency analysis wi h insigh s om ecen ad ancemen s
in deep lea ning models.
A. TRADITIONAL PROPOSALS
Ini ially, he ask o anspo mode classi ica ion ocused on
manual ea u e ex ac ion and aining o machine lea ning
models using adi ional algo i hms such as suppo ec o
machine [22], k-nea es neighbou [14] o decision ees [35].
An example o ha is he Hemminki e Al. [23] p oposal.
This p oposal is based on a hie a chical ea u e-based
sys em composed o h ee s ages. The hie a chy begins wi h
a p oposal ha ini ially dis inguishes be ween pedes ian
mo ion and o he ypes o mo ion a a coa se le el. I he
kinema ic mo ion classi ie ails o de ec signi ican physical
mo emen , such as walking, he p ocess ad ances o a
s a iona y classi ie . This classi ie hen assesses whe he he
use is ei he s a iona y o on some o m o mo o ized ans-
po a ion. When mo o ized anspo a ion is iden i ied, he
classi ica ion p ocess mo es on o a mo o ized anspo a ion
classi ie . This classi ie is esponsible o ca ego izing he
ongoing ac i i y in o one o i e modali ies: bus, ain, me o,
am, o ca , and i employs adap i e boos ing (AdaBoos )
as a s a is ical classi ica ion me a-algo i hm o enhanced
accu acy.
The e a e o he s udies ha , in addi ion o using senso
da a, use GPS da a o imp o e classi ica ion. An example o
his is he app oach p oposed by L. Randle e al. [33]. The
plan was o p ima ily ely on accele ome e da a unless i was
no clea wha mode o anspo a ion was being used. This
way, hey sa e ene gy and keep he algo i hm simple by using
mo e powe -hung y senso s and da a p ocessing only when
necessa y. I is wo h no ing ha he algo i hm was es ed on
a ela i ely small da ase , equi ing use s o label hei ips
wi h bo h de ice o ien a ion and anspo a ion mode du ing
he aining phase. None heless, he app oach o le e aging
addi ional senso s o complemen accele ome e da a when
necessa y is in e es ing.
B. DEEP LEARNING METHODS
In he con ex o machine lea ning and a i icial in elligence,
he impo ance o ea u e selec ion has been a long-
s anding challenge. T adi ionally, he p ocess o iden i ying
and selec ing ele an ea u es om aw da a has been a
c ucial s ep in building e ec i e p edic i e models. Howe e ,
mode n deep lea ning echniques ha e ans o med his ield
by elimina ing he need o manual ea u e selec ion, enabling
he di ec use o aw da a o asks such as ac i i y ecogni ion
and anspo a ion mode p edic ion.
One o he pionee ing echniques in his ega d is he use
o CNNs [18]. CNNs ha e shown ema kable capabili ies in
unde s anding spa ial ela ionships wi hin da a, making hem
highly e ec i e in asks like image ecogni ion. When applied
o ac i i y ecogni ion, CNNs can di ec ly p ocess aw senso
da a, allowing o he au oma ic ex ac ion o ele an ea u es
om he inpu .
On he o he hand, inspi ed by he ans o ma i e success
o deep lea ning in ac i i y ecogni ion and anspo a ion
mode p edic ion, J.V. Jeyakuma e al. p oposed a no el
deep lea ning model known as he Deep Con olu ional
Bidi ec ional LSTM (DCBL) [26]. DCBL akes aw senso
da a as inpu , combining he powe o con olu ional laye s
o cap u e spa ial pa e ns and bidi ec ional LSTMs o
model empo al dependencies. This app oach enables p ecise
p edic ions o anspo a ion modes wi hou he complexi y o
ea u e enginee ing o selec ion.
C. HUMAN ACTIVITY RECOGNITION
Rega ding p oposal speci ically o ien ed o Human Ac i i y
Recogni ion (HAR), he wo k in [40] p oposes a ada -based
HAR app oach using hype -dimensional compu ing (HDC),
speci ically designed o iden i y six human ac i i ies: walk-
ing, walking while ca ying a ca on, walking wi h a s ick,
walking in a squa posi ion, unning, and jumping. The HDC
sys em p o es o be ene gy-e icien , obus , ligh weigh ,
and capable o as lea ning wi h low la ency, making i
ideal o eal- ime, on-chip human ac i i y ecogni ion in
u u e ada sys ems. The wo k [17] de elops Op iMappe ,
a unin o med c oss-subjec ans e lea ning amewo k
o ac i i y ecogni ion ha ex ac s abs ac knowledge
ac oss subjec s and u ilizes his knowledge o de eloping
a pe sonalized and accu a e ac i i y ecogni ion model in
new subjec s. They ocus on daily and spo human ac i i y
ecogni ion (jumping, descending s ai s, walking, unning,
cycling). The wo k [9] p oposes an heu is ic algo i hm
o a decision ee o au oma e a decision-making o
HAR o accu a ely de ec passenge s ge ing in o public
anspo a ion sys ems. They use a public WiFi-based ac i i y
ecogni ion da ase o ex ac human ac i i y ea u es om
Channel S a e In o ma ion powe , which changes caused by
nea by human ac i i y.
D. TRANSPORT MODE DETECTION
The wo k [25] p oposes a combined solu ion o a Long-Sho -
Te m Memo y ne wo k and a Healing algo i hm (used o
co ec misde ec ed modes based on majo i y o ing) capable
o ecognizing 12 anspo a ion modes (walking, unning,
climbing s ai s, descending s ai s, bicycle, mo o cycle, ca ,
subway, ain, high-speed ain, am and me obus). The
VOLUME 13, 2025 3
G. C. Ibañez e al.: Mobili App: A DL-Based Tool o T anspo Mode De ec ion
au ho s in [28] p opose a model able o ecognize 6 ans-
po a ion modes: ca , bus, ain, bicycle, walking and e ical
ac i i ies (ele a o , escala o , and s ai s) om a la ge da ase
wi h o e 16,000 pa icipan s going abou hei daily ac i i y
in he ci y-s a e o Singapo e. The da ase includes ba ome e ,
accele ome e , and Wi-Fi scanne da a. They ocus hei
wo k on iden i ying e ical mobili y a ci y-scale and i s
po en ial o ack and iden i y e ical anspo a ion in a
densely buil u ban en i onmen . The wo k [34] in oduces a
anspo a ion mode ecogni ion h ough using mul imodal
da a om wea able senso s: mo ion, sound and ision.
Th ee independen deep neu al ne wo k (DNN) classi ie s
wo k wi h he h ee ypes o senso s, espec i ely. Then
wo schemes use he classi ica ion esul s om he h ee
mono-modal classi ie s. Resul s show ha while pe o mance
is educed o each indi idual classi ie , he bene i s o
usion a e e ained wi h pe o mance imp o ed by 15%.
While mo ion and sound senso s a e a ailable in cu en
sma phones, ision should be ob ained om body-wo n
came as such as eye-wea compu e s (e.g. Google Glass,
Spec acles by Snap).
Inspi ed by hese and o he app oaches, in his wo k we
ha e designed a no el hie a chical deep lea ning model o
ecognize anspo a ion modes and ela ed human ac i i ies
(s ill and walk). Ou model uses con olu ional neu al
ne wo ks (CNN) o high-le el ea u e ex ac ion and LSTM
ne wo ks o p ocessing sequen ial da a, o de elop a
p edic i e model o ele en ypes o anspo and human
ac i i ies: Bike, bus, ca , mo o bike, un, s a iona y, subway,
ain, am, walk, e-scoo e . Addi ionally, we employ p e-
p ocessing echniques o s anda dize, augmen , and e ine he
da ase , enhancing he aining and e alua ion o machine
lea ning models. Building on his, we de eloped h ee
algo i hms capable o cha ac e izing comple e mul imodal
ips (comp ising mul iple segmen s) each associa ed wi h a
speci ic anspo mode, in eal ime:
•Mic o-s a e algo i hm. I classi ies mo ion wi hin each
20-second sample se , as de ailed in Algo i hm 2.
•Mac o-s a e algo i hm. I classi ies mo ion ac oss e e y
i e mic o-s a es (equi alen o 100 sec.) by employing
a majo i y- alue scheme, as desc ibed in Algo i hm 1.
•S op algo i hm. I au onomously de ec s when a use
concludes hei ip, as i is ou lined in Algo i hm 3.
III. MOBILITY DATASET OF MOBILITY SENSORS
This sec ion ou lines he me hodology used o he c ea ion
and e inemen o he da ase cen al o his s udy. I begins
wi h a de ailed desc ip ion o he senso s employed, ol-
lowed by an explana ion o he da a collec ion p ocess.
Subsequen ly, he p e-p ocessing echniques applied o he
aw da a a e de ailed. To u he enhance he da ase ,
his sec ion also desc ibes se e al da a augmen a ion ech-
niques. Finally, he sec ion add esses ea u e ex ac ion,
explaining how he p ep ocessed da a was ans o med in o
a s uc u ed ea u e se o use in he ou lie de ec ion
p ocess.
A. DISCUSSION ABOUT AVAILABLE DATASETS
REGARDING MOBILITY
Be o e c ea ing ou own mobili y da ase , we i s e al-
ua ed he sui abili y o exis ing public da ase s such as
Sussex-Huawei Locomo ion Da ase (SHL) [19],[38], T ans-
po Mode De ec ion Da ase (TMD) [11] and Collec y
Da ase [16]. Howe e , we de e mined ha none o he
a ailable da ase s sa is ied he speci ic equi emen s o his
p ojec .
The SHL da ase [19] is a ich esou ce o de eloping
ac i i y ecogni ion echnologies, o e ing a wide a ay o
senso da a (2812 h du ing 7 mon hs) ha cap u es a de ailed
iew o use en i onmen s and ac ions (8 anspo modes).
I includes da a collec ed om 15 sma phones, a ied senso
inpu s, in eg a es hi d-pa y da a o benchma king, and
p o ides c ucial in o ma ion on de ice s a us and p ecise
posi ioning. Howe e , i s complexi y may de e simple
applica ions, and edundancies, ha dwa e speci ici y, and
limi ed pa icipan di e si y pose challenges o i s b oade
applicabili y and he gene aliza ion o de i ed models.
The TMD da ase [11] is a obus esou ce o ac i i y
ecogni ion esea ch, ma ked by i s di e se pa icipan base
(13 use s) and igo ous p e-p ocessing, o e ing a solid
ounda ion o de eloping and es ing models ac oss a ious
sma phone models. Howe e , i s limi a ions, including
limi ed da a du a ion (31 h), and a small ange o ac i i y
classes (5 anspo modes), may cons ain i s e ec i eness in
cap u ing he ull complexi y and di e si y o human mo ing
ac i i ies.
Finally, he Collec y da ase [16] p esen s a comp ehensi e
o e iew o anspo a ion habi s, encompassing a wide a ay
o anspo modes (8 anspo modes, 15 use s), high amoun
o da a ga he ed (242 h du ing 5 mon h) and p o iding
de ailed usage du a ion while ensu ing pa icipan anonymi y.
Howe e , i s geog aphical ocus on Zag eb, C oa ia, he
inaccessibili y o he da a collec ion app ( o ga he da a om
o he ci ies), and he lack o de ice-speci ic in o ma ion may
limi i s b oade applicabili y and in oduce challenges in
ensu ing eplica ion and senso da a consis ency in esea ch.
In conclusion, he e is a no able absence o a comp ehen-
si e public da ase sui able o e ec i ely aining machine
lea ning models o p edic anspo a ion modes in u ban
en i onmen s. An ideal da ase would encompass da a om
a di e se ange o u ban esiden s and include a wide a ie y
o anspo a ion modes.
B. SMARTPHONE SENSORS USED IN MOBILITAPP
In ou da ase , da a was collec ed om h ee mo ion senso s
embedded in he sma phone: accele ome e ,gy oscope, and
magne ome e . These h ee senso s ha e he ad an age o
being p esen in nea ly all sma phones on he ma ke ,
p o iding a comp ehensi e iew o he use ’s mo emen . Fo
he gy oscope and magne ome e , da a was collec ed di ec ly
om he aw senso ou pu s. In he case o he gy oscope
and he magne ome e , da a has been collec ed di ec ly om
4VOLUME 13, 2025
G. C. Ibañez e al.: Mobili App: A DL-Based Tool o T anspo Mode De ec ion
he pu e senso s. In he case o he accele ome e , howe e ,
i was decided o use he linea accele a ion alue, since i
p o ides in o ma ion abou accele a ion along he x,y, and
zaxes, excluding he e ec s o g a i y, o a oid he masking
e ec s o g a i y on he p ecise eadings o he accele ome e
senso .
Fo he pu pose o da a collec ion and subsequen da ase
compila ion, an And oid sma phone applica ion was de el-
oped. Called Mobili App [4], his applica ion is able o
collec da a om he phone senso s du ing use s’ ips.
I has wo main unc ionali ies: (i) he i s one is o
collec senso da a om ips labeled by ou olun ee s; and
(ii) he second one is o es he li e anspo de ec ion
model du ing mul imodal a el. A isual explana ion o his
comp ehensi e p ocess o da a acquisi ion is p o ided in
Fig. 1. In addi ion, a de ailed use manual and ideo u o ial
a e a ailable on he Mobili App websi e [4].
A e alida ing he unc ionali y and accu acy o he
Mobili App sys em, we ca ied ou mul iple da a collec ion
campaigns o ob ain su icien senso samples o all
anspo a ion modes included in he Mobili App da ase .
These da a samples we e essen ial o he e ec i e aining
o ou p edic ion model. Consequen ly, he esul ing da ase
inco po a es samples om 50 sma phones, exceeding he
use coun s o p e iously analyzed da abases. This la ge
sample size is expec ed o imp o e he model’s abili y o
gene alize e ec i ely o new use s.
C. PRE-PROCESSING STEPS
The e icien p ocessing and analysis o senso da a collec ed
du ing a ious jou neys p esen a signi ican challenge,
pa icula ly due o he inhe en a iabili y in jou ney du a ion
and condi ions. To add ess his challenge, he subsequen sub-
sec ions examine a se ies o ad anced echniques designed
o s anda dize,augmen , and e ine he da ase , enabling
mo e e ec i e aining and e alua ion o machine lea ning
models.
1) SLIDING WINDOW
The da a collec ed om sma phone senso s du ing a jou ney
a e s o ed in CSV- o ma iles. Howe e , hese da ase s o en
display a iabili y in jou ney du a ions. To s anda dize and
p ep ocess his da a e ec i ely, a echnique known as he
sliding window me hod is applied. Adjus ing he window
size and o e lap ac o is no i ial; hese pa ame e s
signi ican ly in luence he g anula i y and quali y o he
analyzed da a. Smalle windows migh cap u e ine de ails
bu isk missing b oade ends, while la ge windows
migh smoo h o e impo an nuances. Simila ly, highe
o e lap migh p o ide mo e con inuous da a analysis bu
inc ease edundancy and compu a ional load. Con e sely,
lowe o e lap migh miss c i ical ansi ional in o ma ion. Fo
ou use case, a e conside able es ing, we de e mined ha
he op imal alues o window size and o e lapping ac o
we e 512 and 50%, espec i ely.
2) DATA AUGMENTATION
Da a augmen a ion is a c i ical p ocess in he de elopmen o
obus machine lea ning models, pa icula ly when wo king
wi h da ase s ha need o ep esen a complex and a iable
eal-wo ld phenomenon. The objec i e o da a augmen a ion
is o a i icially inc ease he size and enhance he quali y
o he da ase by gene a ing new syn he ic examples de i ed
om he o iginal da a. These modi ica ions enable he model
o lea n om a mo e di e se se o examples, which is
essen ial o enhancing i s gene aliza ion capabili ies. Fo his
pu pose, in his wo k wo da a augmen a ion unc ions ha e
been designed: ob usca ion and o a ion.
On he one hand, ob usca ion echniques y o a oid
ecogni ion o phone-speci ic pa e ns. Essen ially, he p o-
cess in ol es adding noise o he phone’s senso da a. Bo h
addi i e and mul iplica i e noise a e in oduced o enhance
he da ase . Addi i e noise se es o impose a consis en noise
loo ac oss he signal, while mul iplica i e noise modi ies
he signal in a scale-dependen ashion. The applica ion
o whi e noise is pa icula ly ad an ageous because o
i s consis en equency spec um, which e ec i ely masks
de ice-speci ic noise signa u es while p ese ing he signal’s
spec al cha ac e is ics.
On he o he hand, o a ion echniques y o add ess he
a iabili y o phone o ien a ions, which can signi ican ly
a ec senso eadings. Thus, he da ase is augmen ed by sim-
ula ing a ious po en ial phone o ien a ions ha may occu in
eal-wo ld scena ios du ing a use ’s jou ney. This is achie ed
by pe o ming o a ions o he h ee-dimensional senso da a
along he X, Y, and Z axes. Fo his, andom o a ion angles
om 0◦ o 180◦a e gene a ed. Subsequen ly, a co esponding
h ee-dimensional o a ion ma ix is cons uc ed based on
hese angles. The senso da a o each ip is hen ans o med
by his o a ion ma ix, e ec i ely eo ien ing he da a o
e lec a di e en phone posi ion.
These app oaches ensu e ha he model emains ocused
on iden i ying di e en anspo a ion modes, a oiding
dis ac ions caused by mino , i ele an de ails un ela ed o
he p ima y ask.
3) FEATURE EXTRACTION FOR OUTLIER DETECTION
Fea u e ex ac ion o ou lie de ec ion in ol es iden i ying
and selec ing he mos ele an ea u es om a da ase o
de ec ou lie s ha de ia e signi ican ly om he no m and
can ep esen anomalies, e o s, aud o a e e en s. Fea u e
ex ac ion helps algo i hms o iden i y mo e e ec i ely hese
anomalies, imp o ing analysis accu acy and e iciency.
T adi ional me hods like Z-sco e a e o en cons ained by
assump ions o linea i y and no mal dis ibu ion. In con-
as , he Mahalanobis dis ance (MD) [29] o e s a mo e
nuanced app oach, especially in mul i a ia e con ex s. Unlike
Euclidean dis ance, MD accoun s o co ela ions be ween
a iables, making i pa icula ly e ec i e o iden i ying
ou lie s in mul idimensional da ase s. This cha ac e is ic is
especially ad an ageous o his da ase due o he inclusion
o da a om mul iple senso ypes.
VOLUME 13, 2025 5
G. C. Ibañez e al.: Mobili App: A DL-Based Tool o T anspo Mode De ec ion
FIGURE 1. Applica ion in e ace o he Mobili App ool o collec ing senso da a. Fi s , he use ini ia es da a collec ion
om he mobili y senso s, ensu ing ha each segmen is labeled wi h he co esponding anspo mode. A he end o he
ip, he use s ops he da a collec ion p ocess and has he op ion o send he da a o ou se e . [4].
TABLE 1. Summa y o he ea u es ex ac ed om he Mobili App senso
da ase .
To conduc his s udy, we u ilized ea u es ex ac ed om
he aw da a, which we e u he p ocessed using he p incipal
componen analysis (PCA) echnique. Wi h he educed
ea u e se , we calcula ed he Mahalanobis dis ance o
each obse a ion. The inal s ep in ol ed iden i ying ou lie s
based on he calcula ed MDs. We used he Chi-Squa e
Dis ibu ion o de e mine a h eshold o wha cons i u es
an ou lie . A 95% con idence in e al was chosen, meaning
ha obse a ions wi h a MD g ea e han he co esponding
Chi-Squa e alue we e conside ed ou lie s. This s a is ical
app oach p o ides an objec i e c i e ion o ou lie de ec ion.
Fo he ea u e ex ac ion p ocess, we chose o ex ac
h ee ypes o ea u es: s a is ical ea u es, equency domain
ea u es, and ime domain ea u es. Table 1shows a
desc ip ion o he ea u es ex ac ed om he Mobili App
senso da ase .
D. RESULTING MOBILITAPP SENSOR DATASET
A e comple ing a ious da a collec ion campaigns, we ha e
gene a ed he Mobili App Senso Da ase , which is publicly
a ailable [8]. As men ioned abo e, i consis s o he X, Y and
Z axes o h ee mo ion senso s: Accele ome e , Gy oscope
and Magne ome e . The ile is 10,500 lines long and 2.9 GB
in size. No e ha he da a ga he ed in his da ase we e
collec ed du ing 15 mon hs in di e en campaigns, om
Ma ch 2023 o June 2024 om a o al o 50 di e en
sma phones, ep esen ing 298 hou s o da a collec ing ime.
Volun ee s labeled hei unimodal ips wi hin he Ba celona
Me opoli an A ea using nine di e en anspo a ion modes
and h ee human ac i i ies ela ed o hese modes: Bike, Bus,
Ca , Mo o bike, Run, S a iona y, Subway, T am, T ain, Walk,
e-Bicycle, e-Scoo e . Volun ee s collec ed ac i i y samples
while ca ying he sma phone in a ious posi ions, including
in hei hand, pocke , and bag.
The dis ibu ion o hou s by anspo mode, wi hou
including ye any da a augmen a ion, is depic ed in Fig. 2.
The da ase p esen s a signi ican challenge: achie ing a
mo e balanced dis ibu ion o hou s ac oss all anspo a ion
modes, as he da a is une enly dis ibu ed. This imbalance
is p ima ily due o olun ee s’ endency o use he mos
common modes o anspo a ion (bus, subway, ca , walking).
This issue is commonly obse ed in o he simila da ase s
ha we e analyzed [11],[16],[19]. I should be no ed ha ,
in his s udy, he e-bicycle mode o anspo will no be
conside ed o aining he p edic ion model, as we ha e no
ye collec ed su icien samples. Howe e , we plan o epea
he aining in he nea u u e o include his sus ainable
mode o pe sonal anspo . Consequen ly, Fig. 3illus a es
he senso da ase a e applying classical da a augmen a ion
echniques o a i icially expand i s size and enhance i s
6VOLUME 13, 2025
G. C. Ibañez e al.: Mobili App: A DL-Based Tool o T anspo Mode De ec ion
TABLE 2. Compa ison o key ea u es o Mobili App senso da ase
compa ed o simila da ase s.
FIGURE 2. Mobili App senso da ase (July 2024). Class dis ibu ion o he
da ase collec ed using Mobili App [4], a e applying p e-p ocessing
s eps, excluding da a augmen a ion echniques.
FIGURE 3. Mobili App Senso Da ase (July 2024), a e applying da a
augmen a ion echniques, see Sec. III-C2.
quali y by gene a ing new syn he ic examples de i ed om
he o iginal da a, see Sec. III-C2.
Compa ed o he o he e e enced da ase s summa ized in
Sec ion III-A, ou da ase p esen s a highe olume o hou s
eco ded han TDM and simila han Collec y. Howe e ,
i is no ably smalle when con as ed wi h he SHL da ase .
Ne e heless, i is pe inen o no e ha ou da ase includes a
b oade ange o anspo modes han i s coun e pa s and a
highe a ie y o use s. Table 2summa izes he key ea u es
o he da ase s analyzed, as well as hose o ou own da ase ,
Mobili App.
IV. DESIGN OF A HIERARCHICAL PREDICTION MODEL
FOR ACTIVITY AND TRANSPORTATION MODE
RECOGNITION
To de elop he hie a chical p edic ion model, we ocused on
minimizing compu a ional cos s, as Mobili App is designed
o un on sma phones wi h limi ed p ocessing powe . As a
esul , he model was designed o be ligh weigh and easily
op imized, ensu ing e icien ope a ion on mobile de ices.
We expe imen ed wi h di e en a chi ec u es, om ully
connec ed ne wo ks o ecu en a chi ec u es including
con olu ional ne wo ks. A e all he expe imen a ion, he
a chi ec u e ha bes i s ou use case is he one shown in
Fig. 4. Ou p oposal akes inspi a ion om he e icacy o
he mixed model p oposed in [37], and addi ionally i also
includes a hie a chical design. Rega ding he hype pa ame e
op imiza ion p ocess, we conduc ed ex ensi e expe imen s
o ine- une he unc ions and pa ame e s unde lying he
hie a chical deep lea ning model. Below, we summa ize he
inal design choices and key ea u es o he model, which a e
also depic ed in Table 3.
The model is composed o an ini ial unk o con olu ional
blocks. This is he mos impo an pa o he model o ea u e
ex ac ion. The achie emen o hese blocks should allow he
model o ex ac he ea u es om he da a, which allows
he es o he p edic ion model o di e en ia e be ween he
a ious modes o anspo . In his block a ious pa ame e s
and con igu a ions ha e been expe imen ed. As a esul , he
bes con igu a ion ex ac ed om his pa o he s udy is
he one using dila ed con olu ions wi h ke nel sizes o 7,
5 and 3; and wi h he ini ial block’s s uc u e composed
o con olu ional laye s using ReLU ac i a ion (a nonlinea
unc ion ha ou pu s he inpu di ec ly i i is posi i e and ze o
o he wise, enabling e icien aining by in oducing spa si y)
[39], wi h ba ch no maliza ion ( o no malize he inpu s o
hidden s a es wi hin he ne wo k ac oss a ba ch, s abilizing
and accele a ing aining), spa ial d opou (a egula iza ion
echnique ha andomly d ops en i e ea u e maps du ing
aining o educe o e i ing) o 5% [36], and max pooling (a
downsampling echnique commonly used in neu al ne wo ks
o educe he spa ial dimensions o an inpu olume by
cap u ing he mos p ominen ea u es while disca ding less
ele an in o ma ion).
Re aining he con olu ional amewo k desc ibed abo e,
his model inco po a es skip connec ions ha ou e he
ea u e maps di ec ly in o dedica ed LSTM laye s wi h
128 neu ons. Conc e ely, he ou pu s om he i s , second,
hi d, and ou h max-pooling s ages a e each ed in o
sepa a e LSTM uni s. Such a design enables indi idual
LSTMs o cap u e and analyze empo al pa e ns a a ious
le els o ea u e abs ac ion, en iching he model’s capaci y
o disce n complex empo al ela ionships wi hin he da a.
Finally, he e is a dense laye (also known as a ully con-
nec ed laye ) applying a L2 egula ize [13] wi h pa ame e
0.001, which p ocesses he ea u es lea ned by he LSTM;
and also a dense ou pu laye wi h a so max ac i a ion
VOLUME 13, 2025 7
G. C. Ibañez e al.: Mobili App: A DL-Based Tool o T anspo Mode De ec ion
FIGURE 4. A chi ec u e o he p oposed hie a chical p edic ion model
designed in Mobili App.
unc ion which ou pu s he ma ching p obabili ies o each
anspo mode.
A. DESIGN OF HIERARCHICAL MODEL PARAMETERS AND
PERFORMANCE EVALUATION
This sec ion p esen s a pe o mance e alua ion o ou p o-
posed hie a chical deep lea ning model o p edic ing ac i i y
and anspo a ion mode, ained using he con igu a ion
summa ized in Table 3.
No ice ha ou hie a chical deep lea ning model has
been explici ly designed o accoun o he class imbalance
p esen in ou da ase . Fo ins ance, o quan i y he di e ence
be ween he model’s p edic ions and he ac ual labels and
op imize he model du ing aining, we employed a Weigh ed
Ca ego ical C ossen opy wi h 0.1 smoo hing as he loss
unc ion (see Table 3). This app oach includes he ollowing
elemen s:
•Ca ego ical C ossen opy (CCE): A loss unc ion com-
monly used in mul i-class classi ica ion asks, which
measu es he di e gence be ween he p edic ed p oba-
bili y dis ibu ion and he ue one-ho encoded labels.
In a one-ho encoding scheme, he co ec class is
ep esen ed by a p obabili y o 1, while all o he classes
a e assigned a p obabili y o 0.
•Weigh ed CCE: Assigns di e en weigh s o each class,
penalizing misclassi ica ion o unde ep esen ed classes
mo e hea ily, he eby mi iga ing he e ec s o class
imbalance.
•Label smoo hing (0.1 smoo hing): Replaces s ic
one-ho labels wi h so e p obabili y dis ibu ions,
educing o e con idence in p edic ions and imp o ing
gene aliza ion. Fo example, in a ou -class scena io,
a adi ional one-ho encoding o [0, 1, 0, 0] would be
adjus ed o [0.025, 0.9, 0.025, 0.025] wi h a smoo hing
ac o o α= 0.1. This p e en s he model om
becoming o e ly con iden in i s ou pu s and enhances
gene aliza ion, pa icula ly in imbalanced da ase s.
To u he mi iga e class imbalance, we also applied
unde sampling o majo i y classes such as ca and public
anspo , he eby educing hei dominance and imp o ing
o e all da ase balance.
Mo eo e , addi ional da a ebalancing echniques could be
explo ed o enhance ep esen a ion in he bicycle, e-bicycle,
e-scoo e , and mo o cycle classes. Two p omising me hods
o ou da ase include:
•(Syn he ic Mino i y O e -sampling Technique) [12]:
Gene a es syn he ic samples by in e pola ing exis ing
da a poin s om mino i y classes, inc easing hei
ep esen a ion.
•ADASYN (Adap i e Syn he ic Sampling) [21]: Expands
on SMOTE by adap ing sample gene a ion based on he
local dis ibu ion o mino i y classes, ocusing on a eas
wi h highe classi ica ion di icul y.
We plan o in es iga e hese ebalancing s a egies as pa o
ou u u e wo k in he coming mon hs.
We pa i ioned he senso da ase in o h ee subse s:
(i) T aining se used o he aining o he model (80% o he
da a); (ii) Valida ion se used o gi e an unbiased e alua ion
o he model ained by uning he hype pa ame e s o he
model (10% o he da a); and (iii) Tes se used as an unbiased
e alua ion o he inal model (10% o he da a). We ha e
ensu ed ha he alida ion and es se s consis o da a om
use s no included in he aining se . This s a egy aims o
be e emula e eal-wo ld scena ios, whe e he model mus
gene alize o unseen use s, he eby p o iding a mo e obus
e alua ion o i s pe o mance. The o al aining ime o he
model was 2 hou s on he i9-13900F compu e speci ied in
Table 3.
We can see he pe o mance e alua ion o he model bo h
in he con usion ma ix in Fig. 5and in he classi ica ion
epo in Table 4. A con usion ma ix ep esen s he eal class
( ue label) in ows and he p edic ed class (p edic ed label) in
columns. In he con usion ma ix (see Fig. 5), we p o ide an
example o he case o he e-Scoo e : (i) T ue posi i e (TP)
is he numbe o posi i es co ec ly iden i ied (yellow squa e
zone); (ii) ue nega i e (TN) is he numbe o nega i es
co ec ly iden i ied (b own squa e zone); (iii) alse posi i e
(FP) is he numbe o nega i es inco ec ly iden i ied as
posi i e ( ed ec angle zone); and (i ) alse nega i e (FN)
8VOLUME 13, 2025
G. C. Ibañez e al.: Mobili App: A DL-Based Tool o T anspo Mode De ec ion
TABLE 3. Hype pa ame e con igu a ion o he hie a chical model.
is he numbe o posi i es inco ec ly iden i ied as nega i es
(blue ec angle zone).
Table 4shows he mos common me ics o e alua ing
he pe o mance o p edic ion models: (i) P ecision, as he
pe cen age ha he model co ec ly p edic s posi i e when
making a decision, see Eq. (1); (ii) Recall, also known as
sensi i i y, as he pe cen age o posi i es co ec ly iden i ied
ou o all he exis ing posi i es, see Eq. (1); and (iii) F1-
sco e, he ha monic mean o P ecision and Recall as mo e
sui able me ic o e alua ing pe o mance in scena ios wi h
unbalanced classes, see Eq. (2). We calcula ed he model’s
a e age weigh ed Accu acy, see Eq. 3, achie ing a alue o
88%, as de ailed in Table 4. This a e age accu acy alue
was weigh ed by he numbe o samples o each mode o
ai ly e lec he con ibu ion o each anspo mode in i s
compu a ion.
P ecision =TP
TP +FPRecall =TP
TP +FN (1)
F1−sco e =2·P ecision ·Recall
P ecision +Recall (2)
Accu acy =TP +TN
TP +TN +FP +FN (3)
I is clea om he esul s shown in Fig. 5 ha while mos
o he anspo s a e success ully p edic ed wi h an accu acy
abo e 80%, i becomes e iden ha modes wi h a spa se
amoun o aining da a (see Bike, T ain and e-Scoo e in
Fig. 2) end o be misiden i ied mo e equen ly by he model.
This obse a ion highligh s an in iguing poin : Bike and
e-Scoo e a e equen ly misclassi ied as one ano he by he
model: (i) T ue Bike is w ongly p edic ed as e-Sco e wi h
27% p obabili y; while (ii) ue e-Sco e is w ongly p edic ed
as Bike wi h 13% p obabili y. This is bo h in e es ing
and easonable, conside ing ha hese modes o anspo
sha e simila i ies in accele a ion and handling cha ac e is ics.
Ne e heless, as we collec addi ional Bike and e-Scoo e
senso samples om ou olun ee s (see Fig. 2), he p edic i e
FIGURE 5. Con usion ma ix o he hie a chical Mobili App p edic ion
model. Values a e exp essed as ac ions o one. Fo example, Ca is
co ec ly ecognized 96% o he ime, i is con used wi h Bus 1% o he
ime, and i is con used wi h T ain 1% o he ime.
TABLE 4. Classi ica ion epo on he es se o he hie a chical
Mobili App p edic ion model. The a e age weigh ed Accu acy is 88%.
accu acy o hese anspo modes in ou Mobili App model
will ce ainly imp o e. The case o he Mo o bike is also
in e es ing, as i shows a highe accu acy despi e ha ing a
simila numbe o samples o he bike and e-scoo e . This is
due o i s inhe en olling-o e d i ing beha io , which is well
cap u ed by he gy oscope, making i easie o iden i y his
anspo mode.
A e e iewing hese esul s, we a e now eady o conduc
a compa a i e analysis o he ou selec ed p edic ion models,
e alua ed h ough a obus me hodological amewo k. The
app oach adop ed o his compa ison in ol ed g oup- old
c oss- alida ion wi h i e olds [6], execu ed ac oss h ee
di e en andom seeds. This p ocedu e culmina ed in a o al
o 15 uns pe model, ensu ing a comp ehensi e assessmen
o each model’s pe o mance consis ency and esilience o
a ying da a spli s. The models compa ed in his sec ion a e
he ollowing:
VOLUME 13, 2025 9
G. C. Ibañez e al.: Mobili App: A DL-Based Tool o T anspo Mode De ec ion
TABLE 6. Di e ence in s o age size and pe o mance o he o me
classi ica ion model and he cu en model op imized using Tenso low
Li e.
TABLE 7. Compa ison o he ba e y consump ion o he Mobili App
applica ion wi h o he well-known And oid applica ions.
VI. PERFORMANCE EVALUATION OF MOBILITAPP
To e alua e he pe o mance o he Mobili App ool,
we ocused on he in eg a ion o he s a e-based sys em
o de e mining Mac o-s a es (see Sec ion V-A) and he
kine ic mo ion classi ie o Mic o-s a e de e mina ion (see
Sec ion V-B). Fou case s udy examples we e analyzed o
illus a e ansi ions be ween s a es:
Fi s expe imen : T ansi ion WALK o STATIONARY.
Fig. 14 shows accele ome e samples and he s a e clas-
si ica ion. In he obse ed expe imen , wo Mic o-s a es
we e inaccu a ely labeled as MOVING due o po en ially
poo loca ion da a accu acy. No ably, he Mac o-s a es
show esilience agains hese anomalies, unde sco ing hei
obus ness in il e ing ou isola ed inaccu acies in Mic o-s a e
labeling.
Second expe imen : T ansi ion STATIONARY o BUS.
Fig. 15 shows accele a ion peaks in he ea ly samples,
co esponding o he use boa ding and se ling in o he bus.
A his s age, he bus emains s a iona y as he passenge s
ake hei sea s. Va iabili y in he egula i y o he analyzed
windows is no ed, a ibu able o inconsis en GPS signal
quali y which can cause delayed o occasionally p ema u e
loca ion upda es. We obse e he absence o e oneous s a es,
which unde sco es he clean and accu a e cap u e o use
ac i i y du ing he expe imen , along wi h he co ec ac i i y
classi ica ion p edic ed by ou Mobili App ool.
Thi d expe imen : T ansi ion BUS o WALK. The esul s
depic ed in Fig. 16 p esen some i egula windows al hough
no spu ious Mic o-s a es. As in he o he expe imen s, we can
conclude ha he Mobili App ool accu a ely cap u ed he
use ac i i y and co ec ly classi ied bo h he ac i i y and
anspo a ion mode du ing he expe imen .
Fou h expe imen : Real-li e mul imodal ips in
Ba celona. A e es ing he sys em wi h he p e ious
h ee expe imen s, he nex phase in ol es a eal-wo ld
e alua ion o mul imodal ips. Tes ing was conduc ed
using se e al sma phone models no p e iously included
in he da ase o simula e he expe ience o new use s.
Tes s iden i ied occasional inaccu acies in he loca ion on
he map, especially when using he subway. Howe e ,
hese inaccu acies we e la gely mi iga ed by he model’s
obus p edic ion capabili ies o he subway mode o
anspo , ensu ing eliable classi ica ion despi e GPS e o s.
To add ess his issue, we modi ied Mobili App o handle
GPS inaccu acies independen ly o he lowcha depic ed in
Fig. 9. Speci ically, when Mobili App de ec s a signi ican
GPS e o , i immedia ely in okes he machine lea ning
model o p edic he use ’s mode o anspo , allowing he
sys em o iden i y whe he he use is a eling on he subway.
This app oach e ec i ely esol es issues ela ed o GPS
inaccu acies commonly encoun e ed in he subway.
Figu e 17 p esen s wo examples o es s pe o med wi h
wo sma phones ha we e no used du ing aining (Samsung
Galaxy S23 and Xiaomi Redmi No e 12). Simila ly, Fig. 18
shows addi ional es s conduc ed wi h wo o he sma phones
(Xiaomi Hype OS and Samsung Galaxy A54) ha we e no
used du ing aining. Nei he o hese de ices con ibu ed da a
o he da ase used o aining he Mobili App p edic ion
model. Th ough hese and nume ous simila es s wi h a -
ious sma phones o new use s, we obse ed ha Mobili App
accu a ely p edic s he di e en anspo a ion modes used in
expe imen s in ol ing mul imodal ci y a el.
F om ex ensi e es ing, we concluded ha he de ec ion
sys em o S a iona y and Walk ac i i ies is obus and
pe o ms e ec i ely ac oss mos scena ios. An a ea o
imp o emen is iden i ied in high- a ic occasions wi h
equen s ops, whe e he sys em some imes inco ec ly
egis e s he S a iona y ac i i y o semi-s a iona y ehicles
in a a ic jam. The emphasis o ou design on obus ness
and accu acy o e speed in de ec ion p oduces his beha io .
Ne e heless, we easily add essed his issue on he se e
wi h a sc ip ha de ec s semi-s a iona y pa e ns om
consecu i e ehicle p edic ions and classi ies hem as a ic
jam e en s. This app oach ensu es accu a e iden i ica ion o
a ic conges ion while main aining obus anspo mode
classi ica ion.
In conclusion, he sys ema ic e alua ion o he Mic o and
Mac o-s a e sys ems, along wi h he kine ic mo ion classi ie
ac oss a ious scena ios, highligh s Mobili App’s obus ness
as a ool o u ban planne s o analyze ci izens’ mobili y
habi s. The sys em demons a es esilience, pa icula ly in
accu a ely di e en ia ing be ween WALK, STATIONARY,
and MOVING s a es. None heless, challenges we e obse ed
in high- a ic scena ios, whe e b ie ehicle misclassi i-
ca ions as STATIONARY we e obse ed du ing ce ain
seconds. Howe e , he majo i y o ing scheme (see Fig. 7)
a e agg ega ing 100 seconds o da a (5 mic os a es o
20 seconds each), e ec i ely esol ed his issue. The
eal-wo ld mul imodal ip ials showcased s ong sys em
pe o mance o he mos equen ly used anspo modes,
e en hough mino misclassi ica ions we e occasionally
obse ed du ing he ini ial seconds o mo emen . These
indings emphasize he applica ion’s signi ican po en ial o
deploymen in eal-wo ld u ban en i onmen s.
16 VOLUME 13, 2025
G. C. Ibañez e al.: Mobili App: A DL-Based Tool o T anspo Mode De ec ion
FIGURE 14. Tes ing wi h a WALK →STATIONARY ansi ion in he accele ome e samples. The 20-second
Mic o-s a es a e shown in he le image, while he 100-second Mac o-s a es a e displayed in he igh image. To al
expe imen ime: 200 sec.
FIGURE 15. Tes ing wi h a STATIONARY →BUS ansi ion in he accele ome e samples. The 20-second Mic o-s a es
a e shown in he le image, while he 100-second Mac o-s a es a e displayed in he igh image. To al expe imen
ime: 200 sec.
FIGURE 16. Tes ing wi h a BUS →WALK ansi ion in he accele ome e samples. The 20-second Mic o-s a es a e
shown in he le image, while he 100-second Mac o-s a es a e displayed in he igh image.
A. MOBILITAPP USE CASE: MASSIVE MULTIMODAL TRIP
DATA COLLECTION CAMPAIGN ‘‘HOW DO YOU COME TO
THE CAMPUS?’’
Ou wo k ocuses on s udying u ban mobili y h ough he
analysis o ci izens’ mobili y habi s. To his end, he aim is
o he Mobili App ool o se e as a esou ce o esea che s
and public en i ies esponsible o u ban mobili y planning
and managemen . Mobili App allows use s o upload a mul i-
modal ip summa y o he se e included in a ligh ex CSV
ile (each below 300 KB). The Mobili App Mul imodal T ip
VOLUME 13, 2025 17
G. C. Ibañez e al.: Mobili App: A DL-Based Tool o T anspo Mode De ec ion
FIGURE 17. Mobili App es s ca ied ou wi h wo sma phones ha we e
no used du ing aining. Two mul imodal ips a e shown: Ca +Walk in
he image on he le ; Walk+Bus+Walk+T am+Walk in he image on he
igh .
FIGURE 18. Addi ional Mobili App es ing was pe o med wi h wo o he
sma phones ha we e no used du ing aining. Two mul imodal ips
a e shown: Walk+Bus+Walk in he image on he le ;
Walk+Bus+Walk+T am+Walk in he image on he igh .
da ase consis s o iles ha cap u e op ional use da a (such
as age ange and gende ) and eco d de ailed in o ma ion
abou mul imodal ips, including p edic ed anspo modes
o each ip segmen and ansi ion loca ion poin s (la i ude
and longi ude) be ween di e en modes. Gi en he sensi i i y
o his use da a and he associa ed p i acy isks, he da ase
canno be publicly eleased in i s cu en o m. Howe e ,
as men ioned in Sec ion VII, we a e ac i ely esea ching
p i acy-p ese ing echniques ha could be implemen ed o
add ess his issue.
In Fig. 19, an example o a mul imodal ip (2nd and 3 d
ows) is shown, iden i ied wi h he same andom ID (in he
blue ec angle). The ip belongs o a woman aged 45-59, who
i s walked and hen used he subway o he second sec ion.
The ans e poin ( om walking o subway) is ma ked in
he GPS loca ion highligh ed by a g een ec angle, wi h he
ansi ion ime indica ed by he ed ec angle.
The da a om hese mul imodal cap u es a e in aluable
o s udying mobili y lows in any ci y, as exempli ied by
Ba celona in ou case s udy. Two en i ies showing pa icula
in e es in his p ojec a e UPC Sos enible [5] and he
ATM [1]. UPC Sos enible is keen on analyzing popula
anspo combina ions among use s a each UPC campus
(mo e in o ma ion in he UPC ci izen science po al [7]).
Meanwhile, ATM ocuses on imp o ing public anspo
in he Ba celona me opoli an a e. Mobili App acili a es a
mo e comp ehensi e and as e analysis o daily mobili y
pa e ns compa ed o adi ional me hodologies, e.g. su eys
o passenge s abou o igin and end o he jou ney.
Thus, once he Mobili App p edic ion model was ali-
da ed, we decided o ca y ou a Massi e Mul imodal T a el
Da a Collec ion Campaign wi hin he communi y o he Uni-
e si a Poli ècnica de Ca alunya (UPC) o compile a da ase
o mul imodal ips, composed o consecu i e segmen s
p edic ed using he Mobili App ool. Addi ionally, many
o he campaign olun ee s also collabo a ed by gene a ing
unimodal labeled ips o help enla ge he Mobili App Senso
Da ase [8], see Fig. 2.
In collabo a ion wi h UPC Sos enible [5] and he ATM,
we launched a 1-week da a collec ion ini ia i e in he UPC
wi h he pilo da a collec ion campaign ‘‘How do you come
o he Campus?’’2[30], wi hin he amewo k o he open and
ci izen science ini ia i es o he UPC [7].
This campaign included he pa icipa ion o olun ee s
om di e en g oups (s uden s, p o esso s, esea che s
and adminis a i e s a ), who con ibu ed o es ing he
pe o mance o he Mobili App ool. They gene a ed mo e
han 200 mul imodal ips du ing hei ips o and om he
uni e si y, which we s o ed in he Mobili App Mul imodal
T ip Da ase o la e analysis o he mobili y habi s o
he UPC communi y. As men ioned abo e, his da ase
con ains sensi i e in o ma ion ha could po en ially e eal
use loca ions. To comply wi h da a p i acy egula ions
and p o ec use con iden iali y, we canno publish i in
i s cu en o m. Howe e , we a e ac i ely in es iga ing
ajec o y anonymiza ion echniques ha may enable he
secu e sha ing o he da ase in he u u e. In he mean ime,
we can add ess po en ial p i acy conce ns by publishing
agg ega ed da a ins ead o aw da ase s, ensu ing ha
2Mobili y da a collec ion campaign ‘‘How do you come o he Campus?’’
15 h-19 h Ap il 2023.
18 VOLUME 13, 2025
G. C. Ibañez e al.: Mobili App: A DL-Based Tool o T anspo Mode De ec ion
FIGURE 19. Example o mul imodal ips collec ed wi h Mobili App. Gende and Age ange a e asked jus o s a is ical pu poses.
FIGURE 20. Sc eensho o he dashboa d wi h mul imodal ips collec ed wi h he applica ion. The Mobili App Powe Bi dynamic ool is a ailable
online [3].
FIGURE 21. Ou come o he Mobili App Powe Bi dynamic ool [3] il e ed acco ding o he selec ion shown in
Fig. 20 o women, (le image) and o men ( igh image). Examples o wha could be analyzed wi h he mobili y
in o ma ion collec ed wi h ou Mobili App ool.
indi idual use in o ma ion emains p o ec ed. To achie e
his, we ha e de eloped a Mobili App Powe Bi dynamic
ool [3] o analyse mul imodal ips. This dashboa d imple-
men ed in Powe Bi helps us o analyze he mul imodal
ip da a collec ed by he Mobili App ool du ing he
campaign.
Fig. 20 shows a sc eensho o he dashboa d, which
shows a selec ion o 21 ou o he 207 mobili y segmen s.
VOLUME 13, 2025 19
G. C. Ibañez e al.: Mobili App: A DL-Based Tool o T anspo Mode De ec ion
FIGURE 22. Ou come o he Mobili App Powe Bi dynamic ool [3] il e ed showing all agg ega ed
ips o he Les Co s dis ic (whe e he UPC is loca ed).
A mul imodal ip consis s o one o mo e mobili y
segmen s, so ha each phase o he ip co esponds o
a di e en mode o anspo . In he igu e we ha e
il e ed he ips o ma ch he op ions ma ked in black
(gende =woman; age ange =18-29, 45-59; anspo
mode =Bus, Me o, Ca , Walk, o Bicycle; School =all;
p o iles =p o esso s, esea che s, s uden s). The il e ed
esul s, depic ed in Fig. 21, e eal walking as he mos
p e e ed mode o anspo a ion. Addi ionally, gende -based
di e ences in anspo a ion choices we e obse ed: women
ended o ely mo e equen ly on ca s, while men showed
a g ea e inclina ion owa d using buses and subways du ing
hei mul imodal ips. In addi ion, he ool shows he il e ed
ips added o a map, see Fig. 22, in his case om all he
Dis ic s o Ba celona o he Les Co s Dis ic (whe e he
UPC is loca ed).
These a e jus a ew examples o wha can be analyzed
wi h he mul imodal a el in o ma ion collec ed wi h ou
Mobili App ool. The Mobili App Powe Bi dynamic ool can
be accessed in his link [3].
VII. CONCLUSION AND FUTURE WORK
In his s udy, we in oduced Mobili App, a deep lea ning-
based ool o eal- ime anspo a ion mode de ec ion using
sma phone senso da a, designed wi h a s ong emphasis on
minimizing compu a ional cos s. This op imiza ion is c ucial,
as he ool is in ended o ope a e on sma phones wi h limi ed
p ocessing powe . Mobili App p o ides aluable insigh s o
u ban planne s, enabling da a-d i en s a egies o imp o e
mobili y and suppo he ansi ion o mo e sus ainable
ci ies. Ou hie a chical p edic ion model combines CNNs o
ea u e ex ac ion and LSTM laye s o empo al p ocessing,
enhanced wi h skip connec ions o imp o e pe o mance.
To u he op imize compu a ional e iciency on mobile
de ices, a kinema ic mo ion classi ie is in eg a ed o ea ly
mo ion de ec ion, educing dependency on deep lea ning
models.
This esea ch ep esen s a ho ough in es iga ion in o
anspo mode and human ac i i y ecogni ion, using he
ci y o Ba celona as a case s udy o explo e he dynamics
o u ban mobili y. I b idges adi ional analy ical me hods
and ad anced deep lea ning echniques, o e ing a de ailed
pe spec i e on he s a e-o - he-a solu ions in his domain.
Th ough igo ous examina ion, he s udy highligh s he
ans o ma i e po en ial o deep lea ning models in sol ing
complex u ban mobili y challenges. Key indings om his
wo k no only illus a e he in ica e na u e o anspo mode
de ec ion bu also lay ou di ec ions o u u e explo a ion,
pushing he bounda ies o u ban mobili y esea ch.
This esea ch demons a es ha deep lea ning p o ides
a p omising amewo k o accu a ely de ec ing and dis in-
guishing anspo modes, pa icula ly when u ilizing empo-
al da a collec ed om a ious senso s. The implemen a ion
o hese echniques sheds ligh on he complexi ies o
managing and p e-p ocessing senso y da a, emphasizing he
need o obus da a-handling models o ensu e eliable and
high-quali y ou comes. Impo an ly, he di e en ia ion o
use g oups be ween aining and e alua ion phases is c i ical
o simula e eal-wo ld condi ions and e alua e he model
unde ci cums ances simila o i s in ended deploymen .
Addi ionally, i has been shown ha combining ecu en
a chi ec u es such as LSTM wi h con olu ional ne wo ks
helps o achie e a obus anspo mode classi ica ion model
using accele ome e , magne ome e , and gy oscope da a.
A e comple ing a ious da a collec ion campaigns, we ha e
gene a ed he Mobili App Senso Da ase [8], which is
20 VOLUME 13, 2025
G. C. Ibañez e al.: Mobili App: A DL-Based Tool o T anspo Mode De ec ion
publicly a ailable. Finally, he adop ion o a mo e complex
hie a chical a chi ec u e has p o en e ec i e in imp o ing
he gene aliza ion capabili ies o he model, leading o be e
pe o mance and mo e accu a e esul s.
The de eloped DL model success ully p edic s mos
anspo modes wi h an accu acy exceeding 80%, achie ing
a weigh ed a e age accu acy o 88%. These p omising
esul s unde sco e he u ili y o ou app oach o p edic ing
use s’ anspo modes du ing hei a els. This capabili y
is pa icula ly aluable o u ban planne s and public se ice
p o ide s, as he Mobili App ool enables de ailed analysis o
ci izens’ mobili y habi s, con ibu ing o he ad ancemen o
mo e sus ainable u ban mobili y solu ions.
A signi ican achie emen o his s udy is he de elopmen
o an And oid applica ion named Mobili App [4] ha
combines deep lea ning wi h s a is ical me hods, p o ing
obus ness o he anspo mode ecogni ion sys em. This
ool, capable o gene a ing mul imodal ip summa ies,
ep esen s a leap o wa d in unde s anding u ban mobili y
pa e ns, o e ing aluable insigh s o u ban planne s.
The Mobili App ool enables he collec ion o mobili y
da a, which is s o ed in wo da ase s:
•The Mobili App Senso Da ase , see Sec ion III-D
and Fig. 2. This da ase consis s solely o alues
co esponding o he X, Y, and Z axes o h ee mo ion
senso s: Accele ome e , Gy oscope, and Magne ome e .
I does no con ain any pe sonal in o ma ion and
is publicly a ailable a [8].This da ase is used o
ain ou DL-based p edic ion model o de e mine he
anspo a ion mode used by he ci izen.
•The Mobili App Mul imodal T ip Da ase , see Sec-
ion VI-A and Fig. 19. This da ase includes pe sonal and
sensi i e use in o ma ion, such as age ange, gende ,
and he GPS loca ion o each o igin-des ina ion pai o
each segmen o he ip (one segmen pe p edic ed
anspo a ion mode). This da ase includes sensi i e
loca ion da a, so i canno be published in i s cu en
o m due o p i acy egula ions. We a e explo ing
anonymiza ion echniques o secu e sha ing in he
u u e. Meanwhile, only agg ega ed da a will be eleased
o p o ec use con iden iali y. This da ase is designed
o assis mobili y planne s in unde s anding use s’
mobili y habi s and iden i ying po en ial s a egies o
enhance u ban mobili y, os e ing a ansi ion owa d
mo e sus ainable anspo a ion solu ions.
In conclusion, his wo k con ibu es o he exis ing body
o knowledge in he a ea o anspo mode ecogni ion using
deep lea ning and pa es he way o u u e s udies ha will
enhance and expand hese sys ems’ capabili ies. The po en ial
impac o his esea ch ex ends beyond academia, o e ing
p ac ical solu ions and insigh s ha could shape he u u e
o u ban mobili y in ci ies like Ba celona. Mo eo e , he
me hodology de eloped in his wo k aims o con ibu e o
mo e sus ainable u ban mobili y in Ba celona and can be
adap ed o o he ci ies as well. The Mobili App ool can
assis u ban planne s and public se ice p o ide s o be e
unde s and he mobili y habi s o ci izens, assess he impac
o mobili y imp o emen ini ia i es, and enhance public
anspo a ion se ices.
Fo u u e wo k, he eam is encou aged o collec
addi ional da a o unde ep esen ed anspo modes, such
as e-scoo e , Mo o bike, Bike, and e-Bike. In his ega d,
we plan o conduc addi ional da a-ga he ing campaigns
a se e al uni e si ies ha ha e in i ed us o collabo a e
du ing he Eu opean Sus ainable Mobili y Week. The goal
is o aise awa eness among he uni e si y communi y abou
adop ing sus ainable anspo a ion o hei daily commu es
o campus.
Mo eo e , addi ional da a ebalancing echniques, such as
SMOTE [12] and ADASYN [21], will be explo ed o imp o e
he ep esen a ion o hese unde ep esen ed classes wi h a
limi ed numbe o ga he ed samples.
Addi ionally, p i acy conce ns and he sys em’s in e -
p e abili y mus be ca e ully add essed be o e making he
Mobili App Mul imodal T ip Da ase publicly a ailable.
In his ega d, we a e explo ing anonymiza ion echniques,
such as ajec o y ob usca ion and di e en ial p i acy,
o ensu e use p i acy and p o ec sensi i e in o ma ion,
which is a necessa y condi ion be o e making his da ase
publicly a ailable.
To e ec i ely mi iga e p i acy conce ns when he ajec-
o y da ase is public, among a a ie y o p i acy models
a ailable in he li e a u e one migh conside applying a
di e en ially p i a e da a publishing mechanism. Essen ially,
di e en ial p i acy (DP) ensu es ha he p esence o absence
o a single indi idual ajec o y in he da ase does no
signi ican ly a ec he o e all s a is ics. Tha is, he ype
o a acks DP p o ec s agains e e s o hose in which
he ad e sa y aims o de e mine whe he an indi idual
con ibu ed hei da a o he da ase . This is achie ed by
adding ca e ully calib a ed andom noise o he da a o que y
esponses [41].
In he con ex o deep lea ning a chi ec u e esea ch,
we aim o explo e he in eg a ion o T ans o me s in o
ou amewo k. Addi ionally, he inno a i e applica ion o
Gene a i e Ad e sa ial Ne wo ks (GANs) [20] o da a
augmen a ion p esen s new oppo uni ies o enhancing
model pe o mance and expanding esea ch possibili ies.
These s a egies ep esen he ongoing de elopmen o deep
lea ning in in e p e ing complica ed anspo da a, as hey
seek o o e come exis ing limi s and in es iga e no el ways.
On he one hand, T ans o me s a e a deep lea ning a chi-
ec u e ini ially de eloped o Na u al Language P ocessing
(NLP) asks, bu hei e sa ili y ex ends o ields like
compu e ision and audio. Models like GPT (Gene a i e P e-
ained T ans o me ) le e age T ans o me s o asks such
as ansla ion, ex classi ica ion, and language gene a ion.
T ans o me s excel a cap u ing global dependencies, scaling
o la ge da ase s, and adap ing o applica ions beyond ex .
On he o he hand, GANs a e la ely being used o da a
augmen a ion, pa icula ly in scena ios wi h limi ed da ase s.
GANs consis o a gene a o and a disc imina o ha compe e
VOLUME 13, 2025 21
G. C. Ibañez e al.: Mobili App: A DL-Based Tool o T anspo Mode De ec ion
in a minimax game, enabling he c ea ion o syn he ic da a
ha closely esembles he o iginal da ase . This app oach
enhances model pe o mance by inc easing da a di e si y,
add essing class imbalances, and imp o ing gene aliza ion.
GAN-based augmen a ion is pa icula ly e ec i e in appli-
ca ions like image syn hesis, speech gene a ion, and medical
da a augmen a ion.
ACKNOWLEDGMENT
The au ho s would like o hank Miquel Go aneg a Es añol
and Jaume Planas i Planas o hei discussions and con i-
bu ions du ing hei deg ee and mas e ’s heses, espec i ely.
Addi ionally, hey ex end hei g a i ude o he olun ee s who
helped hem de elop he Mobili App ool by sha ing hei
mobili y da a. Finally, hey also would like o hank Albe
Villa oya Saladié om UPC Sos enible [5] and F ancesc
Cal e om he ATM [1] o hei collabo a ion in o ganizing
he da a collec ion campaign ‘‘How do you come o he
Campus?’’.
ACRONYMS
ADASYN Adap i e Syn he ic Sampling.
ATM Au o i a del T anspo Me opoli à.
CCN Con olu ional Neu al Ne wo ks.
DCBL Deep Con olu ional Bidi ec ional LSTM.
DNN Deep Neu al Ne wo k.
DP Di e en ial P i acy.
EWMA Exponen ially Weigh ed Mo ing A e age.
FFT Fas Fou ie T ans o m.
HAR Human Ac i i y Recogni ion.
HDC Hype Dimensional Compu ing.
LSTM Long Sho -Te m Memo y.
ML Machine Lea ning.
MD Mahalanobis dis ance.
PCA P incipal Componen Analysis.
PSD Powe Spec al Densi y.
SHL Sussex-Huawei Locomo ion Da ase .
SMOTE Syn he ic Mino i y O e -sampling Technique.
TMD T anspo Mode De ec ion Da ase .
UPC Uni e si a Poli ècnica de Ca alunya.
REFERENCES
[1] P o . Mónica Aguila Iga ua (UPC) and M . F ancesc Cal e (ATM).
P edic ion o he T anspo a ion Mode Used By he Ci izens in Thei
Mul imodal T ips F om he Analysis o Sma phone Senso s. Collabo-
a ion ag eemen ATM. Accessed: Ap . 17, 2025. [Online]. A ailable:
h ps:// u u .upc.edu/35021523
[2] Mobili App Applica ion Tha P edic s he T anspo a ion Mode Used.
Accessed: Ap . 17, 2025. [Online]. A ailable: h ps://play.google.com/
s o e/apps/de ails?id=com.mobi.mobili app&pcampaignid=web_sha e
[3] Mobili App Powe Bi Dynamic Tool o Analyse Mul imodal T ips.
Accessed: Ap . 17, 2025. [Online]. A ailable: h ps://aco a .link/o0Eszw
[4] Mobili App: Tool o Help Analyze he Mobili y Flow o Ci izens in he
Me opoli an A ea o Ba celona. Accessed: Ap . 17, 2025. [Online].
A ailable: h ps://mobili app.upc.edu
[5] Gabine D’Inno ació I Comuni a De La Uni e si a Poli ècnica De
Ca alunya (UPC). Accessed: Ap . 17, 2025. [Online]. A ailable:
h ps://sos enible.upc.edu
[6] (2021). G oupKFold. [Online]. A ailable: h ps://sciki -lea n.o g/s able/
modules/gene a ed/sklea n.model_selec ion.G oupKFold.h ml
[7] (2023). UPC Ci izen Science Po al. [Online]. A ailable:
h ps://en.cienciaciu adana.upc.edu
[8] M. A. Iga ua, G. C. Ibáñez, and M. G. Es añol, ‘‘Mobili App—Mobili y
senso da ase ,’’ CORA Reposi o i de Dades de Rece ca, Ba celona, Spain,
Dec. 2024, doi: 10.34810/da a1948. Accessed: Ap . 17, 2025.
[9] R. Alizadeh, Y. Sa a ia, and C. Ne guizian, ‘‘Cha ac e iza ion and
selec ion o WiFi channel s a e in o ma ion ea u es o human ac i i y
de ec ion in a sma public anspo a ion sys em,’’ IEEE Open J. In ell.
T ansp. Sys ., ol. 5, pp. 55–69, 2024, doi: 10.1109/OJITS.2023.3336795.
[10] And oid De elope s. (2023). Je pack Compose. Accessed: Sep. 19, 2023.
[Online]. A ailable: h ps://de elope .and oid.com/je pack/compose
[11] C. Ca pine i, V. Lomonaco, L. Bedogni, M. D. Felice, and L. Bononi,
‘‘Cus om dual anspo a ion mode de ec ion by sma phone de ices
exploi ing senso di e si y,’’ in P oc. IEEE In . Con . Pe asi e Compu .
Commun. Wo kshops (Pe Com Wo kshops), Ma . 2018, pp. 367–372, doi:
10.1109/PERCOMW.2018.8480119.
[12] N. V. Chawla, K. W. Bowye , L. O. Hall, and W. P. Kegelmeye , ‘‘SMOTE:
Syn he ic mino i y o e -sampling echnique,’’ J. A i . In ell. Res., ol. 16,
pp. 321–357, Jun. 2002, doi: 10.1613/jai .953.
[13] C. Co es, M. Moh i, and A. Ros amizadeh, ‘‘L2 egula iza ion o lea ning
ke nels,’’ 2012, a Xi :1205.2653.
[14] P. Cunningham and S. J. Delany, ‘‘K-nea es neighbou
classi ie s—A u o ial,’’ ACM Compu . Su ., ol. 54, no. 6, p. 128,
Ap . 2007, doi: 10.1145/3459665.
[15] R. Da id, J. Duke, A. Jain, V. J. Reddi, N. Je ies, J. Li, N. K eege ,
I. Nappie , M. Na aj, S. Rege , R. Rhodes, T. Wang, and P. Wa den, ‘‘Ten-
so Flow li e mic o: Embedded machine lea ning on TinyML sys ems,’’
2020, a Xi :2010.08678.
[16] M. E delić, T. E delić, and T. Ca ić, ‘‘Da ase o mul imodal anspo
analy ics o sma phone use s–Collec y,’’ Da a B ie , ol. 50, Oc . 2023,
A . no. 109481, doi: 10.1016/j.dib.2023.109481.
[17] R. Fallahzadeh, Z. E. Asha i, P. Alinia, and H. Ghasemzadeh, ‘‘Pe son-
alized ac i i y ecogni ion using pa ially a ailable a ge da a,’’ IEEE
T ans. Mobile Compu ., ol. 22, no. 1, pp. 374–388, Jan. 2023, doi:
10.1109/TMC.2021.3071434.
[18] H. Gjo eski, J. Bizjak, M. Gjo eski, and M. Gams, ‘‘Compa ing deep
and classical machine lea ning me hods o human ac i i y ecogni ion
using w is accele ome e ,’’ in P oc. IJCAI Wo kshop Deep Lea n.
A i . In ell., 2016, pp. 1–7. [Online]. A ailable: h ps://api.seman ic
schola .o g/Co pusID:34272612
[19] H. Gjo eski, M. Cilibe o, L. Wang, F. J. O donez Mo ales, S. Mekki,
S. Valen in, and D. Roggen, ‘‘The Uni e si y o Sussex-Huawei
locomo ion and anspo a ion da ase o mul imodal analy ics wi h
mobile de ices,’’ IEEE Access, ol. 6, pp. 42592–42604, 2018, doi:
10.1109/ACCESS.2018.2858933.
[20] I. Good ellow, J. Pouge -Abadie, M. Mi za, B. Xu, D. Wa de-Fa ley,
S. Ozai , A. Cou ille, and Y. Bengio, ‘‘Gene a i e ad e sa ial
ne wo ks,’’ in P oc. Ad . Neu al In . P ocess. Sys . (Neu IPS),
2014, pp. 2672–2680. [Online]. A ailable: h ps://pape s.nips.cc/
pape /2014/hash/5ca3e9b122 61 8 06494c97b1a cc 3-Abs ac .h ml
[21] H. He, Y. Bai, E. A. Ga cia, and S. Li, ‘‘ADASYN: Adap i e syn he ic
sampling app oach o imbalanced lea ning,’’ in P oc. IEEE In . Join
Con . Neu al Ne w. (IEEE Wo ld Cong . Compu . In ell.), Jun. 2008,
pp. 1322–1328, doi: 10.1109/IJCNN.2008.4633969.
[22] M. A. Hea s , S. T. Dumais, E. Osuna, J. Pla , and B. Scholkop , ‘‘Suppo
ec o machines,’’ IEEE In ell. Sys . Thei Appl., ol. 13, no. 4, pp. 18–28,
Jul. 1998, doi: 10.1109/5254.708428.
[23] S. Hemminki, P. Nu mi, and S. Ta koma, ‘‘Accele ome e -based
anspo a ion mode de ec ion on sma phones,’’ in P oc. 11 h ACM
Con . Embedded Ne wo ked Senso Sys ., No . 2013, pp. 1–14, doi:
10.1145/2517351.2517367.
[24] S. Hoch ei e and J. Schmidhube , ‘‘Long sho - e m memo y,’’
Neu al Compu ., ol. 9, no. 8, pp. 1735–1780, No . 1997, doi:
10.1162/neco.1997.9.8.1735.
[25] J. Iskande o and M. A. Gu ensan, ‘‘B eaking he limi s o anspo a ion
mode de ec ion: Applying deep lea ning app oach wi h knowledge-based
ea u es,’’ IEEE Senso s J., ol. 20, no. 21, pp. 12871–12884, No . 2020,
doi: 10.1109/JSEN.2020.3001803.
[26] J. V. Jeyakuma , E. S. Lee, Z. Xia, S. S. Sandha, N. Tausik, and
M. S i as a a, ‘‘Deep con olu ional bidi ec ional LSTM based anspo a-
ion mode ecogni ion,’’ in P oc. ACM In . Join Con . In . Symp. Pe asi e
Ubiqui ous Compu . Wea able Compu ., Oc . 2018, pp. 1606–1615, doi:
10.1145/3267305.3267529.
22 VOLUME 13, 2025
G. C. Ibañez e al.: Mobili App: A DL-Based Tool o T anspo Mode De ec ion
[27] D. P. Kingma and J. Ba, ‘‘Adam: A me hod o s ochas ic op imiza ion,’’
2014, a Xi :1412.6980.
[28] A. Mani annan, E. J. Willemse, B. T. Balamu ali, W. C. B. Chin,
Y. Zhou, B. Tunçe , A. Ba a , and R. Bou anais, ‘‘A amewo k o
he iden i ica ion o human e ical displacemen ac i i y based on mul i-
senso da a,’’ IEEE Senso s J., ol. 22, no. 8, pp. 8011–8029, Ap . 2022,
doi: 10.1109/JSEN.2022.3157806.
[29] G. J. McLachlan, ‘‘Mahalanobis dis ance,’’ Resonance, ol. 4, no. 6,
pp. 20–26, Jun. 1999, doi: 10.1007/b 02834632.
[30] M. Aguila , A. Villa oya, M. Go aneg a, G. Ca a aca, and M. De
Mingo, ‘‘Anàlisi de la campanya de ecollida de dades de mobili a a
la comuni a UPC amb l’eina Mobili App,’’ I em, Re is a De Biblio e-
conomia I Documen ació, ol. 2024, no. 76, pp. 1–22, Jul. 2024, doi:
10.60940/i emn76id430338.
[31] N. Nacha , ‘‘The mann-whi ney U: A es o assessing whe he
wo independen samples come om he same dis ibu ion,’’ Tu s.
Quan . Me hods Psychol., ol. 4, no. 1, pp. 13–20, Ma . 2008, doi:
10.20982/ qmp.04.1.p013.
[32] K. O’Shea and R. Nash, ‘‘An in oduc ion o con olu ional neu al
ne wo ks,’’ 2015, a Xi :1511.08458.
[33] L. R. Randle , J. B. Wansche , and M. Zabic, ‘‘Dis ibu ed a el mode
es ima ion,’’ in P oc. 19 h ITS Wo ld Cong ., Jan. 2012, pp. 1–6.
[34] S. Richoz, L. Wang, P. Bi ch, and D. Roggen, ‘‘T anspo a ion
mode ecogni ion using wea able mo ion, sound, and ision senso s,’’
IEEE Senso s J., ol. 20, no. 16, pp. 9314–9328, Aug. 2020, doi:
10.1109/JSEN.2020.2987306.
[35] L. Rokach and O. Maimon, ‘‘Decision ees,’’ in Da a Mining and
Knowledge Disco e y Handbook, ol. 6. Bos on, MA, USA: Sp inge ,
2005, pp. 165–192, doi: 10.1007/0-387-25465-X_9.
[36] S. Lee and C. Lee, ‘‘Re isi ing spa ial d opou o egula izing con o-
lu ional neu al ne wo ks,’’ Mul imedia Tools Appl., ol. 79, nos. 45–46,
pp. 34195–34207, Dec. 2020, doi: 10.1007/s11042-020-09054-7.
[37] Q. Tang, K. Jahan, and M. Ro h, ‘‘Deep CNN-BiLSTM model o
anspo a ion mode de ec ion using sma phone accele ome e and
magne ome e ,’’ in P oc. IEEE In ell. Vehicles Symp. (IV), Jun. 2022,
pp. 772–778, doi: 10.1109/IV51971.2022.9827275.
[38] L. Wang, H. Gjo eski, M. Cilibe o, S. Mekki, S. Valen in, and
D. Roggen, ‘‘Enabling ep oducible esea ch in senso -based anspo a-
ion mode ecogni ion wi h he sussex-huawei da ase ,’’ IEEE Access,
ol. 7, pp. 10870–10891, 2019, doi: 10.1109/ACCESS.2019.2890793.
[39] B. Xu, N. Wang, T. Chen, and M. Li, ‘‘Empi ical e alua ion o ec i ied
ac i a ions in con olu ional ne wo k,’’ 2015, a Xi :1505.00853.
[40] Y. Yao, W. Liu, G. Zhang, and W. Hu, ‘‘Rada -based human ac i i y
ecogni ion using hype dimensional compu ing,’’ IEEE T ans. Mic ow.
Theo y Techn., ol. 70, no. 3, pp. 1605–1619, Ma . 2022, doi:
10.1109/TMTT.2021.3134992.
[41] À. Mi anda-Pascual, P. Gue a-Balboa, J. Pa a-A nau, J. Fo né, and
T. S u e, ‘‘An o e iew o p oposals owa ds he p i acy-p ese ing
publica ion o ajec o y da a,’’ In . J. In . Secu ., ol. 23, no. 6,
pp. 3711–3747, Dec. 2024, doi: 10.1007/s10207-024-00894-0.
GERARD CARAVACA IBAÑEZ ecei ed he M.S.
deg ee in a i icial in elligence enginee ing om
he Uni e si a Poli ècnica de Ca alunya (UPC),
Ba celona, Spain, in 2024. His expe ise ex ends
ac oss compu e ision, na u al language p ocess-
ing, mobili y, and a spec um o machine lea ning
challenges. His esea ch in e es includes deep
lea ning models o p edic ing mobili y ac i i ies.
LUIS J. DE LA CRUZ LLOPIS (Membe , IEEE)
ecei ed he G adua e and Ph.D. deg ees in
elecommunica ions enginee ing om he Uni e -
si a Poli ècnica de Ca alunya (UPC), Ba celona,
Spain, in 1994 and 1999, espec i ely. He is cu -
en ly an Associa e P o esso wi h he Depa men
o Ne wo k Enginee ing, UPC. He is also pa o
he Sma Se ices o In o ma ion Sys ems and
Communica ion Ne wo ks (SISCOM) Resea ch
G oup. His cu en esea ch in e es s include he
applica ion o machine lea ning echniques in wi eless mul i-hop and C-V2X
ne wo ks and also he de elopmen o IoT sma se ices.
ADRIAN CATALÍN DIACONEASA ecei ed he
deg ee in elecommunica ions enginee ing om
he Uni e si a Poli ècnica de Ca alunya (UPC),
mino in elema ic sys ems. He has expe ience
wi h a ious so wa e de elopmen ools ob ained
h ough pa icipa ion in mul idisciplina y p ojec s.
His esea ch in e es includes de eloping deep
lea ning models o p edic mobili y pa e ns and
ac i i ies.
ALBERTO BAZÁN GUILLÉN ecei ed he Engi-
nee ing deg ee in elecommunica ions and elec-
onics and he M.Sc. deg ee in elema ics om
he Cen al Uni e si y ‘‘Ma a Ab eu’’ o Las
Villas (UCLV), San a Cla a, Cuba, in 2020 and
2022, espec i ely. He is cu en ly pu suing he
Ph.D. deg ee in elema ics enginee ing wi h he
SISCOM Resea ch G oup, Uni e si a Poli ècnica
de Ca alunya (UPC). His esea ch in e es s include
wi eless ne wo ks, ehicula ne wo ks, elec ical
ehicles, u ban mobili y, he IoT, and ede a ed lea ning models.
MÓNICA AGUILAR IGARTUA (Membe , IEEE)
ecei ed he M.Sc. and Ph.D. deg ees in elecom-
munica ion enginee ing om he Uni e si a
Poli ècnica de Ca alunya, Ba celona, Spain, in
1995 and 2000, espec i ely. She is he au ho o
mo e han 30 jou nal a icles, has chai ed se e al
con e ences and is a membe o he Edi o ial
Boa d o he Ad Hoc Ne wo ks jou nal. He
esea ch in e es s include design and pe o mance
e alua ion o ou ing p o ocols o ehicula
ne wo ks, elec ic ehicle, pla ooning, machine lea ning, sma ci y se ices,
and sus ainable u ban mobili y. She belongs o he SISCOM Resea ch G oup,
ocused on sma se ices o in o ma ion sys ems and communica ion
ne wo ks.
VOLUME 13, 2025 23
