Neu al Con olle o PTZ came as based on nonpano amic o eg ound de ec ion
Miguel A. Molina-Cabello∗, Ezequiel L´
opez-Rubio∗, Ra ael Ma cos Luque-Baena∗,
En ique Dom´
ınguez∗and Ka l Thu nho e -Hemsi∗
∗Depa men o Compu e Languages and Compu e Science
Uni e si y o M´
alaga, Bule a Louis Pas eu , 35, 29071 M´
alaga, Spain
Emails: {miguelangel,ezeql , mluque,en iqued,ka lkhade }@lcc.uma.es
Abs ac —In his pape a con olle o PTZ came as based
on an unsupe ised neu al ne wo k model is p esen ed. I
akes ad an age o he o eg ound mask gene a ed by a non-
pa ame ic o eg ound de ec ion subsys em. Thus, ou aim is
o op imize he mo emen s o he PTZ came a o a ain he
maximum co e age o he obse ed scene in p esence o mo ing
objec s. A g owing neu al gas (GNG) is applied o enhance
he ep esen a ion o he o eg ound objec s. Bo h quali a i e
and quan i a i e esul s a e epo ed using se e al widely used
da ase s, which demons a e he sui abili y o ou app oach.
1. In oduc ion
Mos o he o me su eillance sys ems we e buil wi h
a single s a iona y came a o many yea s. Howe e , nowa-
days i is possible o ind di e en ypes o came as and
any su eillance sys em is equen ly composed by mul iple
de ices which y o co e he la ges possible a ea. [1].
Among o he ypes o c i e ia, wo o he mos used ypes
o came a a e he omnidi ec ional and he pan- il -zoom
(PTZ) came as. Con en ional su eillance sys ems usually
comp ise a leas one omnidi ec ional came a and one PTZ
came a.
Su eillance sys ems a e capable o moni o ing he en-
i e scene by using a single omnidi ec ional came a bu ,
due o he limi ed esolu ion o hese pano amic came as,
de ailed in o ma ion o he objec s migh no be acqui ed.
As a esul o ha , a PTZ came a is used o hose asks,
which equi e close-up iews a high esolu ion.
PTZ came as a e well sui ed o objec iden i ica ion
and ecogni ion in a - ield scenes. Howe e , he p ac ical
use o PTZ came as in eal wo ld scena ios is complica ed
due o se e al easons [2]. A con inuous online came a
calib a ion is needed since he absolu e pan, il and zoom
posi ional alues p o ided by he came a ac ua o s a e no
synch onized wi h he ideo s eam in mos cases. Mo eo e ,
some adap i e backg ound ep esen a ion becomes neces-
sa y o make a ge acking, since he scene backg ound is
con inuously changing due o he came a ope a ion [3].
Con en ional came a sys ems a e usually easily cus-
omizable and le use s deploy he sensing in as uc u es
acco ding o hei needs, by adjus ing he a ious came a
pa ame e s such as ield o iew, esolu ion, ope a ing mode
(nigh /day ision, indoo /ou doo ) [4]. In mos cases he
design o he in as uc u e o a su eillance sys em is pe -
o med manually, al hough he wide ange o con igu a ions
and pa ame e se ings leads o subop imal solu ions, which
imply an incomple e co e age o he moni o ed a ea o ,
con e sely, highe deploymen cos s o achie e a sa is ac o y
esul [5]. As a gene al ule, we can assume ha he goal
o a came a planne is o gua an ee he maximum co e age
o he obse ed space, minimizing occlusions and ob aining
he bes isibili y o he objec s o in e es [6].
In his pape , we p opose a me hod o PTZ came as
based on G owing Neu al Gas (GNG) models in o de
o au oma ically de e mine he posi ion and pan- il -zoom
se ings o op imize he co e age o he o eg ound.
T adi ional o eg ound algo i hms iden i y o eg ound
pixels because hei ea u es a e di e en om hose o he
backg ound, bu his leads o alse de ec ion o mo ing
came as. Apa o o he p oposals based on building a
pano amic model o he scene, we a e ocusing on non-
pano amic me hods [7], [8] which a e sui able o ee
mo ing came as. The use o he neu al app oach il e s he
noise and spu ious objec s ob ained in he o eg ound mask.
Fu he mo e he mo ing objec s in he scene a e ep esen ed
wi h highe accu acy and obus ness. These a e he key
elemen s o design an e ec i e PTZ con olle .
GNG models a e a ype o sel -o ganizing neu al ne -
wo ks and one o he mos success ul example o unsu-
pe ised lea ning in a g aph. The o iginal GNG model was
p oposed by F i zke [9] and has become a s anda d o appli-
ca ions in compu e ision [10] and obo ics [11], as well as
o he sel -o ganizing models o o eg ound de ec ion [12]
o objec acking [13] in ideo sequences.
The es o he pape is o ganized as ollows. In sec ion
2, he p oposed con ol sys em o PTZ came as is p esen ed.
In sec ion 3, we epo he esul s achie ed wi h he p oposed
neu al con olle . Finally, sec ion 4 includes some conclud-
ing ema ks.
2. Sys em a chi ec u e
In his sec ion he a chi ec u e o he p oposed PTZ
came a con ol sys em is desc ibed. The sys em is made
o h ee modules, namely a o eg ound de ec ion p ocedu e
(Subsec ion 2.1), an unsupe ised lea ning model o lea n
978-1-5090-6182-2/17/$31.00 ©2017 IEEE 404
he dis ibu ion o he o eg ound objec s (Subsec ion 2.2),
and a con ol module o mo e he came a (Subsec ion 2.3).
2.1. Fo eg ound de ec ion
The i s ask o be accomplished is he de ec ion o he
pixels which belong o he o eg ound. This means ha a
bina y mask mus be compu ed o each incoming ideo
ame, so ha he o eg ound pixels a e ma ked as ue in
ha mask. In o de o do his, we ha e used ou p e ious
backg ound model o mo ing came as [8]. The eade is
kindly di ec ed o he e e ence o mo e de ails. A each
ime ins an , he ou pu o his algo i hm is a bina y lag
i,j ∈ { ue, alse} o each pixel, whe e (i, j)a e he
pixel coo dina es. F om his a aining se is buil , which
comp ises he coo dina es o all o eg ound pixels:
S ={(i, j)| i,j = ue} ⊂ R2(1)
This aining se is p o ided o a modi ied e sion o he
GNG neu al model, as speci ied nex .
2.2. Neu al model
In o de o loca e he mos ele an o eg ound objec s in
he scene, we p opose o ain a GNG [9] online by episodes,
so ha each episode is an incoming ideo ame. The GNG
ea u es a a iable numbe o neu ons H, which a e inse ed
and emo ed om he ne wo k as he lea ning p ocess is
execu ed. The neu ons a e connec ed by undi ec ed links,
so ha he esul ing g aph migh ha e se e al connec ed
componen s. We modi y he GNG o p ocess he incoming
inpu da a by episodes, so ha a episode a new aining
se S is p esen ed, which is made o D-dimensional eal
alued ec o s, S ⊂RD. As seen in Subsec ion 2.1, o
ou applica ion D= 2. Each neu on i∈ {1, ..., H}has an
associa ed cen oid wi∈RD, an age (which is a na u al
numbe ) and an e o a iable ei∈R,ei≥0. Each
connec ion also has an age. We will no e A he se o all
connec ions, A⊆ {1, ..., H}×{1, ..., H}.
The lea ning algo i hm is gi en by he ollowing s eps:
1) A he ini ial episode = 0 s a wi h wo neu ons
(H= 2) joined by a connec ion. Each p o o ype is
ini ialized o a sample d awn a andom om S0.
The e o a iables a e ini ialized o ze o. The age
o he connec ion and he neu ons a e ini ialized o
ze o, oo.
2) D aw a aining sample x∈RDa andom om
om S .
3) Find he nea es neu on qand he second nea es
neu on sin e ms o Euclidean dis ance:
q= a g min
i∈{1,...,H}kwi−xk(2)
s= a g min
i∈{1,...,H}−{q}kwi−xk(3)
4) Inc emen he age o all edges depa ing om q.
5) Add he squa ed Euclidean dis ance be ween x
and he nea es neu on q o he e o a iable eq:
∆eq=kwq−xk2(4)
6) Upda e qand all i s di ec opological neighbo s
wi h s ep size b o neu on qand n o he
neighbo s, whe e b> n:
(i) = 




bi i=q
ni (i6=q)∧(i, q)∈A
0 i (i6=q)∧(i, q)/∈A
(5)
∆wi=(i) (x−wi)(6)
7) I qand sa e connec ed by an edge, hen se he
age o his edge o ze o. O he wise, c ea e i .
8) Se he age o qand s o ze o, and inc emen he
age o all he o he neu ons.
9) Remo e edges wi h an age la ge han amax. Then
emo e all neu ons which ha e no ou going edges,
and hose neu ons whose age is la ge han amax.
10) I λsamples ha e been p ocessed since he las
neu on c ea ion and he cu en numbe o neu ons
His lowe han he maximum Hmax, hen inse
a new neu on as ollows. Fi s de e mine he
neu on wi h he maximum e o and he neu on
zwi h he la ges e o among all di ec neighbo s
o . Then c ea e a new neu on k, inse edges
connec ing kwi h and z, and emo e he o iginal
edge be ween and z. A e ha , dec ease he e o
a iables e and ezby mul iplying hem wi h a
cons an α, and ini ialize he e o a iable ek o he
new alue o e . Finally, se up he p o o ype o k
o be hal way be ween hose o and z, as ollows:
wk=1
2(w +wz)(7)
11) Dec ease all e o a iables eiby mul iplying hem
by a cons an d.
12) Remo e all neu ons which ha e no won du ing he
las Nks eps, and hei connec ions.
13) I he maximum numbe o samples o be p ocessed
o he cu en episode has been eached, hen go
o s ep 13. O he wise, go o s ep 2.
14) I he las episode has been p ocessed, hen s op.
O he wise, inc emen he episode coun e , load
he nex episode and go o s ep 2.
2.3. Came a con ol
The las s ep o he sys em is he came a con ol module.
A each ime ins an , he se o connec ed componen s o
he di ec ed g aph associa ed o he GNG is compu ed. F om
he se o all connec ions A⊆ {1, ..., H}×{1, ..., H}, he
se o connec ed componen s S ∈2{1,...,H}is a pa i ion o
405
he se o all neu ons {1, ..., H}, so ha each se Si, ∈ S
comp ises neu ons which a e linked by a chain o connec-
ions in A. These connec ed componen s a e associa ed o
o eg ound objec s which appea in he scene a ime .
F om hem we choose he la ges one, i.e. he connec ed
componen associa ed o he la ges o eg ound objec . Then
he cen oid o ha componen is compu ed:
µ =1



ˆ
S 


X
i∈ˆ
S
wi(8)
whe e ˆ
S s ands o he la ges connec ed componen o S .
Finally, he came a is mo ed owa ds he cen oid µ .
In addi ion o his, we ha e assumed ha he size o he
connec ed componen is calcula ed like a ci cle, conside ing
he cen oid as he cen e o he ci cle and i s adius as he
mean dis ance om each neu on o he cen oid:
=1



ˆ
S 


X
i∈ˆ
S
kwi−µ k(9)
The e o e he size o he connec ed componen is es i-
ma ed as ollows:
ω =π 2
(10)
The came a con ol module, gi en he cen oid and
he size o he la ges connec ed componen , sends o he
came a he commands ha i mus execu e in o de o
ollow he acked o eg ound objec s. The commands he
module can send o he came a a e ho izon al, e ical and
zoom mo emen s. Wi h he ho izon al mo emen he came a
mo es o he le o o he igh ; he came a can mo e
upwa ds o downwa ds wi h he e ical mo emen ; and i
can apply zoom in o zoom ou wi h he zoom mo emen .
The amoun s o mo emen o each kind a e quan ized, so
ha he e is a minimum mo emen o size ρδ o each kind
o mo emen δ∈ {pan, il , zoom}. Fu he mo e, we ha e
o a oid o bidden mo emen s. Fo example, he came a
can no apply zoom inde ini ely because i has de ined a
maximum and a minimum zoom alue. On he o he hand,
no mo emen can be applied in each kind o mo emen .
Besides, he came a will do one mo emen pe each kind
o mo emen a each ime ins an o he ideo. Thus, i
we ep esen s he ho izon al, e ical and zoom posi ion
o he came a in hese ame as (α ,β ,γ )as sphe ical
coo dina es, in he ( +1)- h ame he posi ion o he came a
posi ion will be:
(σ +1,β +1,γ +1)=(σ ,β ,γ ) + (∆σ ,∆β ,∆γ )
(11)
whe e he a ia ion o each kind o mo emen δis δ ∈
{−ρδ,0, ρδ} o a decision o he con ol module o de-
c ease, s op o inc ease he posi ion o he came a in ha
kind o mo emen , espec i ely.
Fu he mo e, each kind o mo emen has a maximum
and a minimum possible alue as a physical limi o he
came a. Fo example, we canno apply zoom in o zoom
ou all we wan . These maximum and minimum alues a e
no ed Ψδand ψδ, espec i ely.
In o de o ca y ou as ew mo emen s as possible, we
ha e de ined some scena ios whe e no mo emen is applied
o he came a. So ha , when he ho izon al and e ical
coo dina es o he cen oid o he acked objec a e no a
om he e ical and ho izon al coo dina es o he cen e o
he ame, hen he con ol module does no issue a e ical
and ho izon al mo emen command, espec i ely. Mo eo e ,
no zoom mo emen is applied i he size (in pixels) o
he a ge objec is be ween a minimum and a maximum
alue o pe cen age espec o he o e all numbe o pixels
o he ame. So ha , o each kind o mo emen δwe
ha e speci ied a minimum and a maximum alue: φδand
Φδ, espec i ely. The alues o he ho izon al and e ical
mo emen s indica e he dis ance (in deg ees) om he cen e
o he ame o he cen oid; and he alues o he zoom
mo emen indica es he pe cen age o he numbe o pixels
o he ame which a e occupied by he a ge .
Finally, when no a ge is ound he con ol module
ac i ely ies o ind a a ge by mo ing he came a in o de
o a oid sys em s a es whe e he came a will ne e ind a
a ge .
3. Expe imen al esul s
In his sec ion we show he compu a ional expe imen s
we ha e ca ied ou and i s esul s. The so wa e and ha d-
wa e ha ha e been used a e epo ed in Subsec ion 3.1.
Then, he es ed ideo sequences a e speci ied in Subsec ion
3.2. The uned pa ame e s o he so wa e a e shown in
Subsec ion 3.3. Finally, he ob ained esul s om he ex-
pe imen s a e desc ibed in Subsec ion 3.4.
3.1. Me hods
We ha e used a nonpano amic o eg ound objec de ec-
ion algo i hm o ou con ol sys em. The me hod we ha e
employed is a ela ed wo k om ou esea ch g oup [8].
This me hod ha we no e nonpan is implemen ed in Ma lab
and i uses MEX iles w i en in C++ o he mos CPU
ime demanding pa s. The implemen a ion is a ailable on
i s websi e 1. On he o he hand, ou implemen a ion o he
GNG model is w i en in Ma lab.
The came a con ol module is based on he i ualp z
lib a y [14]. I simula es he ope a ion o a PTZ came a om
a 360-deg ee pano amic ideo, and i is accessible om i s
websi e 2. I s implemen a ion is w i en in C++ and uses he
OpenCV lib a y3.
The epo ed expe imen s ha e been ca ied ou on a 64-
bi Pe sonal Compu e wi h an eigh -co e In el i7 3.60 GHz
CPU, 32 GB RAM and s anda d ha dwa e. The implemen-
a ion o ou app oach does no use any GPU esou ces, so
i does no equi e any speci ic g aphics ha dwa e.
1. h p://www.lcc.uma.es/∼ezeql /nonpan/nonpan.h ml
2. h ps://bi bucke .o g/pie e luc s cha les/ i ualp z s andalone
3. h p://openc .o g/
406
TABLE 1. CONSIDERED PARAMETER VALUES.
Me hod Pa ame e s
Nonpan Fea u es, F={[19 20 22]}
S ep size, = 0.03
Th eshold, τ= 0.999
GNG model Max uni s, Hmax = 100
Lambda, λ= 100
Numbe o s eps, N= 20000
Epsilon B, b= 0.2
Epsilon N, n= 0.006
Alpha, α= 0.5
A max, amax = 50
D, d= 0.995
S eps o emo e non ac i e neu ons, Nk= 1000
Vi ualp z Va ia ion o ho izon al mo emen , ρσ= 2
Va ia ion o e ical mo emen , ρβ= 2
Va ia ion o zoom mo emen , ργ= 2
Minimum ho izon al limi , ψσ=−180
Maximum ho izon al limi , Ψσ= 180
Minimum e ical limi , ψβ= 0
Maximum e ical limi , Ψβ= 180
Minimum zoom limi , ψγ= 40
Maximum zoom limi , Ψγ= 140
Minimum ho izon al dis ance, φσ=−10
Maximum ho izon al dis ance, Φσ= 10
Minimum e ical dis ance, φβ=−20
Maximum e ical dis ance, Φβ= 20
Minimum zoom dis ance, φβ= 0.10
Maximum zoom dis ance, Φβ= 0.40
3.2. Sequences
Th ee ideos ha e been used in o de o ca y ou he
expe imen s, which a e a ailable on he i ualp z websi e.
The h ee sequences a e indoo scenes and hey a e named
scena io3,scena io4 and scena io5. The wo i s ones
co espond o he same loca ion (a spacious hall wi h people
walking on in) and hey a e e y simila , so we ha e jus
epo ed he esul s o he scena io3 sequence (3500x1750
pixels and 566 ames). On he o he hand, he scena io5
sequence (3500x1750 pixels and 1957 ames) shows a oom
wi h people mo ing on in and doing di e en ac ions.
3.3. Pa ame e selec ion
A se o uned alues o he pa ame e s o he me hods
ha e been de ined o ca y ou he expe imen s. These ixed
pa ame e s ha e been chosen om he ecommenda ions o
he GNG au ho s and ou expe ience. They a e epo ed in
Table 1. In pa icula , he alues ela ed o he PTZ came a
con ol ha e been chosen in o de o add ess he low ame
a e o he benchma k ideos.
3.4. Resul s
The ope a ion o ou p oposal wi h one o he selec ed
ideos is epo ed in Fig. 1. In his igu e, a cap u ed ame
wi h he PTZ came a is shown, along wi h he ob ained
esul a e he execu ion o he nonpan algo i hm, he s a e
o he GNG model in ha momen and he la ges connec ed
componen which is o be acked. This las image also
shows he cen oid o he la ges ( acked) componen .
A e wa ds, depending o he cen oid and he size o
he selec ed componen , he con ol module will indica e he
di e en s eps (ho izon al, e ical and zoom mo emen s) o
he came a.
The GNG s a e depends di ec ly om i s p e ious s a e
and he nonpan ou pu . The as e he mo emen o he
came a and he ac ions o he agen s, he less close he
GNG will be wi h espec o he desi ed one. This can be
be e app ecia ed in Fig. 2. The e olu ion o he GNG s a e
keeps connec ed componen s o neu ons e en hough hey
belong o di e en componen s. This e ec is s onge o
inc eased agen ac i i y. Ne e heless, he own na u e o he
GNG p o ides an e icien app oxima ion o he cen oid o
he a ge and a sui able app oxima ion o i s size.
Fu he mo e, in he wo i s ows i can be obse ed how
he esul and he e olu ion o he ideo can be di e en
depending o he GNG neu al ne wo k ini ializa ion. This
has an in luence on he con olle module and i s decision
could be di e en wi h he same ame. This no mally occu s
when he ame p esen s mo e han one pe son o a high
le el o noise in he ou pu p oduced by he nonpan.
We ha e chosen some ames as g ound u h o he pu -
pose o de e mining he pe o mance due o he sequences
no inco po a ing a g ound u h mask. Because o his we
ha e used a acking g ound u h added o he es ed ideos,
which con ains he in o ma ion abou he cen oid and he
bounding box o a pe son. Thus, he selec ed benchma k
ames p esen only one pe son in di e en si ua ions. The
commands issued by he con olle acco ding o he g ound
u h o he acked objec posi ion and he GNG es ima ed
posi ion a e epo ed.
The app oach has been un wi h he wo es ed ideos
and wo di e en ini ializa ions o he pu pose o ha ing
a wide ange o benchma k ames, so we ha e ou se-
quences. Wi h each one we ha e selec ed 25 andom ames
om he benchma k ame se . As i can be obse ed in Fig.
3 he quali a i e esul s o e ed by he GNG a e simila o
he g ound u h. In addi ion, he decision o he con ol
module is qui e simila o he g ound u h and he GNG
da a in each benchma k ame.
In o de o ge a quan i a i e poin o iew abou he
pe o mance o ou app oach we ha e selec ed he accu acy,
by compu ing he hi s o he decision and he a emp s. The
ob ained pe o mance is epo ed in Table 2. We can con-
side he mos impo an pe o mance in his es ed ideos
is he accu acy in he ho izon al mo emen and, o a lesse
ex en , he zoom mo emen . This is because he people a e
mo ing on om le o igh and ice e sa, u he away
o close , bu almos always a ound he came a. Acco ding
o hese esul s, mos o he mis akes in he decision o
he mo emen a e p oduced by he noise o he ob ained
esul o he nonpan me hod and he GNG s a e. This can
be obse ed in he las wo ows o Fig.3: he GNG has
some well-dis inguished pa s bu hey a e connec ed, so he
GNG conside s a highe a ge and i s cen oid is displaced,
p oducing a di e en mo emen decision om he g ound
407
TABLE 2. ACCURACY RESULTS. FIRST COLUMN CORRESPONDS TO A
BENCHMARK VIDEO (THE TWO TESTED SEQUENCES WITH TWO
DIFFERENT INITIALIZATION)AND REMAINING COLUMNS INDICATE THE
ACCURACY (HITS/ATTEMPTS)FOR THE HORIZONTAL,VERTICAL AND
ZOOM MOVEMENT CONSIDERING THE 25 SELECTED BENCHMARK
FRAMES FOR EACH VIDEO. EACH ROW SHOWS A VIDEO AND ITS
ACCURACY PERFORMANCES AND THE LAST ROW INDICATES THE
TOTAL ACCURACY CONSIDERING THE FOUR VIDEOS THAT THEY HAVE
BEEN CARRIED OUT.
Video Ho izon al Ve ical Zoom
Scena io3 (1) 16/25 (0.64) 22/25 (0.88) 22/25 (0.88)
Scena io3 (2) 23/25 (0.92) 18/25 (0.72) 19/25 (0.76)
Scena io5 (1) 11/25 (0.44) 12/25 (0.48) 11/25 (0.44)
Scena io5 (2) 17/25 (0.68) 18/25 (0.72) 7/25 (0.28)
To al 67/100 (0.67) 70/100 (0.70) 59/100 (0.59)
u h decision and i p oduces a lowe accu acy. Mo eo e ,
he ac ha opposi e e o s ( o example, he g ound u h
indica es a le mo emen and he GNG a igh one) ha e
only appea ed a ew imes mus also be highligh ed: 3 imes
o he 100 es ed ho izon al mo emen decisions and 2
imes o he 100 es ed zoom mo emen decisions.
4. Conclusion
A neu al con olle o PTZ came as, which is based
in a g owing neu al gas (GNG) app oach, was p esen ed
in o de o op imize he maximum co e age o he a ea o
he scene in p esence o o eg ound objec s. The objec s
in mo ion a e de ec ed using a nonpa ame ic o eg ound
de ec ion algo i hm which yields a o eg ound bina y mask.
The GNG model ep esen s he o eg ound mask wi h he
aim o a oiding noise and spu ious objec s, in addi ion o
p o ide highe obus ness in he came a con ol module. The
i ualp z lib a y was used o simula ed he pe o mance o a
eal PTZ came a. Se e al publicly a ailable ideo sequences
has been conside ed in ou s udy. In pa icula , some quan i-
a i e esul s a e ob ained in compa ison o he g ound u h
(pan, il o zoom mo emen in each ame) o he scene.
Wi hin his scheme some p omising esul s (a ound 65%
o accu acy) a e ob ained. I mus be aken in o accoun
ha pe o ming he exac same mo emen as he g ound
u h is no always necessa y o ob ain he bes co e age
o he scene. This means ha e alua ing he pe o mance
o his kind o sys ems is di icul . In la e wo ks, new
quan i a i e measu es should be p oposed o be e cap u e
he pa icula i ies o he con ol p ocess o a PTZ came a.
Acknowledgmen s
This wo k is pa ially suppo ed by he Minis y o Econ-
omy and Compe i i eness o Spain unde g an TIN2014-
53465-R, p ojec name Video su eillance by ac i e sea ch
o anomalous e en s. I is also pa ially suppo ed by he Au-
onomous Go e nmen o Andalusia (Spain) unde p ojec s
TIC-6213, p ojec name De elopmen o Sel -O ganizing
Neu al Ne wo ks o In o ma ion Technologies; and TIC-
657, p ojec name Sel -o ganizing sys ems and obus es-
ima o s o ideo su eillance. All o hem include unds
om he Eu opean Regional De elopmen Fund (ERDF).
The au ho s hank ully acknowledge he compu e esou ces,
echnical expe ise and assis ance p o ided by he SCBI (Su-
pe compu ing and Bioin o ma ics) cen e o he Uni e si y
o M´
alaga. They also g a e ully acknowledge he suppo o
NVIDIA Co po a ion wi h he dona ion o he Ti an X GPU
used o his esea ch.
Re e ences
[1] Y. Xu and D. Song, “Sys ems and algo i hms o au onomous and
scalable c owd su eillance using obo ic p z came as assis ed by a
wide-angle came a,” Au onomous Robo s, ol. 29, no. 1, pp. 53–66,
2010.
[2] G. Lisan i, I. Masi, F. Pe nici, and A. Del Bimbo, “Con inuous
localiza ion and mapping o a pan il zoom came a o wide a ea
acking,” Machine Vision and Applica ions, ol. 27, no. 7, pp. 1071–
1085, 2016.
[3] H. Sajid, S.-C. Cheung, and N. Jacobs, “Appea ance based back-
g ound sub ac ion o p z came as,” Signal P ocessing: Image Com-
munica ion, ol. 47, pp. 417–425, 2016.
[4] K. Konda, N. Conci, and F. De Na ale, “Global co e age maximiza-
ion in p z-came a ne wo ks based on isual quali y assessmen ,”
IEEE Senso s Jou nal, ol. 16, no. 16, pp. 6317–6332, 2016.
[5] E. H¨
o s e and R. Lienha , “On he op imal placemen o mul iple
isual senso s,” in P oceedings o he 4 h ACM In e na ional
Wo kshop on Video Su eillance and Senso Ne wo ks, se . VSSN
’06. New Yo k, NY, USA: ACM, 2006, pp. 111–120. [Online].
A ailable: h p://doi.acm.o g/10.1145/1178782.1178800
[6] Gil Whoan Chu and Myung Jin Chung, “Selec ion o an op imal
came a posi ion using isibili y and manipulabili y measu es o an
ac i e came a sys em,” in P oceedings In e na ional Con e ence on
In elligen Robo s and Sys ems (IROS 2000), ol. 1. IEEE, 2000,
pp. 429–434.
[7] S. Kim, K. Yun, K. Yi, S. Kim, and J. Choi, “De ec ion o mo ing
objec s wi h a mo ing came a using non-pano amic backg ound
model,” Machine Vision and Applica ions, ol. 24, no. 5, pp. 1015–
1028, 2013.
[8] F. J. L´
opez-Rubio and E. L´
opez-Rubio, “Fo eg ound de ec ion o
mo ing came as wi h s ochas ic app oxima ion,” Pa e n Recogni ion
Le e s, ol. 68, pp. 161–168, 2015.
[9] B. F i zke, “A g owing neu al gas ne wo k lea ns opologies,” in
Ad ances in Neu al In o ma ion P ocessing Sys ems 7. MIT P ess,
1995, pp. 625–632.
[10] E. J. Palomo and E. L´
opez-Rubio, “The g owing hie a chical neu al
gas sel -o ganizing neu al ne wo k,” IEEE T ansac ions on Neu al
Ne wo ks and Lea ning Sys ems, ol. PP, no. 99, pp. 1–10, 2016.
[11] P. M. Yanik, J. Manganelli, J. Me ino, A. L. Th ea , J. O. B ooks,
K. E. G een, and I. D. Walke , “A Ges u e Lea ning In e ace o
Simula ed Robo Pa h Shaping Wi h a Human Teache ,” IEEE T ans-
ac ions on Human-Machine Sys ems, ol. 44, no. 1, pp. 41–54, eb
2014.
[12] R. M. Luque-Baena, E. L´
opez-Rubio, E. Dom´
ınguez, E. J. Palomo,
and J. M. Je ez, “A sel -o ganizing map o imp o e ehicle de ec ion
in low moni o ing sys ems,” So Compu ing, ol. 19, no. 9, pp. 2499–
2509, 2015.
[13] R. M. Luque-Baena, J. M. O iz-de Lazcano-Loba o, E. L´
opez-Rubio,
E. Dom´
ınguez, and E. J. Palomo, “A compe i i e neu al ne wo k
o mul iple objec acking in ideo sequence analysis,” Neu al
P ocessing Le e s, ol. 37, no. 1, pp. 47–67, 2013.
[14] G. Chen, P.-L. S -Cha les, W. Bouachi , G.-A. Bilodeau, and
R. Be ge in, “Rep oducible e alua ion o pan- il -zoom acking,” in
Image P ocessing (ICIP), 2015 IEEE In e na ional Con e ence on.
IEEE, 2015, pp. 2055–2059.
408

F ame Nonpan GNG Selec ed componen
Figu e 1. G aphical desc ip ion o he ope a ion o he p oposed me hod. F om le o igh , he columns show a ame o a sequence, he bina y mask
p oduced by he nonpano amic o eg ound de ec ion me hod, he s a e o he neu al ne wo k model in ha momen ( ed ci cles o show he neu ons
and blue line segmen s o ep esen he connec ions among he neu ons), he selec ed componen (i.e. he la ges connec ed componen ) and i s cen oid
( ep esen ed wi h a g een as e isk). The i s and second ows show ame 78 o scena io3, each ow wi h a di e en GNG ini ializa ion in he i s ame.
409
Figu e 2. G aphical e olu ion o he GNG. Fi s and second columns co espond o he ames 176 o 180 o he ideo scena io3 and i s co esponding
GNG s a e. Thi d and ou h columns co espond o he ames 506 o 510 o he ideo scena io3 and i s co esponding GNG s a e.
410
F ame wi h acking GNG GT decision Con olle decision
Le mo e
No e ical mo e
Zoom in mo e
Le mo e
Down mo e
Zoom in mo e
Righ mo e
Down mo e
Zoom ou mo e
No ho izon al mo e
No e ical mo e
Zoom ou mo e
Righ mo e
No e ical mo e
No zoom mo e
Righ mo e
No e ical mo e
No zoom mo e
Righ mo e
Down mo e
No zoom mo e
Righ mo e
No e ical mo e
No zoom mo e
Le mo e
Down mo e
Zoom ou mo e
No ho izon al mo e
No e ical mo e
No zoom mo e
Righ mo e
No e ical mo e
No zoom mo e
No ho izon al mo e
No e ical mo e
Zoom ou mo e
Figu e 3. Resul s o some benchma k ames. Fi s column shows a ame wi h he cen oid and he bounding box o he g ound u h a ge , bo h colou ed
in yellow, and he cen oid and he size o he a ge indica ed by he con ol module, in g een. Second column ep esen s he s a e o he GNG a hese
momen . The las wo columns epo s he di ec ions gi en by he con olle o he g ound u h a ge and he a ge de ec ed by he GNG, espec i ely.
Fi s and second ows a e co esponding o he ames 58 and 510 o he ideo scena io3, and he hi d and ou h ows ep esen he ames 171 and
192 o he same sequence wi h a di e en ini ializa ion and g ound u h a ge . The i h and six h ows show he ames 165 and 249 o he sequence
scena io5.
411