Senso s 2015, 15, 19369-19392; doi:10.3390/s150819369
senso s
ISSN 1424-8220
www.mdpi.com/jou nal/senso s
A icle
Au oma ed T acking o D osophila Specimens
Rubén Chao 1, Ge mán Macía-Vázquez 1, Edua do Zalama 2, Jaime Gómez-Ga cía-Be mejo 2,*
and José-Ramón Pe án 3
1 Uni e si y o Valladolid, Paseo del Cauce 59. Valladolid 47011, Spain;
E-Mails: chaos. [email protected] (R.C.); g.macia. [email protected] (G.M.-V.)
2 Uni e si y o Valladolid, Ins i u o de las Tecnologías A anzadas de la P oducción,
Paseo del Cauce 59. Valladolid 47011, Spain; E-Mail: ezalam[email p o ec ed]
3 Fundación Ca i , Pa que Tecnológico de Boecillo, Valladolid 47151, Spain;
E-Mail: [email p o ec ed]
* Au ho o whom co espondence should be add essed; E-Mail: [email protected] a.es;
Tel.: +34-983-423-398; Fax: +34-983-423-310.
Academic Edi o : Vi o io M.N. Passa o
Recei ed: 7 Ap il 2015 / Accep ed: 27 July 2015 / Published: 6 Augus 2015
Abs ac : The ui ly D osophila Melanogas e has become a model o ganism in he s udy
o neu obiology and beha io pa e ns. The analysis o he way he ly mo es and i s
beha io is o g ea scien i ic in e es o esea ch on aspec s such as d ug ole ance,
agg ession o ageing in humans. In his a icle, a p ocedu e o de ec ing, iden i ying and
acking nume ous specimens o D osophila by means o compu e ision-based sensing
sys ems is p esen ed. This p ocedu e allows dynamic in o ma ion abou each specimen o be
collec ed a each momen , and hen o i s beha io o be quan i a i ely cha ac e ized. The
p oposed algo i hm ope a es in h ee main s eps: a p e-p ocessing s ep, a de ec ion and
segmen a ion s ep, and acking shape. The p e-p ocessing and segmen a ion s eps allow
some limi s o he image acquisi ion sys em and some isual a i ac s (such as shadows and
e lec ions) o be deal wi h. The imp o emen s in oduced in he acking s ep allow he
p oblems co esponding o iden i y loss and swaps, caused by he in e ac ion be ween
indi idual lies, o be sol ed e icien ly. Thus, a obus me hod ha compa es a o ably o
o he exis ing me hods is ob ained.
Keywo ds: mo ing objec sensing; compu e ision; acking; p edic ion me hods
OPEN ACCESS
Senso s 2015, 15 19370
1. In oduc ion
The D osophila Melanogas e has become a powe ul model sys em o analyzing he ela ionship
be ween genes, neu ons and beha io . Impo an esea ch e o s ca ied ou wi h his insec ha e
allowed an unde s anding o many beha io s o medical in e es such as subs ance abuse [1,2],
agg essi i y [3,4], sleep dep i a ion [5], ageing [6] and memo y loss [7], among o he s. The wide ange
o gene ic manipula ions in he D osophila makes his animal an ideal gene ic sys em o analyzing he
gene al p inciples o neu oscience. Howe e , he analysis o he beha io al e ec s o hese manipula ions
is hampe ed by he lack o e ec i e me hods o measu e he lies’ beha io p ecisely and quan i a i ely.
In his con ex , he app oaches based on he use o compu e ision-based senso s ep esen a
p omising s a egy o he au oma ed acking and beha io cha ac e iza ion o he specimens unde
s udy. Howe e , he esul s ob ained wi h hese senso s a e highly dependen , no only on he quali y o
he image acquisi ion sys em, bu also on many o he aspec s ela ed o he in e ac ion be ween he
insec s (o e lapping, occlusions, ajec o y c ossing, collisions) and he p esence o e en ual op ical
a i ac s (shadows, e lec ions). All his hinde s he obus acking o he lies h ough ime.
In his pape , a me hod o dealing wi h he abo e-men ioned e ec s, using an op imized s a egy, is
p oposed in o de o achie e an e icien , au oma ed ly acking.
The pape is o ganized as ollows. The ela ed wo k is e iewed in Sec ion 2. The p oposed algo i hm
is p esen ed in Sec ion 3. Sec ion 4 shows some expe imen al esul s. Finally, concluding ema ks a e
gi en in Sec ion 5.
2. Rela ed Wo k
The s udy o he beha io o he D osophila ly and o he animals equi es de ailed obse a ion,
anno a ion and subsequen analysis o he da a ob ained. The numbe o indi iduals may ange om a
ew o a g ea numbe (say, hund eds), depending on he expe imen . Mo eo e , he indi iduals may be
s udied ei he alone o in g oups, and wi h o wi hou in e ac ion. F equen ly, ideo ma e ial is eco ded
o an ul e io , o -line analysis o he ajec o ies, aimed a ob aining con iden esul s.
Specimen obse a ion and anno a ion is gene ally ca ied ou by a human ope a o , which en ails
equen e o s, gi en ha con inued a en ion is equi ed o long pe iods o ime.
Fo example, neu obiological esea ch aimed a quan i ying he mo o dec ease o specimens a e a
gene ic modi ica ion is shown in [8]. To achie e his objec i e, specimens we e measu ed a egula
in e als, using a ch onome e , h ough ime pe iods anging om a ew hou s o se e al days.
This ep esen s a conside able e o and, e en ually, he esul ing con idence may be comp omised.
In his con ex , compu e ision sensing is a p omising app oach o he au oma ed acking and analysis
o he beha io o D osophila and o he animals h ough ime.
The acking o a mo ing elemen wi hin a gi en scene equi es a sui able disc imina ion be ween
his elemen and he backg ound, i.e., a sui able segmen a ion. Some common segmen a ion echniques
a e e iewed in [9,10], while some segmen a ion echniques based speci ically on backg ound modeling
a e discussed in [11].
Senso s 2015, 15 19371
Some p ac ical examples o mo ing objec segmen a ion upon backg ound modeling a e: ideo
su eillance [12,13], mo ion cap u e [14,15], indus ial quali y con ol [16] and he mul imedia,
en e ainmen and cinema indus y in [17,18].
Roughly speaking, backg ound modeling echniques a e aimed a ob aining an image o he scene
wi hou he mo ing objec s. This may ep esen a challenging objec i e, gi en ha he backg ound
appea ance may a y h ough ime (due o ligh ing changes), some elemen s o he scene may
appea /disappea a (o du ing) a gi en ime, quasi-s a iona y elemen s may also be p esen , e c. Di e en
echniques aimed a ob aining obus and adap i e esul s unde hese ci cums ances ha e been p oposed.
These can be classi ied in o: basic backg ound models [19,20], s a is ic backg ound models [11], uzzy
models [21,22], and models based on es ima ion [23,24]. The models can also be s udied in e ms o
p edic ion [25], ecu si eness [12], adap abili y [26], o modali y [27], e c.
All hese echniques sha e some common ea u es: a backg ound model is i s ob ained; hen he
model is ini ialized and upda ed h oughou he expe imen ; he o eg ound is ob ained; he elemen s in
he o eg ound a e disc imina ed on he basis o hei size in he image; and he main ea u es o he
desi ed a ge elemen a e chosen (colo , shape, mo emen , ex u e, e c.).
Once he a ge elemen has been de ec ed, i mus be acked h ough he image sequence. T acking
me hods ha e expe ienced a conside able ad ance in ecen yea s, gi en hei in e es wi hin he ield
o au oma ic ideo analysis. Some applica ion examples a e a ic moni o ing [12,13], mo ion
analysis [14,15], human-machine in e ac ion [28], and many o he s.
A e iew o acking echniques can be ound in [29,30]. They can be classi ied acco ding o he way
he shape and/o appea ance o he objec s is ep esen ed (poin s, geome ic shape, silhoue es o
con ou s, a icula ed models, empla es, e c.) and he ea u es o be acked (colo , shape, op ical low,
ex u e, e c.). Silhoue es a e used in [31] o acking objec s in a ideo su eillance applica ion. Objec s
a e ep esen ed as ellipses in [32], om which his og ams modeling hei appea ance a e ob ained.
Objec s a e modeled h ough ec angles and he co esponding eigen ec o s in [33]. Objec s a e
cha ac e ized using poin s ha a e acked h oughou he image sequence in [34]. A gene al pu pose
acking algo i hm ha ope a es on lida da a, ins ead o images, can be ound in [35]. Mo e ecen ly, [36]
de eloped ano he algo i hm ha is able o ack a g oup o animals. In his case, se e al images o a
eco ded ideo a e used o iden i y each indi idual by modeling i s appea ance.
Conce ning D osophila, some ep esen a i e wo ks a e [37,38], whe e he analysis o he beha io o
specimens is add essed, and [39] whe e a comple e analysis o he ajec o ies is pe o med. A acking
me hod o moni o ing isola ed specimens o D osophila (and o he animals) is p esen ed
in [40–42]; while a me hod o acking g oups o specimens in e ac ing wi hin a plana scene is
p esen ed in [29]. This is one o he mos ep esen a i e me hods o da e and has been used as he
e e ence in he cu en wo k. O he wo ks ha e ocused on he 3D acking o lies [43], wi h a high
se -up cos , and [44], which uses 3D acking o luminescen molecules o lies in a ial. None heless,
2D acking is o en p e e ed, gi en ha i p o ides simila in o ma ion a a lowe se -up cos . Mo e
ecen ly, [45] ha e p oposed a simple image p ocessing algo i hm o a ge ing a single ly wi h a lase
in o de o manipula e he ly’s ne ous sys em.
Some o he me hods ci ed abo e ha e been used widely, e en in comme cial sys ems, bu a e s ill
open o imp o emen in some aspec s. Conce ning [40,41,45] he analysis o indi iduals in e ac ing wi h
each o he wi hin a g oup is no unde aken. Conce ning [29], he acking e o could be imp o ed
Senso s 2015, 15 19372
h ough a e ined acking s a egy, as is p oposed in he cu en wo k. In addi ion, ha me hod is highly
dependen on he ly pla e, which esul s in a limi ed lexibili y.
Wi h espec o sys ems ha use some o he me hods e iewed in his sec ion, i is wo h men ioning
Noldus E hno ision, BioT ack, Id acke and C-T ax.
Noldus E hno ision [46] is a comme cial sys em able o ack se e al animals mo ing o e a la
en i onmen . The sys em can ope a e a o ably in he case o la ge animals (such as mice, a s and
o he s), gi en ha mul iple poin s on each specimen a e used o acking. Howe e , in he case o
D osophila, he sys em canno deal success ully wi h some common si ua ions, such as ack c ossing
and specimen occlusions. The sys em can ack mul iple indi iduals only when hey do no in e ac .
Bio ack [47], based on [35], is a mul i- acking so wa e ha can be used o ee. I is a gene al
pu pose solu ion ha can ack se e al kinds o animals in a la a ena bu deals pa ly wi h c ossings
and occlusions.
Id acke [48] is an open sou ce so wa e ha uses an algo i hm able o ack indi iduals in a
g oup [36]. None heless, his so wa e canno make co ec ions in eal ime in o de o a oid p opaga ing
iden i ica ion e o s.
C-T ax [49] is a eely a ailable so wa e o D osophila acking, de eloped by Cal ech (Cali o nian
Ins i u e o Technology). This sys em su e s om e en ual acking e o s (loss and swap o indi idual
iden i ies) de i ed om he selec ed acking p ocedu e, which uses a weakly cha ac e ized dynamic
model, based on [29]. Mo eo e , he ly pla e is shape- es ic ed and he ly acking is no pe o med in
eal ime.
To summa ize, he solu ions p oposed o da e o D osophila au oma ed acking su e om se e al
d awbacks. Some sys ems can ack isola ed lies, which se e ely limi s he analysis o ce ain beha io s.
O he sys ems can ack se e al lies bu canno hold indi idual iden i ies when wo o mo e lies
app oach o o e lap. Mo eo e , some p ac ical es ic ions a e o en p esen , such as a educed lexibili y
o adap o new expe imen s, o an inabili y o ope a e in eal ime and, in some cases, an impo an cos .
In he p esen wo k, a new model o D osophila acking is p oposed ha allows se e al in e ac ing
specimens o be iden i ied and acked unde noisy condi ions and e en ual specimen o e lapping.
The main ad an ages o he p oposed me hod de i e om he p oposed acking me hod. A acking
s a egy based on he op imal Kalman il e , which minimizes he p edic ion e o , hus signi ican ly
educing he iden i y swaps, is p oposed. In p e ious wo ks, such as [29], ob aining de ailed spa ial
in o ma ion gained p io i y o e ob aining a dynamic model o he lies. The cu en app oach is also
aimed a ob aining de ailed spa ial in o ma ion. Howe e , he use o his in o ma ion is imp o ed
signi ican ly by using he Kalman il e , which esul s in an imp o ed p edic ion o he sys em’s s a e
and, he e o e, a signi ican educ ion o he e o a e (gi en ha indi idual iden i ies a e be e
p ese ed hough ime). Fu he mo e, he p oposed me hodology is obus and adap able: he me hod
does no ely on a co ec de ec ion o he ly pla e and e en ly e lec ions a he pla e bounda ies a e
au oma ically emo ed. This con e s he sys em a signi ican lexibili y, gi en ha di e en pla es can
be used wi hou equi ing any p e ious adap a ion o he sys em.
Senso s 2015, 15 19373
3. The Algo i hm
The algo i hm p oposed in his pape ope a es in h ee main s eps: p e-p ocessing, p ocessing
and acking.
3.1. P ep ocessing
The images a e con e ed in o g ayscale and p ocessed h ough a Gaussian il e . Then, he pla e is
de ec ed and a backg ound model is compu ed. Finally, a shape model o he specimens is compu ed o
hei ul e io de ec ion and acking.
3.2. Pla e De ec ion
The pla e is he egion o in e es whe e he mo emen de ec ion will be ca ied ou . In he p esen
wo k, he pla e is de ec ed by using a Canny il e [50] ollowed by a ci cle Hough ans o m [51].
This is app op ia e o ci cle-shaped pla es, which is he mos common case. Simila p ocedu es could
be applied o o he pla e geome ies.
3.3. Backg ound Modeling
In gene al, ideo p ocessing sys ems seek o ex ac mo ing elemen s ( o eg ound) om s a iona y
elemen s (backg ound) in he images [9,10]. This can be achie ed by compu ing a p ope model o
he backg ound.
In ou case, he Simple Gaussian me hod (SG), wi h selec i e upda ing [9], has been used, gi en he
unimodal na u e o he backg ound. Small changes o backg ound pixel b igh ness a e modeled h ough
a unimodal Gaussian de ined upon es ima es o he b igh ness mean and de ia ion. In ou case, median
and median absolu e de ia ion MAD [52] ha e been used o an inc eased obus ness. A co ec ion e m
is hen applied o i he co ec da a in o one s anda d de ia ion:
μ
(,)=medI(,) (1)
(,)=cMAD=cmedI(,)−med(,) (2)
whe e x and y a e he pixel coo dina es, ∈󰇝0,∆,2∆,…,} and =1.4826. This alue ensu es ha he
co ec ac ion o da a is wi hin one s anda d de ia ion a ound he median [52].
This p o ides a backg ound model whe e he alue o each pixel co esponding o he Gaussian
de e mined by μ(x,y) and σ(x,y) is ob ained. This model is upda ed pe iodically du ing he
p e-p ocessing s ep, so ha only hose pixels ha do no belong o he backg ound a e e ised. Thus, he
backg ound will emain s able o e en ual ligh ing a ia ions and/o image quali y de iciencies, hus
p e en ing phan om a i ac s de i ed om s a ic lies o eme ge [24]. The upda e pe iod will depend on
he acquisi ion sys em quali y.

Senso s 2015, 15 19374
3.4. Shape Model
Once he backg ound model is compu ed, he elemen s o in e es in he images a e ob ained by
sub ac ing he backg ound (Sec ion 3.2), and hei a eas  a e compu ed. The p ocess is ca ied ou
h oughou an image sequence o compu e he co esponding mean μ and a iance σ
,
μ
=




 (3)
σ
=


−
μ

 (4)
whe e N is he numbe o elemen s. These wo alues will be used o de e mine whe he each de ec ed
elemen co esponds o a ly o no upon a p obabilis ic c i e ion based on he elemen a ea, hus a oiding
alse posi i es.
3.5. P ocessing: Elemen De ec ion and Segmen a ion
The de ec ion, segmen a ion and alida ion o he mo ing elemen s is add essed in he cu en
s ep [9,53].
3.6. Backg ound Sub ac ion
The o eg ound, i.e., he mo ing objec s, is ob ained by sub ac ing he (μ,σ)backg ound model
om he cu en image, (,)=255−(),()–
μ
())>σ()
0,((()–
μ
(()))≤0 (5)
whe e  a e he pixel coo dina es, 255 is he maximum alue possible o he in ensi y o a pixel and 0
is he minimum. The alue o N depends on he quali y o he da a acquisi ion sys em, and N σ (p) is he
on de ec ion h eshold, o {N = 0, 1,2,…n}. In ou case, ini ially N = 10. The be e he sys em we
ha e, he smalle he equi ed alue (because o he lowe backg ound noise). This alue is inc eased o
dec eased (wi h a hys e esis) du ing he alida ion phase o spli o me ge ellipses.
Those pixels whe e he di e ence be ween he b igh ness alue and he backg ound model is unde
he on de ec ion h eshold a e assumed o belong o he o eg ound (o o he backg ound o he wise).
An example is shown in Figu e 1: he inpu ame (Figu e 1a) and he esul a e backg ound
sub ac ion (Figu e 1b). (Backg ound is displayed in black (0) and o eg ound in ligh g ay
(255−()).
Senso s 2015, 15 19375
(a)
(b)
Figu e 1. Example o o eg ound de ec ion. (a) Inpu ame; (b) De ec ed o eg ound.
I is wo h men ioning ha images ha e been acqui ed using backligh ing. Pixels by he cen e o he
lies a e da ke han he su ounding pixels and he pixel b igh ness inc eases as he pixel app oaches he ly
bounda y. This will allow he i ing o he ly b igh ness by 2D Gaussians, and subsequen ly, o 2D ellipses.
The elemen s de ec ed in he image a e i s classi ied acco ding o hei a ea,
ℎ<

⇒
ℎ⇒ (6)
In his exp ession, hmin is he h eshold alue ha is se o μ−σ, , whe e μ and
σ a e he alues compu ed in Equa ions (3) and (4). This h eshold, se o c = 15, is used o clean
he o eg ound (in o de o ob ain a o eg ound mask ee o noise). Mo eo e , he c alue selec ed
should be la ge enough o deal wi h image noise. In gene al, he mo e noise- ee he acquisi ion sys em
is, he la ge c should be (in ou case c = 15). I is wo hwhile no ing ha his cleaning p ocess could be
done a he alida ion s ep, bu i is ac ually done in he p ocessing s ep o educe he compu ing e o
o he alida ion s ep.
3.7. Segmen a ion
The connec ed componen s in he o eg ound a e hen i ed o ellipses so ha hese componen s can
be classi ied o subsequen acking [28,31]. The pa ame e s o he ellipses a e compu ed, esul ing in
a lis =󰇝(,,
θ
,,)} (7)
whe e , , a e he coo dina es o he cen e o each elemen , θ is he o ien a ion (wi hin ±π), a is hal
he majo axis and b hal he mino axis. A, b and θ a e compu ed upon he co a iance ma ix o he
Gaussian model, which is ob ained om he mean and co a iance alues, he pixels being weigh ed by
hei di e ence o he backg ound le el. Adding weigh o each pixel is impo an because his weigh
allows connec ed componen s o be spli in he alida ion phase: he on de ec ion h eshold is inc eased
in a Region O In e es (ROI) a ound he ly; he e o e, he a ea o he ly dec eases. This can be
app ecia ed in Figu e 2, whe e he le els o a gi en ROI a ound a ly (a e sub ac ing he backg ound
and compu ing he absolu e alue) and he esul s co esponding o wo h eshold alues, 50 and 70, a e
shown. The a ea dec eases signi ican ly as he h eshold inc eases.
Senso s 2015, 15 19376
(a) (b) (c)
Figu e 2. The a ea o a gi en ly dec eases as he on h eshold inc eases. (a) Values in an
ROI a ound a ly (a e sub ac ing he backg ound and compu ing he absolu e alue);
(b) Resul co esponding o a h eshold alue o 50; (c) Resul co esponding o a h eshold
alue o 70.
In his way, i he e a e wo lies close o each o he , ini ially de ec ed as a single objec , his objec
can be spli in o he wo lies by means o inc easing he o eg ound h eshold in he alida ion phase.
=|()−()|/σ() (8)
=
 (9)
μ
=1
 (10)
In his exp ession 󰇝1,2,….} a e he coo dina es o he pixels o a gi en connec ed elemen
wi hin he image;  is he no malized di e ence om () o he a e age b igh ness o his pixel
wi hin he backg ound model, μ(); Z is he sum o hese di e ences; and μ hei a e age alue.
This allows he co a iance ma ix,
=1(−
μ
)(−
μ
)
 (11)
o be ob ained, which, in u n, co esponds o
=20
02 (12)
whe e =󰇡
θ

θ
−
θ

θ
󰇢 (13)
The eigen decomposi ion ∑= p o ides a and b, and θ is ob ained om R:
=2 (14)
=2 (15)
θ
=(
θ
/cos
θ
) (16)
Senso s 2015, 15 19377
The ellipse cen e loca ion(,) co esponds o he cen al pic o he Gaussian dis ibu ion.
3.8. Valida ion
In his s ep, undesi able a i ac s in he images a e il e ed. In pa icula , weakly connec ed
componen s (de i ed om nea by lies) will be spli in o he co esponding specimens and spu ious
a i ac s will be emo ed.
The p oblem is o mula ed in e ms o a p obabilis ic model: he ly loca ions ha bes explain he
cu en ame a e sea ched. In pa icula , he se o ellipse loca ions ha maximizes
(

|)(

)(|

) (17)
is sea ched, whe e X is he said se and I is he ac ual ame.
In gene al, he loca ion o he di e en lies may be assumed o be independen , which leads o
he o mula ion
(

)=()

 (18)
()=󰇝||/} (19)
whe e () is he likelihood e m associa ed o a no mal dis ibu ion model o each ly i, o
i = {1, 2, 3...N} (N being he o al numbe o lies), μ and σ a e he shape model pa ame e s,
and π is he a ea o he ellipse co esponding o ly i.
Finding an analy ical solu ion o his p oblem would be imp ac ical because p(X│I) has local maxima
and X is a disc e e se . The e o e, a heu is ic s a egy has been used. When a connec ed componen has
a weak p obabili y, he numbe o ellipses is inc eased o dec eased o ob ain an imp o ed p obabili y.
Mo e speci ically, in o de o ind he connec ed componen s co esponding o a gi en ly, he ellipse
enclosing his ly o a la ge p obabili y (in ela ion o ()) has o be compu ed. This p obabili y may
be low because he a ea is ei he oo la ge o oo small. I he a ea is oo la ge, his connec ed componen
may co espond o se e al nea by specimens. On he con a y, i he a ea is oo small, his componen
would co espond o a spu ious a i ac . The o eg ound de ec ion h eshold will hen be inc eased, in
he o me case, o dec eased in he la e . The new elemen s appea ing a e his h eshold uning will
be assumed o co espond o ac ual lies i hei () is high enough, o o a spu ious a i ac o he wise.
I is wo hwhile no ing ha he numbe o lies can ei he be in oduced by he use o es ablished by
he sys em au oma ically in he p e-p ocessing s ep. I he alue is es ablished by he sys em, i can
change i ano he ly is de ec ed in he acking phase. In any case, we seek o maximize he p obabili y
o P(X/I) by maximizing he p obabili y o () in he alida ion phase. Summa izing, knowing he
numbe o lies can be use ul bu is no equi ed by he algo i hm.
An example o he alida ion p ocess is shown in Figu e 3. The de ec ed elemen in Figu e3b
co esponds o wo ac ual lies, as can be seen in Figu e 3c.
Senso s 2015, 15 19384
Figu e 10. B igh ness o a backg ound pixel agains he ame numbe , wi hou ambien
ligh . (The pixel is loca ed a he midpoin o he ed line in Figu e 9, and ± N σ lines ha e
also been plo ed.) A p ecise backg ound model can be ob ained in his case.
(a) (b) (c)
Figu e 11. T acking ly in e ac ions (enci cled in yellow and ed). F ames a , + 1 and
+ 2 a e shown in (a–c), espec i ely. The ins an aneous speeds, hei mo ing a e age and
he Kalman il e ou pu s a e displayed ed, blue and g een, espec i ely. The iden i y o he
lies is p ese ed a e he in e ac ions hanks o he op imal il e combined wi h
he Hunga ian.
In he nex ame (Figu e 11b), a + 1, lies 1 and 2 a e o e lapped and lies 7 and 10 ha e al eady
sepa a ed. Finally, in he las ame (Figu e 11c), a + 2, he wo ly couples ha e sepa a ed comple ely.
I can be seen ha he Kalman il e combined wi h he Hunga ian p e en ly iden i ies om being los
o swapped du ing he whole sequence, i.e., he lies a e acked co ec ly, as expec ed.

Senso s 2015, 15 19385
4.3. Re lec ions
Heu is ic s a egies ha e been inco po a ed o he gene al p ocedu e o deal wi h some special
si ua ions ha may occasionally happen.
An example is shown in Figu e 12. As men ioned in a p e ious sec ion, he au oma ic pla e de ec ion
implemen ed in his wo k can be manually disabled when an accu a e de ec ion is no possible.
Un o una ely, some undesi ed e lec ions may appea in his case a he pla e bounda y. Fly 9 is nea
he bounda y a ame (Figu e 12a). A ly e lec ion a i ac ha has no been emo ed in he
p e-p ocessing s ep appea s in he image, a ame + 1 (enci cled in blue, in Figu e 12b). This a i ac
is assigned a new acke , acke 11. Each ime a new acke becomes ac i e, i s s abili y is es ed du ing
a ew ames (say 50 ames o a 15 ps a e). Then, undesi ed si ua ions a e de ec ed, such as: he
acke loses i s ly o he acke is assigned a ly ha has al eady been assigned ano he acke . In his
example, he new acke has been ac i e o oo sho a ime. The e o e, i has been emo ed
au oma ically and is no longe conside ed in subsequen ames (Figu e 12c). The same p ocedu e can
deal wi h simila si ua ions, such as spu ious a i ac s caused by poo image quali y.
(a) (b)
(c)
Figu e 12. Fly e lec ion a he pla e bounda y. (The ins an aneous speeds, hei mo ing
a e age and he Kalman il e ou pu s a e displayed ed, blue and g een, espec i ely.)
F ames a , + 1 and + 2, in (a–c) espec i ely, show ha he ly e lec ion (enci cled in
blue) is co ec ly emo ed by he algo i hm.
4.4. Quan i a i e Resul s
The esul s co esponding o six ideos o eigh D osophila in e ac ing wi hin a Pe i pla e, sized a
90 mm in diame e and 6mm in heigh , a e epo ed. The a e age ly popula ion was 0.12 lies/cm2.
The ideos we e eco ded in he a e noon, coinciding wi h he peak ac i i y o he lies, a a empe a u e
o 28 °C.
Some si ua ions we e es ed du ing he expe imen s, such as a pla e being placed pa ially ou side he
image o ly ood being p esen in he images. Mo eo e , di e en e o a es we e quan i ied. The
Senso s 2015, 15 19386
ob ained esul s a e epo ed in Table 1. The expe imen al condi ions a e desc ibed in Table 2
(along wi h he condi ions co esponding o ano he me hod aken as e e ence).
Table 1. T acking e o co esponding o six es ideos. The i s column shows he ideo
id. The second and hi d columns show he o al numbe o ames and he numbe o ames
wi h wo o mo e lies close o each o he (occlusion) epo ed by he so wa e, espec i ely.
Columns 4 o 6 show he numbe o e o s. The las columns epo he e o s wi h espec
o occlusions and ly densi y (and ime), espec i ely.
Video Numbe o F ames E o s E o F equency
To al Wi h Occlusion Swap Los Spu ious E o /Occlusions (%) E o /ρ/ (%)
1 4095 1782 0 1 0 0.84 3.05
2 5025 2181 2 0 0 1.37 4.97
3 5012 1533 2 0 0 1.95 4.98
4 5022 2366 1 0 0 0.63 2.48
5 5044 1834 0 0 0 0 0
6 5005 2024 3 0 0 2.2 7.4
Table 2. Compa ison be ween he sensing sys ems used in he expe imen s.
(Cu en Wo k) Re e ence [29]
Ligh ing Backligh LED 800 lumens
(wi h ambien ligh sc eening)
Ceni al LED 480 lumens
(wi h ambien ligh sc eening)
Op ics 1280 × 720
Mic oso ® Li eCam Cinema™
1280 × 1024-pixel, i ewi e came a (Basle A622 )
equipped wi h an 8 mm lens (PENTAX)
Resolu ion 4 pix/mm 4 pix/mm
FPS 15 20
Pla e adius 4.5 cm 12.5 cm
Videos 1, 2, 3, 4, 5 i, ii
To al popula ion (N) 8 50
Popula ion densi y (ρ) 0.12 lies/cm2 0.106 lies/cm2
The ideo id numbe and he numbe o ames a e anno a ed in he i s and second columns o
Table 1, espec i ely. The numbe o ames wi h occlusion ( epo ed by he so wa e), i.e., he numbe
o ames in which wo o mo e lies a e me ged in a single blob, is shown in column 3.
The acking e o s a e epo ed in columns 4, 5 and 6. Column 4 e e s o iden i y swaps; column 5
o iden i y losses; and column 6 o spu ious a i ac s no emo ed by he algo i hm. These igu es ha e
been de e mined manually by wa ching he p ocessing esul s.
The las wo columns e e o he iden i y swaps wi h espec o he numbe o ly in e ac ions and he
numbe o e o s (iden i y losses plus iden i y swaps) wi h espec o he ly popula ion densi y and he
ideo leng h (in seconds), espec i ely. Bo h esul s a e gi en in %.
I can be seen ha he sys em beha es sui ably conce ning ly occlusions and iden i y losses. In sum,
he wo s case co esponds o ideo sequence 6 ( he las ow in Table 1). This is a 5005- ame ideo
sequence and he e a e 2024 ames wi h o e lapping si ua ions. Howe e , only h ee o e lapping
si ua ions ha e esul ed in iden i y swaps. Mo eo e , he e has been no iden i y loss.
Senso s 2015, 15 19387
The p oposed me hod has also been compa ed o a s a e o he a one. The wo k [29] has been
selec ed, gi en ha i is conside ed a e e ence pape on D osophila acking. Quan i a i e compa ison
is a challenging issue, gi en ha he sensing sys ems used in ha wo k a e di e en om ou s. Mo eo e ,
en i onmen al condi ions such as empe a u e, ime o day, humidi y, e c., signi ican ly in luence he
beha io o he lies. The e o e, he a iabili y om one expe imen o he o he may be la ge.
Assuming he said a iabili y, he ob ained esul s a e epo ed in Tables 3 and 4. Table 3 shows he
p ocessing ime (in ames), he numbe o occlusions and he e o a es co esponding o wo ideo
sequences aken om he supplemen a y Table 1 o [29]. In conc e e, we ha e aken he ideos p esen ed
in ows 5 and 6 (labeled i and ii in he p esen wo k). These ideos ha e been selec ed because hei ly
densi y is simila o ou s (while he ly densi y in he o he ideos o ha wo k is lowe ). I is wo hwhile
no ing ha he ly densi y is a ca dinal aspec , gi en ha he la ge his densi y is, he mo e in e ac ions
he e will be be ween lies. Mo eo e , he e o a es ha e been compu ed in a di e en way wi h espec
o he said pape : E o /Occlusions has been compu ed by conside ing bo h he numbe o losses and
swaps (gi en ha , in ou opinion, an iden i y loss is also a acking e o and should be conside ed as
such); and he e o in he las column has been compu ed wi h espec o he ly popula ion densi y
ins ead o he numbe o lies used in he e e ence me hod. Wha is mo e, we ha e compu ed he e o
using he ime (in seconds) because he ame a e is di e en in he compa ed expe imen s. In ac , his
ame a e is lowe in ou case, which esul s in mo e acking di icul y. Finally, we ha e epo ed he
e o a e in %, which is an easily unde s andable measu e.
Table 3. E o equency. Figu es ha e been es ima ed om he da a shown in he e e ence
pape , acco ding o he c i e ia desc ibed in he cu en sec ion.
Supplemen a y Table 1 [29] Numbe o F ames E o s E o F equency
Video To al Wi h Occlusion Swap Los Spu ious Swap/Occlusions (%) E o /ρ/ (%)
i 6123 2135 0 2 0 1.87 6.16
ii 6587 4545 1 0 0 0.44 3.08
Table 4. Compa ison be ween he e o equency o he wo me hods.
Video (s) E o
(Swap + Los )
E o F equency
To al wi h Occlusion E o /Occlusions (%) E o /ρ/ (%)
i, ii [29] 6355 334 3 0.898 4.4
1,2.3,4,5,6 (P oposed me hod) 1946 984 8 0.813 3.4
Table 4 shows a quan i a i e compa ison be ween he wo me hods o a numbe o expe imen s
(using he e o a es de ined abo e). I can be seen ha he e o equency wi h espec o he numbe
o occlusions is simila o bo h me hods. This e o quan i ies he beha io o he p e-p ocessing s ep.
Howe e , he e o equency wi h espec o he popula ion densi y and he ime is clea ly lowe in
he case o he p oposed me hod. This e o quan i ies he beha io o he ly acking. The e o e, om
he ob ained esul s, we can deduce ha he use o he Kalman il e and he o he imp o emen s
p oposed in he cu en pape esul in signi ican p og ess.
Senso s 2015, 15 19388
Some execu ion examples o he p oposed me hod, ope a ing in eal ime, can be downloaded
a [58]. Some u he ma e ial (sou ce code, so wa e lib a ies, demos and documen a ion) can be
downloaded a [59].
A aluable ea u e o his me hod is ha i can wo k in eal ime. P ocessing akes only abou 7 o
15 ms pe ame in a (modes ) In el Pen ium 4, 3 GHz compu e . Mo eo e , he algo i hm is designed in
such a way ha he p e-p ocessing phase and he acking phase could be execu ed in pa allel in a dual
co e a chi ec u e.
5. Conclusions
In his pape , an au oma ic sys em o acking D osophila melanogas e mo emen s has been
p esen ed. The p oposed me hod deals e icien ly wi h p oblems de i ed om limi ed image quali y and
he in e ac ion among lies.
The limi ed image quali y esul s in he modi ica ion o he aspec o he lies ha , in u n, hinde s
hei de ec ion and would cause iden i y losses du ing acking. The desc ibed p e-p ocessing and
p ocessing s eps deal sui ably wi h hese p oblems and allow an accu a e o eg ound o be ob ained wi h
low compu a ional e o . Shadows, e lec ions and phan oms a e also a oided. Thus, he cos o he
equi ed sys em ha dwa e is kep low as a as bo h he acquisi ion sys em and he compu ing esou ces
a e conce ned.
Mo eo e , he p oposed acking algo i hm allows an accu a e p edic ion o he sys em s a e o be
ob ained wi h espec o o he exis ing algo i hms, hus esul ing in a signi ican ly educed a e o ly
iden i y losses and swaps. Mo eo e , he me hod has been comple ed wi h some heu is ics ha deal wi h
p oblems ha may e en ually happen, such as spu ious a i ac s and ly e lec ions a he pla e bo de .
Fu he mo e, he co ec acking does no equi e an accu a e de ec ion o he pla e, which eases he
me hod adap a ion o di e en expe imen s. The whole p ocedu e—p e-p ocessing, p ocessing and
acking eal ime—can ope a e in eal ime.
The me hod allows de ailed and p ecise in o ma ion on he lies o be ob ained conce ning bo h hei
dynamics and hei beha io . The dis ance a elled by each ly, he ins an and a e age eloci ies, he
numbe o jumps e c., can be eadily compu ed om he ob ained da a. Fu he mo e, he di e en lies
can be classi ied by hei le el o ac i i y ( h oughou di e en ime pe iods) o he ime hey ha e s ayed
in a gi en ac i i y le el, o by a gi en pla e egion. The a e age beha io o all he lies in he expe imen
can also be compu ed.
This no able amoun o da a is use ul o analyzing he beha io o he lies in a g ea numbe o
di e en expe imen s.
Acknowledgmen s
This esea ch has been pa ially suppo ed by he Jun a de Cas illa y León (Agencia de In e siones y
Se icios, y P og. de Apoyo a P oys. de In es igación (g an VA036U14) and Minis y o Science and
Inno a ion (g an DPI2011-25489). We also hank he IBGM (U a-CSIC) o p o iding he specimens
and pla es.
Senso s 2015, 15 19389
Au ho Con ibu ions
Rubén Chao and Ge mán Macía-Vázquez ha e implemen ed he sys em so wa e and ha dwa e.
Edua do Zalama, and Jaime Gómez-Ga cía-Be mejo, a he Uni e si y o Valladolid, ha e iden i ied he
need and ha e designed he main lines o he me hod. José-Ramón Pe án has e ised he manusc ip .
Con lic s o In e es
The au ho s decla e no con lic o in e es .
Re e ences and No es
1. Wol , F.W.; Hebe lein, U. In e eb a e models o d ug abuse. J. Neu obiol. 2003, 54, 161–178.
2. Gua nie i, D.J.; Hebe lein, U. D osophila melanogas e , a gene ic model sys em o alcohol
esea ch. In . Re . Neu obiol. 2003, 54, 199–228.
3. Chan, Y.B.; K a i z, E.A. Speci ic subg oups o F uM neu ons con ol sexuallydimo phic pa e ns
o agg ession in D osophila melanogas e . P oc. Na l. Acad. Sci. USA 2007, 104, 19577–19582.
4. Hoye , S.C.; Ecka , A.; He el, A.; Za s, T.; Fische , S.A.; Ha die, S.L.; Heisenbe g, M.
Oc opamine in male agg ession o D osophila. Cu . Biol. 2008, 18, 159–167.
5. Ho, K.S.; Sehgal, A. D osophila melanogas e : An insec model o undamen als udies o sleep.
Me hods Enzymol. 2005, 393, 772–793.
6. Shaw, P.; Oco , K.; Bodme , R.; Oldham, S. D osophila aging. Exp. Ge on ol. 2008, 43, 5–10.
7. Konsolaki, M.; Song, H.J.; Dobbs, W.; Ga za, D. P2–109 D osophila modelso Alzheime ’s- ela ed
pa hways. Neu obiol. Aging 2004, 25, S255–S255.
8. Manuela, C.E.; Diego, S.M.; Dolo es, G. Qué hace Laza illo Glial? P egun émosle a las A axias
Espinoce ebelosas, Ins i u o de Biología y Gené ica Molecula -Depa amen o de Bioquímica y
Biología Molecula y Fisiología, Uni e sidad de Valladolid-CSIC, Valladolid, Spain, 2010.
9. Picca di, M. Backg ound sub ac ion echniques: A e iew. In P oceedings o he 2004 IEEE
In e na ional Con e ence on Sys ems, Man and Cybe ne ics, The Hague, The Ne he lands, 10–13
Oc obe 2015, 3099–3104.
10. B adski, G.; Kaehle , A. Lea ning OpenCV: Compu e Vision wi h he OpenCV Lib a y;
O’Reilly Media, Inc.: Sebas opol, CA, USA, 2008.
11. Cheung, S.; Kama h, C. Robus backg ound sub ac ion wi h o eg ound alida ion o U ban
T a ic Video. J. Appl. Signal P oc. 2005, 14, 2330–2340.
12. Mau in, B.; Masou , O.; Papanikolopoulos, N.P. T acking all a ic: compu e ision algo i hms
o moni o ing ehicles, indi iduals, and c owds. IEEE Robo . Au om. Mag. 2005, 12, 29–36.
13. Ca anza, J.; Theobal , C.; Magno , M.; Seidel, H. F ee-Viewpoin Video o Human Ac o s.
ACM T ans. G aph. 2003, 22, 569–577.
14. Ho p ase , T.; Ha i aoglu, I.; W en, C.; Ha wood, D.; Da is, L.; Pen land, A. Real- ime 3D mo ion
cap u e. In P oceedings o he Wo kshop on Pe cep ual Use In e aces, San F ancisco, CA, USA,
5–6 No embe 1998.
15. B osnan, T.; Sun, D.W. Imp o ing quali y inspec ion o ood p oduc s by compu e ision––A
e iew. J. Food Eng. 2004, 61, 3–16.

Senso s 2015, 15 19390
16. Wa en, J. Unencumbe ed Full Body In e ac ion in Video Games. Mas e ’s Thesis, MFA Design
and Technology, Pa sons School o Design, New Yo k, NY, USA, 2003. (Unpublished)
17. Semani, D.; Bouwmans, T.; F élico , C.; Cou ellemon , P. Au oma ic ish ecogni ion in in e ac i e
li e ideos. In P oceedings o he IVRCIA2002, O lando, FL, USA, 14–18 July 2002.
18. Lee, B.; Hedley, M. Backg ound es ima ion o ideo su eillance. In P oceedings o he Image and
Vision Compu ing New Zealand (IVCNZ 2002), Auckland, New Zealand, 22–28 No embe 2002;
pp. 315–320.
19. Zheng, J.; Wang, Y.; Nihan, N.; Hallenbeck, E. Ex ac ing oadway backg ound image: A mode
based app oach. J. T ansp. Res. Rep. 2006, 1944, 82–88.
20. El Ba , F.; Bouwmans, T.; Vachon, B. Fuzzy in eg al o mo ing objec de ec ion. In P oceedings
o he IEEE In e na ional Con e ence on Fuzzy Sys ems (FUZZ-IEEE2008), Hong Kong, China,
1–6 June 2008.
21. Siga i, M.; Mozayani, N.; Pou eza, H. Fuzzy unning a e age and uzzy backg ound sub ac ion:
Concep s and applica ion. In . J. Compu . Sci. Ne w. Secu . 2008, 8, 138–143.
22. Messelodi, S.; Modena, C.M.; Sega a, N.; Zanin, M. A kalman il e based backg ound upda ing
algo i hm obus o sha p illumina ion changes. In Image Analysis and P ocessing–ICIAP 2005;
Sp inge : Be lin Heidelbe g, Ge many, 2005.
23. Cucchia a, R.; G ana, C.; Picca di, M.; P a i, P. De ec ing Mo ing Objec s, Ghos s, and Shadows
in Video S eams: IEEE T ans. Pa e n Anal. Mach. In ell. 2003, 25, 1337–1342.
24. Wang, H.; Su e , D. A no el obus s a is ical me hod o backg ound ini ializa ion and isual
su eillance. In Compu e Vision–ACCV 2006; Sp inge : Be lin/Heidelbe g, Ge many, 2006.
25. Po ikli, F.; Tuzel, O. Human body acking by adap i e backg ound models and mean-shi analysis.
In P oceedings o he IEEE In e na ional Wo kshop on Pe o mance E alua ion o T acking and
Su eillance, Nice, F ance, Oc obe 2003; pp. 1–9.
26. Po ikli, F.; Tuzel, O. Bayesian backg ound modeling o o eg ound de ec ion. In P oceedings o
he Thi d ACM In e na ional Wo kshop on Video Su eillance & Senso ne wo ks, New Yo k, NY,
USA, 1–2 Augus 2005; pp. 55–58.
27. Bulling, A.; Gelle sen, H. Towa d Mobile Eye-Based Human-Compu e In e ac ion. IEEE Pe asi e
Compu . 2010, 9, 1536–1268
28. B anson, K.; Robie, A.A.; Bende , J.; Pe ona, P.; Dickinson, M.H. High-Th oughpu e homics in
la ge g oups o D osophila. Na . Me hods 2009, 6, 451–457.
29. Yilmaz, A.; Ja ed, O.; Shah, M. Objec acking: A su ey. ACM Compu . Su . 2006, 38, 13.
30. Ha i aoglu, I.; Ha wood, D.; Da is, L.S. W 4: Real- ime su eillance o people and hei ac i i ies.
IEEE T ans. Pa e n Anal. Mach. In ell. 2000, 22, 809–830.
31. Comaniciu, D.; Mee , P. Mean shi analysis and applica ions. In P oceedings o he IEEE
In e na ional Con e ence on Compu e Vision (ICCV), Ke ky a, G eece, 20–27 Sep embe 1999;
pp. 1197–1203.
32. Black, M.; Jepson, A. Eigen acking: Robus ma ching and acking o a icula ed objec s using a
iew-based ep esen a ion. In . J. Compu . Vis. 1998, 26, 63–84.
33. Veenman, C.; Reinde s, M.; Backe , E. Resol ing mo ion co espondence o densely mo ing
poin s. IEEE T ans. Pa . Analy. Mach. In ell. 2001, 23, 54–72.
Senso s 2015, 15 19391
34. Feldman, A.; Hybine e, M.; Balch, T. The Mul i-I e a i e Closes Poin T acke : An online
algo i hm o acking mul iple in e ac ing a ge s. J. Field Robo . 2012, 29, 258–276.
35. Pé ez-Escude o, A.; Vicen e-Page, J.; Hinz, R.C.; A ganda, S.; de Pola ieja, G.G. idT acke :
T acking indi iduals in a g oup by au oma ic iden i ica ion o unma ked animals. Na . Me hods
2014, 11, 743–748.
36. Ma in, J.R. A po ai o locomo o beha iou in D osophila de e mined by a ideo- acking
pa adigm. Beha . P ocess. 2004, 67, 207–219.
37. Ramazani, R.B.; K ishnan, H.R.; Be geson, S.E.; A kinson, N.S. Compu e au oma ed mo emen
de ec ion o he analysis o beha io . J. Neu osci. Me hods 2007, 162, 171–179.
38. Valen e, D.; Golani, I.; Mi a, P.P. Analysis o he T ajec o y o D osophila melanogas e in a
Ci cula Open Field A ena. PLoS ONE 2007, 2, e1083.
39. Danke , H.; Wang, L.; Hoop e , E.D.; Ande son, D.J.; Pe ona, P. Au oma ed moni o ing and
analysis o social beha io in D osophila. Na . Me hods 2009, 6, 297–303.
40. Donelson, N.; Kim, E.Z.; Slawson, J.B.; Vecsey, C.G.; Hube , R.; G i i h, L.S. High-Resolu ion
Posi ional T acking o Long-Te m Analysis o D osophila Sleep and Locomo ion Using he
“T acke ” P og am. PLoS ONE 2012, 7, e37250.
41. Giles o, G.F. Video acking and analysis o sleep in D osophila melanogas e . Na . P o oc. 2012,
7, 995–1007.
42. A dekani, R.; Biyani, A.; Dal on, G.E.; Sal z, G.B.; A bei man, M.N.; Towe , J.; Nuzhdin, S.;
Ta a é, S. Th ee-Dimensional acking and beha iou moni o ing o mul iple ui lies.
J. R. Soc. In e ace 2013, 10, 20120547.
43. G o e , D.; Yang, J.; Fo d, D.; Ta a é, S.; Towe , J. Simul aneous acking o mo emen and gene
exp ession in mul iple D osophila melanogas e lies using GFP and DsRED luo escen epo e
ansgenes. BMC Res. No es 2009, 2, 58.
44. Ba h, D.E.; S owe s, G.R.; Hö mann, D.; Poehlmann, A.; Dickson, B.J.; S aw, A.D. FlyMAD:
Rapid he mogene ic con ol o neu onal ac i i y in eely walking D osoph. Na . Me hods 2014,
11, 756–762.
45. Noldus E hno isión. A ailable online: h p://www.noldus.com/animal-beha io - esea ch/p oduc s/
e ho ision-x (accessed on 29 July 2015).
46. Bio ack, 2012. A ailable online: h p://www.bio- acking.o g/ca ego y/so wa e/ (accessed on
29 July 2015).
47. idT acke . A ailable online: h p://www.id acke .es (accessed on 29 July 2015).
48. C-T ax. A ailable online: h p://c ax.sou ce o ge.ne / (accessed on 29 July 2015).
49. Canny, J. A compu a ional app oach o edge de ec ion. IEEE T ans. Pa e n Anal. Mach. In ell.
1986, 8, 679–714.
50. Kimme, C.; Balla d, D.H.; Sklansky, J. Finding ci cles by an a ay o accumula o s.
Commun. Assoc. Compu . Mach. 1975, 18, 120–122.
51. Rousseeuw, P.J.; C oux, C. Al e na i es o he median absolu e de ia ion. J. Am. S a . Assoc. 1993,
88, 1273–1283.
52. B oida, T.; Chellappa, R. Es ima ion o objec mo ion pa ame e s om noisy images. IEEE T ans.
Pa . Analy. Mach. In ell. 1986, 8, 90–99.
Senso s 2015, 15 19392
53. G eg, W.; Ga y, B. An In oduc ion o he Kalman Fil e ; Depa men o Compu e Science,
Uni e si y o No h Ca olina a Chapel Hill: Chapel Hill, NC, USA, 2005.
54. Maybeck, P.S. S ochas ic Models, Es ima ion, and Con ol; Academic P ess: New Yo k, NY,
USA, 1982.
55. Pe ez, M. Mé odo de cuan i icación del uido en el Fil o de Kalman, Obse ado es de Es ado en
P oceso Es ocás icos, Fil o de Kalman. s.l., Depa amen o de Elec ónica, Uni e sidad de Alcalá
de Hena es, Mad id, Spain, 2009.
56. Papadimi iou, C.H.; S eigli z, K. Combina o ial Op imiza ion: Algo i hms and Complexi y;
P en ice Hall: Englewood Cli s, NJ, USA, 1982.
57. Supplemen a y Downloads. A ailable online: h p://www.eii.u a.es/%7Eeduzal/D osophila/
(accessed on 29 July 2015).
58. DT ack. A ailable online: h ps://gi hub.com/RubenChao/ acking-d osophila (accessed on
29 July 2015).
59. Video demons a ion. A ailable online: h ps://www.you ube.com/wa ch? =ANLp IK59Qc
(accessed on 29 July 2015).
© 2015 by he au ho s; licensee MDPI, Basel, Swi ze land. This a icle is an open access a icle
dis ibu ed unde he e ms and condi ions o he C ea i e Commons A ibu ion license
(h p://c ea i ecommons.o g/licenses/by/4.0/).