IAENG Iteratoal Joural of Computer Scece, 45:3, IJCS_45_3_0 A Improved Mea Sft Ug Adaptve Fuzzy Gaua Kerel for Idoea Vecle Lcee Plate Trackg Bauk Ramat, Member, IAENG, Edra Joelato, I Ketut Eddy Purama, ad Maurd Hery Puromo, Member, IAENG Abtract A ew approac toward Idoea vecle lcee plate trackg baed o vdeo recordg of vecle o te gway, propoed. Te trackg tecque ued to mprove te performace of a tadard Mea Sft wt a Gaua kerel by electg te approprate kerel radu ug a adaptve fuzzy mecam. Te purpoe of kerel radu varato of Parze wdow to keep or maxmze te mea of te mlarty fucto output wc mple a ucceful trackg proce. Te expermetal reult ow tat Improved Mea Sft ug Adaptve Fuzzy Gaua Kerel proved to ave better effect a compared to te Stadard Mea Sft. Idex Term Improved, Mea Sft, Adaptve, Fuzzy, Gaua, Kerel radu, Parze wdow O I. INTRODUCTION BJECT trackg oe of te mot mportat compoet tat ca be appled te feld of computer vo-baed vdeo. Te trackg-baed vdeo te tak of etmatg tme poto of te object aalyzed a equece of mage. Object trackg a alway bee exctg ad callegg for reearcer wo wat to aalyze vdeo-baed object. Vdeo-baed object trackg reearc cotue to be developed, amog oter [] [7]. Vdeo trackg purug a object or object a ere of mult-equece of vdeo frame. A eetal problem tat ofte faced computer vo applcato object trackg. T due to te dffculte trackg object tat ca are from teral ad exteral factor uc a dfguremet, moto camera, moto-blur, ad occluo [8]. Te major callege tat mut be codered we degg a object trackg ytem te appearace of te object model ad te caddate model te cee recorded o vdeo tat ofte mlar, ad cludg te varato Maucrpt receved May 7t, 07; reved July 8t, 08. T work wa upported by Drectorate Geeral of Hger Educato, Mtry of Reearc ad Educato, Idoea (doctoral fellowp program). Bauk Ramat a tudet doctoral program of Electrcal Egeerg Departmet, Ittut Tekolog Sepulu Nopember (ITS), Surabaya, Idoea. He work a a lecturer of Iformatc Egeerg Departemet of Uverta Pembagua Naoal Vetera Jawa Tmur, Idoea. (emal: baukramat.f@upjatm.ac.d). Edra Joelato wt te Itrumetato ad Cotrol Reearc Group, Faculty of Idutral Tecology, Ittut Tekolog Badug, Badug 403, Idoea (e-mal: ejoel@tf.tb.ac.d). I Ketut Eddy Purama ad Maurd Hery Puromo are wt te Electrcal Egeerg Departmet, Ittut Tekolog Sepulu Nopember (ITS), Surabaya, Idoea, 60 (e-mal: {ketut, ery@ee.t.ac.d}). appearace of te object telf. I geeral, everal metod of vecle lcee plate trackg or object trackg ca be categorzed baed o te feature tat are ued, amog oter tg: a. Color Feature [9], [0]. b. Superpxel Feature []. c. Oreted FAST ad Rotated BRIEF (ORB) Feature Matcg []. d. Edge Feature, Optcal flow Feature, ad Texture Feature [0]. Tree major part are developed te proce of vecle lcee plate trackg ad recogto ytem troug vdeo urvellace a ow Fg., amely: - Lcee Plate Extracto part to get te vecle lcee plate locato cotaed te vdeo frame. - Lcee Plate Trackg part to track te vecle lcee plate locato alog multple vdeo frame, ad - Caracter Extracto part, wc tere are problem of lcee plate caracter egmetato ad recogto. T paper clude te tudy o lcee plate trackg by ug a Improved Mea Sft wt Gaua kerel wc govered by a adaptve fuzzy mecam. Te ma problem aocated wt lcee plate trackg vdeo data te poblty of object oter ta te lcee plate telf appearg a backgroud mage tat mgt be mlar to te target appearace wc could caue terferece wt obervato. I uc cae, t may be dffcult to dtgu te feature of te expected target, wc produce a peomeo kow a clutter. I addto to te callege due to clutter, t alo wort otg ome of tee followg factor [3]: a. Pog Varato. No-tatoary target wll ave dfferet varato we t projected oto te mage plae, uc a we turg or cagg drecto. b. Ambet llumato. Drecto, tety, ad color of te lgt from urroudg evromet affect te appearace of te target. Alo, global llumato cage ofte gve problem for vdeo recordg data te outdoor. For example, te lgt evromet wll ave a dfferet effect we te lgt from te u obcured by cloud. I addto, te agle betwee te lgt drecto ad te ormal to te urface of te object poe dfferet object, wc wll affect vo troug te camera le. c. Noe. Te acquto proce cot of a ere of mage te vdeo frame may troduce oe, depedg o te qualty of te camera ued to capture. (Advace ole publcato: 8 Augut 08)
IAENG Iteratoal Joural of Computer Scece, 45:3, IJCS_45_3_0 Ma proce Metod / Algortm Lcee Plate Extracto Localzato of plate Plate extracto Ug: Boudary/Edge Iformato, Global Image Iformato, Texture Feature, Color Feature, Caracter Feature, etc Vdeo Acquto: Capturg vdeo Lcee Plate Trackg Foregroud etmato Image tream Image tream Image tream Ug :Color Feature,Superpxel Feature,Oreted FAST ad Rotated BRIEF ( ORB )Feature Matcg,Edge Feature,Optcal flow Feature, Texture Feature, etc Lcee plate eac frame Improved Mea Sft Preproceg: Graycale, Black & Wte, Meda flterg, Dlato, Eroto, Covoluto, Flood fll, etc. Caracter Extracto Caracter egmetato Segmetato Ug: Pxel Coectvty, Projecto Profle, Pror Kowledge of Caracter, Caracter Cotour, Combed Feature, etc Caracter recogto Recogto Ug: Raw Data, Extracted Feature, etc Fg.. Vdeo-baed vecle lcee plate trackg ad recogto ytem. Object obervato due to oe could ave drupted ome of te data wc degraded ytem performace. d. Occluo. Obervato falure could alo appe we te object partally or completely clogged (blocked) by oter object. Falure to oberve te target due to obcurato or cloggg by oter object tat may be preet te cee. Alo, tere are oter mportat ue related to te lcee plate telf. Tg to ote from te problem of te vecle lcee plate or te evromet are te detecto ad te vecle lcee plate recogto. Te problem are decrbed a follow [4]: ). Vecle lcee plate varato a) locato (poto): a plate of te vecle a dfferet poto for eac vecle; b) quatty: a pcture of te vecle wc mgt ave more ta oe plate; c) ze: vecle lcee plate could ave a varety of ze depedg o te dtace captured by te lee.it alo mportat to take to accout zoom factor; d) color: vecle lcee plate could ave may dfferet caracter (letter ad umber) ad backgroud color accordg to te type of plate or magg equpmet; e) letter: vecle lcee plate letter could vary oe coutry ad aoter due to dfferet uage of fot ad laguage; f) tadard ad otadard: eac coutry a t ow tadard rule of lcee plate umberrg, owever, may foud a lot of vecle are ug o-tadardzed lcee plate; g) occluo: vecle umber plate may be rouded by drt; ) tred: vecle lcee plate ca be tlted; ) oter: te addto of caracter, vecle lcee plate, ad crew frame. ). Evrometal varato a) llumato: put mage may ave a varato of lgtg due to evromet ad vecle lgt; b) backgroud: te backgroud mage may ave a patter mlar to te vecle lcee plate, uc a te umber o te vecle, te bumper wt a vertcal patter, ad formed groud. Oe algortm tellget ytem tat cotatly beg developed to date ad lkely developed teadly for te foreeeable future to olve problem epecally for vdeo-baed object trackg Mea Sft algortm. Some example of te applcato of t algortm to olve te problem of vdeo-baed object trackg, amog oter, ca be foud te followg paper [],[5],[5] [7]. Te Mea Sft algortm erve a a teratve algortm, ad t powerful ad veratle. It a a oparametrc ature tat allow eaer combato ad tegrato wt oter algortm. Some prevou reult ave maaged to mprove te performace of Mea Sft by addg teccal or oter algortm, amog oter ca be foud [9], [], [8] [6]. T paper erve a a propoal to aceve better reult of te tadard Mea Sft wt te Gaua kerel by electg approprate kerel radu ug adaptve fuzzy mecam. Some of te reao for ug adaptve fuzzy clude flexble ytem, capable of modelg complex olear fucto, ad workg wt oter tecque [7] [3]. Tee Mea Sft ad adaptve fuzzy combato algortm are to olve te problem of vdeo-baed Idoea vecle lcee plate trackg. Improved Mea Sft performace ca be accompled by ug adaptve fuzzy ytem mecam baed o te Probablty Dety Fucto (PDF) of te object model ad (Advace ole publcato: 8 Augut 08)
IAENG Iteratoal Joural of Computer Scece, 45:3, IJCS_45_3_0 PDF of a caddate model. Te propoed metod baed o te coderato tat te radu of Parze wdow kerel oe of te mportat ad fluetal part te geeratg of te PDF of te object model ad te caddate model. By cagg te radu, te propoed algortm attempt to mata te get performace trackg a object term of mlarty evaluato decrbed by te accuracy of trackg called percetage accuracy of object trackg (PAOT). II. PRELIMINARIES A. Probablty Dety Fucto (PDF) Te matematcal defto of te cotuou probablty fucto, p(), atfe te followg properte [3]:. Te probablty tat betwee two pot a ad b. p (a b) p ( ) d B. Kerel Dety Etmato Kerel Dety Etmato (KDE) a metod to etmate PDF of radom varable troug o-parametrc way. Te kerel dety etmato a matter of bac data mootg were populato cocluo are made, baed o a lmted ample of data. Area uc a gal proceg ad ecoometrc, t metod alo called te Parze-Roeblatt wdow metod, amed after Emauel Parze ad Murray Roeblatt, wc credted wt makg t depedet t curret form [33], [34]. It furter codered tat ℜ a cetered ypercube (-D quare), a ow Fg.. Te legt of te edge of te ypercube ca be repreeted by, wc V = for te -D quare, ad V = 3 for a 3-D cube [3]. ( /, / ) ( /, / ) () a. Value are ot egatve for all real. 3. Te tegral of te probablty fucto oe: p() d () Te mot commoly ued probablty fucto te Gaua fucto (alo kow a te ormal dtrbuto): p() ( c) exp (3) Extedg to te cae of vector, owed by o-egatve p() wt te followg properte:. Te probablty tat wt ℜ rego : (4) P p( ) d. Te tegral of te probablty fucto oe, tat : p() d (5) Wt a et of ample of data,...,, We ca ue a metod called dety etmato wc te dety fucto p() ca be etmated, te te output of p() for eac upcomg ample obtaed. Te bac dea bed may metod to etmate te probablty of a ukow dety fucto very mple. T tecque deped o te probablty P tat a vector fall a rego ℜ a gve by Eq.(4). If t aumed tat ℜ o mall tat p() ot muc dfferet t, te t ca be wrtte: (6) P p( ) d p( ) d p( )V Were V te volume of ℜ. Suppoe ample,..., are pulled depedetly accordg to te probablty fucto of dety p(), ad tere k of ample wt te rego ℜ, te te relatop obtaed: (7) P k / So come to te followg obvou etmate for p(): p( ) k / V (8) ( /, / ) ( /, / ) Fg.. Hypercube quare -D. A equato wa troduced [3]: g 0 k k /, k, oterwe (9) Idcate weter te box (cetered o, wt wdt ) or ot. Te umber of ample k fallg wt te rego ℜ, from, gve by: k g (0) So by ug Eq.(8), ad V = for a -D quare, te Parze probablty dety etmato formula (for -D) obtaed: k / V g p( ) g () called wdow fucto [3]. Te for te 3-D cube ypercube, t aumed tat te area ℜ wc ecloe te k ample from, te ypercube wt te log de ℎ cetered o, a ow Fg.3, wt volume V = 3. Te te total umber of pot de te ypercube, a Eq.(0). Ug Eq.(8), ad V = 3 for 3-D cube, te Parze probablty dety etmato formula (for 3-D) obtaed: (Advace ole publcato: 8 Augut 08)
IAENG Iteratoal Joural of Computer Scece, 45:3, IJCS_45_3_0. Uform k / V 3 g p( ) /, for u g u 0, for u oter () (6). Tragular u, for u g u 0, for u oter 3. Epaeckov 3 / 4( u ), for u g u 0, for u oter g u Tu geeral wt dmeo D, te Parze probablty dety etmato formula obtaed [35]: D g (3) Prevouly metoed, tat g((-)/) called wdow fucto. I practce, g( ) t wdow fucto, uually te kerel fucto ued. T kerel, wc correpod to a ceter-baed ypercube ut, kow a te Parze wdow or aïve etmator [35]. Te quatty g((-)/) equal te uty f te ypercube wt te de ℎ cetered o, ad equal to zero f t outde te ypercube. Wt g( ) te wdow fucto ug te kerel fucto, te Parze probablty dety etmate wt dmeo D Eq.(3) alo called Kerel Dety Etmato (KDE), expreed a Eq.(4) [35]. pkde p( ) D g (4) exp ( u ), for u Te real dety = 0.05 = 0.337 = p() C. Kerel Fucto Etmato of PDF by ug KDE alway volve te kerel fucto. Te kerel fucto g( ) atfe te properte of te cotuou probablty fucto,.e. te real oegatve value, ad te tegral of te probablty fucto oe: (5) (9) D. Kerel Radu or Badwdt Selecto A metoed prevouly, from Eq.(4) tere a parameter te PDF etmato by ug KDE. T parameter very mportat ad fluece te ucce of PDF etmato. Te parameter of t kerel fucto called kerel badwdt [36], [37] or ometme alo called kerel radu [9]. To llutrate te effect of te parameter o te ucce of te PDF etmato, a radom ample of tadard ormal dtrbuto a gve Eq.(3) ca be ee o te x-ax (orzotal), a ow Fg.4. Te gray curve te actual dety (ormal dety average 0 ad varat ). We compared, te red curve look wore becaue tere are a lot of artfcal data artfact tat comg from te ue of kerel radu = 0.05, wc too mall. Te over moot gree curve becaue ug te kerel radu = roud may of te foudato tructure. Te black curve wt te kerel radu of = 0.337 optmally refed becaue te approxmate dety cloe to te actual dety. Te de wll geerally be f te wdow fucto ue kerel fucto, t called kerel badwdt [36], [37] or kerel radu [9]. g (u) du (8) 4. Gaua Fg. 3. Hypercube cube 3-D. p( ) (7) Fg. 4. Ifluece of kerel radu cage. Some commoly ued kerel fucto clude: (Advace ole publcato: 8 Augut 08)
IAENG Iteratoal Joural of Computer Scece, 45:3, IJCS_45_3_0 p() p() If te kerel radu expreed a a percet, a 3-D vualzato, te kerel radu effect o te mak of a Parze wdow ug Gaua kerel, wt te repectve radu, = 65%, 75%, 85 % ad 95% a ow Fg. 5. (b) p() p() (a) (c) (d) Fg. 5. Te ape of te mak of a Parze wdow ug Gaua kerel, wt te repectve radu, = 65%, 75%, 85 % ad 95%. E. Mea Sft Bac Mea Sft, propoed by Fukuaga ad Hotetler 975, a o-parametrc etmato trategy [38]. It ca be exteded to locate te deal matc betwee te object model ad te caddate model lgt of te gradet etmato of te PDF etmato te kerel [39]. Te mlarty fucto ued to ae te mlarty of PDF of te object model ad te caddate model. To udertad ow te Mea Sft work, t aumed tutvely a to te detcal dtrbuto of bllard ball a ow Fg.6 [40]. Te goal to fd te deet or te get dtrbuted bllard dtrbuto dety. For example, a arbtrary locato coe,.e. te rego of teret (ROI) a ow Fg.6. From te poto at te ceter pot of te ROI, te ceter of ma of te percal dtrbuto kow. Furtermore, from te ceter pot of ROI ft toward te ceter of te ma. Leavg te Mea Sft vector trace. Te prevou ma dtrbuto ceter of te ball ow te ceter of ROI. From t poto at te ew ceter of ROI, te ma ceter of te ball dtrbuto kow aga. Furtermore, te ew ceter of ROI wa fted aga toward te ew ceter of ma wc leave te Mea Sft vector trace aga. Ad o o utl te ceter pot of ROI equal to te ma ceter of pere dtrbuto, o tere o more ft. T mea tat te get dety of te get dtrbuto of bllard ball a bee foud. Mea Sft Vector ceter of ma Sftg mea coderg feature pace a a emprcal PDF. Aumg tat te put for te Mea Sft a et of pot, t te ca be tougt a a ample of te uderlyg PDF. If a old area (or cluter) ext te feature pace, te t correpod to te (or maxma local) mode of te PDF. Group aocated ca alo be detfed troug te aocato wt te gve mode ug Mea Sft. Coderg eac data pot, t mportat for te Mea Sft to aocate te earet peak of te PDF dataet. Meawle, Mea Sft wll te aalyze te data pot average. Te, t move te ceter of te wdow to te average. Repeat utl coverget. At eac terato, t may be codered tat te wdow ft to a deer area of te dataet. Tu, Mea Sft work a follow:. Fx te wdow aroud eac data pot.. Calculated average data te wdow. 3. Slde te wdow toward te average. 4. -3 repeated utl coverget. F. Mea Sft Trackg For object trackg problem, color feature baed Mea Sft tecque ave bee wdely ued. Te PDF a te color probablty fucto obtaed by ug Parze wdow ug covero from RGB to dexed color. Te probablty of te color feature u of te object model qu, ad te probablty of te color feature u from te caddate model pu(), ca be expreed a Eq.(0) ad Eq.() [40]. qu C g( b ( ) u (0) ) pu ( ) C g ( b ( ) u ) () were u : color feature. : te object ceter pot te prevou vdeo frame. : te pxel locato wt te kerel te curret frame. g( ) : kerel fucto. C : ormalzato factor of te object model. C : ormalzato factor of te caddate model. b() : te color dex (..m) of. : kerel radu. qu : te probablty of te color feature u of te object model. pu() : te probablty of te color feature u from te caddate model. Te probablty of te color feature u of te object model qu, ad te probablty of te color feature u of te caddate model pu(), ometme alo expreed by te Kroecker delta fucto a te followg equato [9]: qu C Rego of Iteret (ROI) Fg. 6. Te dtrbuto of detcal bllard ball. g( b ( ) u pu ( ) C b ( ) u (Advace ole publcato: 8 Augut 08) ) b( ) u g( () ) b( ) u (3)
IAENG Iteratoal Joural of Computer Scece, 45:3, IJCS_45_3_0 Were te Kroecker delta fucto. Te relatop betwee te two PDF wc te probablty of te color feature ca be expreed by te mlarty fucto of f () or Battacaryya coeffcet f () a Eq.(4) [9] ad Eq.(5) [36]: m ) qu / pu ( ) u f f ( (4) m ) pu ( ) q u u ( (5) Te caddate model would be fted toward Mea Sft a repeatable way to maxmze te mlarty parameter. Mea Sft a local coverget terato re toward ger dety of a gve probablty dtrbuto. Te terato proce repeatedly goe utl te get dety of te optmal etmato of te object locato aceved wc dcate te object a bee tracked uccefully. Te computato of Mea Sft vector to tralate te Kerel wdow by m(), expreed a Eq.(6) [4]. were m() g g (6) m() : te dered Mea Sft vector oe terato wt te kerel. : te object ceter pot te prevou vdeo frame. : Te curret frame pxel pot de te kerel. g( ) : kerel fucto. : kerel radu. Next, to get te Mea Sft gradet baed o te lope of te dety cotour gradet. Te geerc formula for gradet lope a Eq.(7) [4]: f ( ) (7) 0 ' 0 If appled to te KDE equato a Eq.(4) te te relatop wll be obtaed a follow: p KDE p KDE D g D Settg t to 0, we get [4]: g' g' g' Tu, te Mea Sft gradet a follow [4]: g' g ' (8) Furtermore, to obta te Parze wdow mak ad t gradet agat te x-ax ad y-ax, wt varou kerel fucto ued: Uform, Tragular, Epaeckov, ad Gaua, ug te Parze_wdow fucto a Eq.(9) [43]. [km, gx, gy] = Parze_wdow (tm, lm,, g( ), grap) (9) Were km : Parze wdow mak. gx, gy : Parze wdow gradet agat te x ad y-axe. tm, lm : egt ad wdt of Parze wdow mak ze. rm : Parze wdow mak radu. g( ) : kerel fucto. grap : plot te mage mak f grap =. Fally to get te Mea Sft trackg by ug te Meaf_trackg fucto a Eq.(30) [43]. [, lo, f, f_dx] = MeaSft_trackg (q u, I,... Lmap, egt, wdt, f_tre, max_t,,... (30) tm, lm, km, gx, gy, f, f_dx, lo) were : te ceter locato of te object te prevou vdeo frame. : te pxel locato wt te kerel te curret frame. lo : te object out of trackg (f, f_dx) : te torage of te reult of mlarty fucto durg te trackg proce. q u : te probablty of te color feature u of te object model. I : ext frame. egt, wdt : ze of I. Lmap : colormap legt. f_tre : te treold value of te mlarty fucto. max_t : maxmum terato. tm, lm : egt ad wdt of Parze wdow mak ze. km : Parze wdow mak. gx, gy : Parze wdow gradet agat te x ad y-axe. III. PROPOSED METHODOLOGY Te ucce of te object trackg proce geeral or te lcee plate trackg proce, t cae, marked by te ucce of matag te value of mlarty fucto trougout te vdeo frame. I t paper, we propoe a ew metod to keep te value of te dered mlarty fucto by electg approprate kerel radu durg te trackg proce alog te vdeo frame. Te electo proce of te kerel radu follow te cage te value of te mlarty (Advace ole publcato: 8 Augut 08)
IAENG Iteratoal Joural of Computer Scece, 45:3, IJCS_45_3_0 fucto. Te tecque for obtag te approprate kerel radu by ug te adaptve fuzzy ytem adjutmet mecam. So far te ue of kerel radu tatc durg te trackg proce alog te vdeo frame. So te ucce of te trackg proce ot cotet depedg o te electo of kerel radu. T paper cotrbute maly te electo of approprate ad dyamc kerel radu troug a adaptve fuzzy mecam baed o te readg of mlarty fucto value. Te reult of te trackg proce expected to be better ta te ue of tatc kerel radu. Sce te kerel radu electo ue a adaptve fuzzy mecam ad te kerel fucto ued Gaua, t metod te called Adaptve Fuzzy Gaua kerel. Furtermore, to ae te mlarty fucto, t ould be compared to a arbtrary value, e.g., Ɛ. Value (Ɛ) coe ad compared to te mea of te output mlarty fucto oe frame of vdeo. Te value of Ɛ obtaed from te mea of mlarty fucto output from te tadard Mea Sft trackg proce ug Gaua kerel fucto. Te cage of te kerel radu are eeded to keep te value of te mlarty fxed or mproved compared wt a certa value, Ɛ. Icreag or decreag te value of te mlarty to te value of Ɛ te ba for creag or decreag te kerel radu. By Eq.(4), te value of Ɛ obtaed from tadard Mea Sft trackg proce by ug Gaua kerel fucto te ame vdeo. Wt Z a a umber of te frame, te te value of Ɛ defed a Eq.(3). Step 4: Te mea value of mlarty fucto output compared wt te value of Ɛ. If te mea value of mlarty fucto output te ame or greater ta te value of Ɛ, te te trackg proce cotued. Coverely, f t maller ta te value of Ɛ, te te dfferece value obtaed. Te dfferece value wll be ued a te error ad te delta error, ad performg adaptve fuzzy ytem proce. By ug te adaptve fuzzy mecam, te te ew kerel radu value wll be obtaed. After tat, te proce cotued ug te ew kerel radu value. Step 5: Te proce repeated utl te maxmum terato a bee reaced. Te flow dagram of te propoed metod ow Fg. 7. Start Italzato Draw te elected target te frt frame Calculato of probablty of feature u object model Z f () Z (3) Te mea value of mlarty fucto output compared wt te value of Ɛ. If te mea value of mlarty fucto output te ame or greater ta te value of Ɛ, te te trackg proce cotued. Coverely, f t maller ta te value of Ɛ, te value of dfferece wll te be obtaed. Te value of te dfferece te ued a a error ad a delta error value. Te error ad delta error value are ued a a ba for te ue of te adaptve fuzzy ytem. Wt te mecam of fuzzy rule, te te ew kerel radu value wll be obtaed. Te proce te cotued ug te ew kerel radu value. Te proce repeated utl te maxmum terato a bee reaced. Calculato of probablty of feature u caddate model Next Frame Parze wdow ug Gaua Kerel wt ew radu Mea Sft Trackg Adaptve Fuzzy Sytem Mea of Smlarty Fucto Output <? A. Algortm Step Te algortm proce explaed are: Step : Italzato of te Mea Sft trackg algortm. Oce Lcee plate object area elected, te te probablty dtrbuto of togram determed for te object model. Step : Te caddate model a te etmato determed for te curret frame. Step 3: Mea-Sft trackg proce. Te mlarty fucto betwee PDF of te object model ad te caddate model calculated. Te mea value of mlarty fucto output obtaed. Ye get error ad delta error No Lat Frame Ed Fg. 7. Flow cart propoed metodology. (Advace ole publcato: 8 Augut 08)
IAENG Iteratoal Joural of Computer Scece, 45:3, IJCS_45_3_0 B. Adaptve Fuzzy Deg Te deg of te adaptve fuzzy ytem for te mproved Mea Sft ug adaptve fuzzy Gaua kerel for vdeo-baed Idoea vecle lcee plate trackg ow Fg. 8. Fuzzfcato Te te ape of te put fuzzy memberp fucto ow Fg. 9. µ SE ME BE Defuzzfcato Iferece Ege error (a) µ SDE MDE BDE Rule Bae Output Iput delta error (b) Fg. 9. Fuzzy memberp fucto of put error ad delta error. error delta error ew kerel radu Fg. 8. Adaptve Fuzzy Sytem. Tee term repreet: mall error (SE), medum error (ME), bg error (BE), mall delta error (SDE), medum delta error (MDE) ad bg delta error (BDE), ad te Gaua fucto ued, deged a te memberp fucto. Te Gaua memberp fucto are defed by Eq.(3). SE gaua (error;, c ) ME gaua (error;, c ) BE gaua (error; 3, c3 ) SDE gaua (delta error; 4, c4 ) MDE gaua (delta error; 5, c5 ) BDE gaua (delta error; 6, c6 ) (3) Te ue of fuzzy memberp fucto parameter a Eq.(34), adaptve accordace wt error ad delta error reult of a dfferece of te mea of mlarty fucto wt a value of Ɛ. So te Gaua fuzzy memberp fucto Eq.(3) alo adaptve. Cage error ad delta error wc mea cage to te fuzzy memberp fucto reult a cage te ew kerel radu. Te goal to retur te deal value of te mea of mlarty fucto, uder te value of Ɛ (a Fg. 7). Furtermore, te ape of te output fuzzy memberp fucto deged a ow Fg. 0. µ out out out 3 85 00 Te formulato of Gaua wt te memberp fucto parameter, c ad x a error or delta error gve by Eq. (33). x c, were..6 gaua ( x;, c ) exp Were te value of parameter, c (33) 70 Kerel Radu (%) Fg. 0. Output fuzzy memberp fucto. repectvely defed a Eq. (34). = max (error); c = max (error) / 3; = max (error); c = max (error) / ; 3 = max (error); c 3 = max (error) / ; 4 = max (delta error); c 4 = max (delta error) / 3; 5 = max (delta error); c 5 = max (delta error) / ; 6 = max (delta error); c 6 = max (delta error) / ; (34) Accordg to Fg. 0, te crp value of eac out, out ad out3 defied a te gleto fuzzy output te form of a value coce of te kerel radu ug Gaua kerel (%), wc gve a Eq. (35). out = 70; out = 85; out3 = 00; (35) Te rule evaluato of adaptve fuzzy output deged a ow Table I. (Advace ole publcato: 8 Augut 08)
IAENG Iteratoal Joural of Computer Scece, 45:3, IJCS_45_3_0 TABLE I RULE EVALUATION OF ADAPTIVE FUZZY OUTPUT Error SE ME BE SDE Delta Error out3 out TABLE II MEAN SHIFT PARAMETERS Parameter Value out te treold for te mlarty fucto 0.6 maxmum umber of terato 00 MDE out3 out out BDE out3 out out Defuzzfcato calculato a fuzzy ytem output ug Ceter Average Defuzzfer [44]. Suppoe tat te output are te form of two fuzzy memberp fucto a Fg.. Were μ ad μ are te Gaua memberp fucto value of te put error ad delta error, out ad out are correpodg fuzzy output value, ad N te umber of Gaua memberp fucto proceed. Te te output of te fuzzy ytem y* ug Ceter Average Defuzzfer ca be calculated ug Eq.(36) [44]. N y * out (36) N Te followg fgure gve a overvew of te mulato reult. For Vdeo, te followg reult are obtaed, a ow Fg. troug Fg. 5. Te tadard Mea Sft geeratg percetage accuracy of object trackg of 50.4673%, wle te Mea Sft wt adaptve fuzzy Gaua kerel yeld 6.68% better accuracy. From te ample frame to 5, 50, 75,00 ow Fg. 5, frame to 5, 00 of te tet reult ug a tadard Mea Sft occur out of te trackg. Wle applyg te propoed metod, oly te 00t frame out of trackg wle te ret of te frame are very clear. It ow tat te propoed metod produce better trackg performace a compared to metod utlzg te tadard Mea Sft. I t example a defuzzfcato output obtaed: y * out Te llutrato ow Fg.. µ µ out Fg.. Smlarty Fucto Output, Mea of Smlarty Fucto Output ad Kerel Radu Plot repectvely of Vdeo Lcee Plate Trackg. out Kerel radu (%) Fg.. Sample overvew of ceter average defuzzfer. IV. RESULTS AND DISCUSSION Improved Mea Sft algortm wt Gaua kerel fucto tat a bee added Fuzzy Adaptve mecam teted o four vdeo recordg of vecle rug o te gway. Vecle ued for tetg are vecle wt lcee plate umber of Idoea. Te algortm appled for trackg te lcee plate of te vecle alog te vdeo frame. Te parameter ued for te tet are preeted Table II. Fg. 3. Ital Memberp Fucto of Error ad Delta Error for Adaptve (Advace ole publcato: 8 Augut 08)
IAENG Iteratoal Joural of Computer Scece, 45:3, IJCS_45_3_0 SE ME BE SDE MDE BDE Fg. 4. Fal memberp Fucto of Error ad Delta Error for Adaptve Fg. 6. Smlarty Fucto Output, Mea of Smlarty Fucto Output ad Kerel Radu Plot repectvely of Vdeo Lcee Plate Trackg. Fg. 7. Ital Memberp Fucto of Error ad Delta Error for Adaptve SE ME BE Fg. 5. Te reult of te proce of trackg Vdeo left ug te Statc Gaua kerel ad te rgt of ug Adaptve Fuzzy Gaua kerel. Trackg reult (from te top) of frame 5, 50, 75,00 are dplayed. For Vdeo, te followg reult are obtaed. Te mulato reult are ow Fg. 6 troug Fg. 9. Te tadard Mea Sft geerate percetage accuracy of object trackg of 76.3780%, wle te Mea Sft wt adaptve fuzzy Gaua kerel yeld 80.350% (better accuracy). From te frame ample a ow Fg. 9, o te 90t frame, te tet reult ug Stadard Mea Sft appear out of te track very clearly. It ow tat te propoed metod produce better trackg performace a compared to metod utlzg te tadard Mea Sft. SDE MDE BDE Fg. 8. Fal memberp Fucto of Error ad Delta Error for Adaptve (Advace ole publcato: 8 Augut 08)
IAENG Iteratoal Joural of Computer Scece, 45:3, IJCS_45_3_0 Fg.. Ital Memberp Fucto of Error ad Delta Error for Adaptve SE ME BE Fg. 9. Te reult of te proce of trackg Vdeo left ug te Statc Gaua kerel ad te rgt of ug Adaptve Fuzzy Gaua kerel. Trackg reult (from te top) of frame 30, 60, 90,0 are dplayed. For Vdeo 3, te followg reult are obtaed, a ow Fg. 0 troug Fg. 3. Te tadard Mea Sft geeratg percetage accuracy of object trackg 6.9630%, wle te Mea Sft wt adaptve fuzzy Gaua kerel yeld 64.848% (better accuracy). From tee reult, te dfferece ot gfcat. From te ample frame to 7, 54, 8, 08 a ow Fg. 3, te trackg reult are almot te ame. Oly te frame to 08, of te tet reult ug a tadard Mea Sft, occur out of te trackg. It ow tat te propoed metod produce better trackg performace a compared to metod utlzg te tadard Mea Sft. Fg. 0. Smlarty Fucto Output, Mea of Smlarty Fucto Output ad Kerel Radu Plot repectvely of Vdeo 3 Lcee Plate Trackg. SDE MDE BDE Fg.. Fal memberp Fucto of Error ad Delta Error for Adaptve Fg. 3. Te reult of te proce of trackg Vdeo 3 left ug te Statc Gaua kerel ad te rgt of ug Adaptve Fuzzy Gaua kerel. Trackg reult (from te top) of frame 7, 54, 8,08 are dplayed. (Advace ole publcato: 8 Augut 08)
IAENG Iteratoal Joural of Computer Scece, 45:3, IJCS_45_3_0 For Vdeo 4, te followg reult are obtaed, a ow Fg. 4 troug Fg. 7. Fg. 4. Smlarty Fucto Output, Mea of Smlarty Fucto Output ad Kerel Radu Plot repectvely of Vdeo 4 Lcee Plate Trackg. Fg. 7. Te reult of te proce of trackg Vdeo 4 left ug te Statc Gaua kerel ad te rgt of ug Adaptve Fuzzy Gaua kerel. Trackg reult (from te top) of frame 4, 48, 7, 96 are dplayed. Fg. 5. Ital Memberp Fucto of Error ad Delta Error for Adaptve From te tet reult Vdeo 4, ample frame to 4, 48, 7, 96, t ca be ee te trackg reult ug te Mea Sft tadard (left), 3 of te four ample of te object lcee plate occur lot track. Wle te trackg reult ug te mproved Mea Sft wt adaptve fuzzy Gaua kerel (rgt), all lcee plate object ca be tracked properly. It ow tat te propoed metod produce better trackg performace a compared to metod utlzg te tadard Mea Sft. Determato of trackg accuracy, t codered Percetage Accuracy of Object Trackg (PAOT) defed Eq. (37). N PAOT Object o track Frame SE ME BE N Frame x 00 % (37) Te tet reult are ummarzed Table III. TABLE III PERCENTAGE ACCURACY OF OBJECT TRACKING (%) Vdeo MS ug Statc Gaua Kerel MS ug Adaptve Fuzzy Gaua Kerel 50.4673 6.68 76.3780 80.350 3 6.9630 64.848 4 8.88 57.5758 Average 54.58 66.0970 SDE MDE BDE Fg. 6. Fal memberp Fucto of Error ad Delta Error for Adaptve (Advace ole publcato: 8 Augut 08)
IAENG Iteratoal Joural of Computer Scece, 45:3, IJCS_45_3_0 From te tet reult, te mproved Mea Sft wt adaptve fuzzy Gaua kerel gave te average percetage of 66.0% trackg accuracy. Wt te overall tral of four vdeo (00%) a "ead to ead," t a uperor performace compared to te tadard Mea Sft. V. CONCLUSION Te paper propoed a mproved Mea Sft ug adaptve fuzzy Gaua kerel for vdeo-baed Idoea vecle lcee plate trackg. Te accuracy of trackg wa determed by ug te Percetage Accuracy of Object Trackg (PAOT). Te expermetal reult owed tat te Improved Mea Sft ug Adaptve Fuzzy Gaua Kerel provded a better average of Percetage Accuracy of Object Trackg baed o te tral of four vdeo comparo wt te Stadard Mea Sft. ACKNOWLEDGMENT Te frt autor grateful to te doctoral fellowp program fuded by te Drectorate Geeral of Hger Educato, Mtry of Reearc ad Educato, Idoea. REFERENCES [] M. A. A. Azz, J. Nu, X. Zao, ad X. L, Effcet ad Robut Learg for Sutaable ad Reacquto-Eabled Had Trackg, IEEE Tra. Cyber., vol. 46, o. 4, pp. 945 958, Apr. 06. [] V. Bata, L. Marcearo, ad C. S. Regazzo, Ole Noparametrc Bayea Actvty Mg ad Aaly From Survellace Vdeo, IEEE Tra. Image Proce., vol. 5, o. 5, pp. 089 0, May 06. [3] J. Dg, Y. Huag, W. Lu, ad K. Huag, Severely Blurred Object Trackg by Learg Deep Image Repreetato, IEEE Tra. Crcut Syt. Vdeo Tecol., vol. 6, o., pp. 39 33, Feb. 06. [4] J. Meg, J. Yua, J. Yag, G. Wag, ad Y. P. Ta, Object Itace Searc Vdeo va Spato-Temporal Trajectory Dcovery, IEEE Tra. Multmed., vol. 8, o., pp. 6 7, Ja. 06. [5] S. F. Razav, H. Sajed, ad M. E. Sr, Itegrato of colour ad uform terlaced dervatve patter for object trackg, IET Image Proce., vol. 0, o. 5, pp. 38 390, 06. [6] A. S. Slva, F. M. Q. Severg, M. L. Olvera, V. M. S. Mede, ad Z. M. A. Pexoto, Object Trackg by Color ad Actve Cotour Model Segmetato, IEEE Lat. Am. Tra., vol. 4, o. 3, pp. 488 493, Mar. 06. [7] J. Su, S. Zag, ad L. Zag, Object Trackg Wt Spatal Cotext Model, IEEE Sgal Proce. Lett., vol. 3, o. 5, pp. 77 73, May 06. [8] K. R. Reddy, K. H. Prya, ad N. Neelma, Object Detecto ad Trackg -- A Survey, 05 Iteratoal Coferece o Computatoal Itellgece ad Commucato Network (CICN), 05, pp. 48 4. [9] S. P. Saoo ad S. Ar, Automated uma trackg ug advaced mea ft algortm, Commucato ad Sgal Proceg (ICCSP), 05 Iteratoal Coferece o, 05, pp. 789 793. [0] M. Owczarek, P. Barak, ad P. Strumllo, Pedetra trackg vdeo equece: A partcle flterg approac, Computer Scece ad Iformato Sytem (FedCSIS), 05 Federated Coferece o, 05, pp. 875 88. [] S. Ca, X. Zou, ad S. Ce, Ole learg for clafcato ad object trackg wt uperpxel, 05 IEEE Iteratoal Coferece o Robotc ad Bommetc (ROBIO), 05, pp. 758 763. [] Y. Yag, X. Wag, J. Wu, H. Ce, ad Z. Ha, A mproved mea ft object trackg algortm baed o ORB feature matcg, Te 7t Cee Cotrol ad Deco Coferece (05 CCDC), 05, pp. 4996 4999. [3] E. Maggo ad A. Cavallaro, Vdeo trackg: teory ad practce. Jo Wley & So, 0. [4] S. Du, M. Ibram, M. Seata, ad W. Badawy, Automatc Lcee Plate Recogto (ALPR): A State-of-te-Art Revew, Crcut Syt. Vdeo Tecol. IEEE Tra., vol. 3, o., pp. 3 35, Feb. 03. [5] Y. Sog, S. L, C. Zu, S. Jag, ad H. Cag, Ivarat foregroud occupato rato for cale adaptve mea ft trackg, IET Comput. V., vol. 9, o. 4, pp. 489 499, 05. [6] W. Yu, X. Ta, Z. Hou, Y. Za, ad Y. Yag, Mult-cale mea ft trackg, IET Comput. V., vol. 9, o., pp. 0 3, 05. [7] S. Zag, C. Wag, S. C. Ca, X. We, ad C. H. Ho, New Object Detecto, Trackg, ad Recogto Approace for Vdeo Survellace Over Camera Network, IEEE Se. J., vol. 5, o. 5, pp. 679 69, May 05. [8] N. L, D. Zag, X. Gu, L. Huag, W. Lu, ad T. Xu, A mproved mea ft algortm for movg object trackg, Electrcal ad Computer Egeerg (CCECE), 05 IEEE 8t Caada Coferece o, 05, pp. 45 49. [9] S. B. Setyawa, D. Purwato, ad R. Mardyato, Vual object trackg ug mproved Mea Sft algortm, 05 Iteratoal Coferece o Iformato Tecology Sytem ad Iovato (ICITSI), 05, pp. 7. [0] M. N. Mokt ad R. A. Salam, Hybrd of Mea-ft ad meda-cut algortm for f egmetato, 008 Iteratoal Coferece o Electroc Deg, 008, pp. 5. [] G. q. L ad Z. z. Ce, A ybrd algortm for detectg cotour of movg object baed o mergg Mea Sft ad GVF Sake model, 0 4t Iteratoal Cogre o Image ad Sgal Proceg, 0, vol. 3, pp. 87 9. [] R. He ad Y. Zu, A Hybrd Image Segmetato Approac Baed o Mea Sft ad Fuzzy C-Mea, 009 Aa-Pacfc Coferece o Iformato Proceg, 009, vol., pp. 05 08. [3] E. Maggo ad A. Cavallaro, Hybrd Partcle Flter ad Mea Sft tracker wt adaptve trato model, Proceedg. (ICASSP 05). IEEE Iteratoal Coferece o Acoutc, Speec, ad Sgal Proceg, 005., 005, vol., pp. 4. [4] A. S. Kattak, G. Raja, N. Ajum, ad M. Qam, Itegrato of Mea-Sft ad Partcle Flter: A Survey, 04 t Iteratoal Coferece o Froter of Iformato Tecology, 04, pp. 86 9. [5] Z. Ll, F. Baoje, L. Bej, ad T. Yadog, A Hybrd Trackg Metod Baed o Actve Cotour ad Mea Sft Algortm, 009 Secod Iteratoal Coferece o Itellget Network ad Itellget Sytem, 009, pp. 70 73. [6] M. Azga, A. Agagolzade, S. Gaem, ad M. Kouzegar, Itellget modfed mea ft trackg ug geetc algortm, 00 5t Iteratoal Sympoum o Telecommucato, 00, pp. 806 8. [7] Y. W. B. G. Zogda Ta Sujag L, Prorty Scedulg of Networked Cotrol Sytem Baed o Fuzzy Cotroller wt Self-tug Scale Factor, IAENG Iteratoal Joural of Computer Scece, vol. 44, o. 3, pp. 308 35, 07. [8] A. M. N. A. A. A. Hery N. Afq A. Dala ad Sumeru, Idoor Temperature Cotrol ad Eergy Savg Potetal of Splt Ut Ar Codtog Sytem ug Fuzzy Logc Cotroller, IAENG Iteratoal Joural of Computer Scece, vol. 43, o. 4, pp. 40 405, 06. [9] V. Jayalakm ad T. A. Razak, Trut Baed Power Aware Secure Source Routg Protocol ug Fuzzy Logc for Moble Adoc Network, IAENG Iteratoal Joural of Computer Scece, vol. 43, o., pp. 98 07, 06. [30] C.-T. L. S. P. J. S. Muke Praad Dog-L L ad S. Jo, Degg Mamda-Type Fuzzy Reaog for Vualzg Predcto Problem Baed o Collaboratve Fuzzy Cluterg, IAENG Iteratoal Joural of Computer Scece, vol. 4, o. 4, pp. 404 4, 05. [3] Kall Kaba ad S. R. Agabozorg, Adaptve Tme-Varat Model Optmzato for Fuzzy-Tme-Sere Forecatg, IAENG Iteratoal Joural of Computer Scece, vol. 4, o., pp. 07 6, 05. [3] X. Hog, Patter Recogto, Mac. Itell., 0. [33] M. Roeblatt, Remark o Some Noparametrc Etmate of a Dety Fucto, A. Mat. Stat., vol. 7, o. 3, pp. 83 837, 956. [34] E. Parze, O Etmato of a Probablty Dety Fucto ad Mode, A. Mat. Stat., vol. 33, o. 3, pp. 065 076, 96. [35] R. Guterrez-Oua, Patter Aaly, CSCE 666, 03. [36] D. Comacu, V. Rame, ad P. Meer, Kerel-baed object trackg, IEEE Tra. Patter Aal. Mac. Itell., vol. 5, o. 5, pp. 564 577, May 003. [37] K. Ce, S. Fu, K. Sog, ad C. G. Ju, A Meaft-baed mbedded computer vo ytem deg for real-tme target (Advace ole publcato: 8 Augut 08)
IAENG Iteratoal Joural of Computer Scece, 45:3, IJCS_45_3_0 trackg, Computer Scece Educato (ICCSE), 0 7t Iteratoal Coferece o, 0, pp. 98 303. [38] K. Fukuaga ad L. Hotetler, Te etmato of te gradet of a dety fucto, wt applcato patter recogto, IEEE Tra. If. Teory, vol., o., pp. 3 40, Ja. 975. [39] D. Comacu ad P. Meer, Mea ft: a robut approac toward feature pace aaly, IEEE Tra. Patter Aal. Mac. Itell., vol. 4, o. 5, pp. 603 69, May 00. [40] Y. Ukratz ad B. Sarel, Mea Sft Teory ad Applcato, Wezma It. Sc. ttp//www. wdom. wezma. ac. l/~ vo/coure/004_/fle/mea_ft/mea_ft. ppt, 004. [4] K. Ce, S. Fu, K. Sog, ad C. G. Ju, A Meaft-baed mbedded computer vo ytem deg for real-tme target trackg, 0, pp. 98 303. [4] S. Trumurugaata, Itroducto To Mea Sft Algortm, 00. [43] S. Berardt, Mea-ft vdeo trackg. 06. [44] L.-X. Wag, A Coure Fuzzy Sytem ad Cotrol, Lodo Pretce-Hall It. Ic., 997. Bauk Ramat, He a lecturer at bacelor degree Uverta Pembagua Naoal Vetera Jawa Tmur. He receved te bacelor degree Itrumetato Pyc from Ittut Tekolog Sepulu Nopember Surabaya 995. He receved a mater degree Itrumetato ad Cotrol from Ittut Tekolog Badug, 000. Curretly, e a P.D. caddate Electrcal Egeerg at Ittut Tekolog Sepulu Nopember, Surabaya. H reearc teret are a tellget ytem, oft computg, mage ad vdeo proceg, tellget cotrol, MATLAB, PHP, Pyto ad Delp Programmg. Edra Joelato, receved te bacelor degree Egeerg Pyc from Ittut Tekolog Badug (ITB), Idoea 990. He receved P.D. Egeerg, from Te Autrala Natoal Uverty (ANU), Autrala 00. Curretly, e te taff of Itrumetato ad Cotrol Reearc Group, Faculty of Idutral Tecology, Ittut Tekolog Badug (ITB), Idoea ad Reearc Profeor at Cetre for UMaed Sytem Stude (CetrUMS), ITB, Idoea. H reearc teret are Hybrd/Dcrete Evet Cotrol Sytem, Advaced Cotrol, Embedded Cotrol Sytem ad Itellget Sytem. He a member of IEEE. I Ketut Eddy Purama, receved te bacelor degree Electrcal Egeerg from Ittut Tekolog Sepulu Nopember (ITS), Surabaya, Idoea 994. He receved Mater of Tecology from Ittut Tekolog Badug, Badug, Idoea 999. He receved a P.D. degree from te Uverty of Groge, Neterlad 007. Curretly, e te taff of Electrcal Egeerg Departmet of Ittut Tekolog Sepulu Nopember, Surabaya, Idoea. H reearc teret are Data Mg, Medcal Image Proceg, ad Itellget Sytem. Maurd Hery Puromo, receved te bacelor degree from Ittut Tekolog Sepulu Nopember (ITS), Surabaya, Idoea 985. He receved M.Eg., ad P.D. degree from Oaka Cty Uverty, Oaka, Japa 995, ad 997, repectvely. He a joed ITS 985 ad a bee a Profeor ce 003. H curret teret clude tellget ytem applcato, mage proceg, medcal magg, cotrol, ad maagemet. He a Member of IEEE ad INNS. (Advace ole publcato: 8 Augut 08)