Indonesian Journal of  Electrical  Engineer ing and  Computer Science   V o l. 10 , No . 3, Jun e   20 18 , pp . 98 9 ~ 99 ISSN: 2502-4752,  DOI: 10. 11591/ij eecs.v10 .i3.pp989-999          9 89     Jo urn a l  h o me pa ge : http://iaescore.c om/jo urnals/index.php/ijeecs  A New Method for Ball  Tracking Based on  α - β , Linear Kalm an  and Extended Kalman Filters  Via Bubble Sort Algorithm       Ha thi r am  Ne na va th,  R a vi   Kum a Ja to th   Department o f  Electronics  and C o mm unication Engineer ing, Na ti onal Inst itut e  of   Tec hno log y , Warangal, Ind i     Article Info    A B STRAC T Article histo r y:  Received  Ja n 19, 2018  Rev i sed   Mar  27 , 20 18  Accepted Apr 11, 2018      Object  tracking  is one of th e ch allenging  issues in computer vision  and vid e o   processing, whic h has several pot enti al a ppl ica tio ns. In this paper,  initia ll y,  a   moving object is selected b y  fr ame  differen c in g method and extracted the  object b y  segment threshold i ng.  The  bubble sort algorithm (BSA) arranges   the regions (l arg e  to sm all) to m a ke sure  th at th ere is at least one big region   (object) in  object detection pro c ess.  To track  th e object, a motion model is  construct e d to set the s y stem  models of Alph a-Beta ( α - β ) filter, Linear   Kalm an filte r (L KF) and Extend ed Kalm an filt er  (EKF). Man y  e xperim e nts  have been cond ucted on balls  with di fferen t  s i zes  in im age s e quences  and   compared th eir  tracking p e rf ormance in  no rmal light and  bad light  conditions . The  parameters obtained are  th e root  m ean s quare err o r (RM S E),  absolute er ror (AE), obje c t tr ac king e rror (OTE), Tr acking d e tection rate  (TDR), and peak signal-to-noise ratio (PSNR) an d they  ar e compared to fin d   the  algorithm th at p e rforms the b e st for  two cond itions.   K eyw ords :   Back gr oun d sub t r action     Bu bb le so rt al go rith m    EKF   LKF   Object trac king e r ror    α - β  filter    Copyright ©  201 8 Institut e  o f   Ad vanced  Engin eer ing and S c i e nce.  All rights re se rve d Co rresp ond i ng  Autho r Hath iram  Nen a v a th   Depa rt m e nt  of  El ect roni cs  an d C o m m uni cati on  En gi nee r i n g,   Natio n a In stitu te of Techn o l o g y   Warang al,  W a r a ng al, Ind i a 50 600 4.  Em a il: h a th iram . iisc@g m ai l.co m       1.   INTRODUCTION  C o m put er  vi si on  i s  t h e  a dva ncem ent  i n  sci e nce a n d t ech n o l o gy  t h at  i n cl ude  m e t hods  f o r  acq ui ri ng ,   pr ocessi ng , an al y z i ng, an d u nde rst a n d i n g i m ages. As a sci e nt i f i c  di sci p l i n e, com put er  vi si on i s  fret f ul  wi t h   th e th eory b e hin d  artificial syste m s th at ex cerp t  inform at ion  fr om  im ages. The i m age dat a  can t a ke  vari ous   fo rm s, such as vi deo se qu e n ces, vi e w s fr om   m a ny  cameras, o r  m u l t i - di m e nsi onal  d a t a  from  a  m e di cal   scann e r.  As a  scien tific d i scip lin e, co m p u t er v i sion  pursu es to  relate its t h eories and  mo d e ls to  t h e d e sig n   of  co m p u t er v i si o n   syste m s. Man y  research  ex ten d s to   in clu d e u s es  in  security and  su rv eillan ce, v i d e com m uni cat i on an d c o m p ressi on , Na vi g a t i on,  di spl a y  t echn o l o gy Hi g h -Le v el   V i deo  Anal y s i s ,  t r affi c   co n t ro l, Metro l og y, v i d e o ed itin g ,  augmen ted  reality an d  Hu m a n - C o m p u t er in terfaces to  med i cal     im agi ng [ 1 - 2 ] .   Gi ven, t h e m o st  im port a nt  st at e (e.g.,  po si t i on an vel o ci t y ) of a t a rget   o b ject  i n  t h e l e a d i n im age, the objective of tracki ng is to  estim a te the states  of the target  i n  the su bse que nt  fram e s. Even t h o u g h   object tracki n g has bee n  studied for m o re than a fe dec a des an d si g n i f i cant  ev ol ut i o n has  been m a de i n   recent yea r s [3-7], it  rem a in s a challenging  problem .           Ob j ect illu st ratio n  is  on of th k e y wo rks in  an p i ctorial track ing  algo rith m ,  an d man y  ou tlin es  have  been  pr o pos ed [ 8 ] .  Si nc e t h e pri m ary   wo rk  of L u cas  and  Kana de [ 9 ] ,  based o n  t h e  raw co ncent r a t i on o f   val u es  ha ve  be en  br oadl y   use d  f o r t r ac ki n g   [1 0] . T h e L u c a s an Kana de  ap pr oac h  [ 1 1] , d o   n o t  t a ke  great   occurre nce c h angea b ility into account and thus, do  not  succee d due t o  the  dissim i l a rity of the c h rom a tic   p r op erties  o f  a targ et  o b j ect.  Matth ews et al. [12 ]  estab lish e d a tem p late  up d a te m e th od   b y  exp l o iting  th inform ation of the first fram e  to accurate points.  Towa rds  better acc ount  for c h anges i n  the  presence  of t h e   ob ject , s u bs pa ce-base d t r a c k i ng a p pr oach   [7 ,1 3]  ha ve  b een  pr op ose d .  In  [ 14] Ha g e r an d B e l h u m eur  Evaluation Warning : The document was created with Spire.PDF for Python.
                        I S SN 2 502 -47 52  I ndo n e sian  J Elec Eng  & Com p  Sci, V o l. 10 No 3 ,  Jun e   2 018  :   98 9 – 99 99 0 recom m ended  a pr o f i c i e nt  L u cas a n d  Ka n a de al g o r i t h m  an used  l o di m e nsi onal  desi g n s  f o r t r acki n g   ch ang i ng  illu min a tio n  co nditio n s . Sev e ral  tech n i qu es h a v e  b e en  an ticip ated  fo v i d e o  ob j ect seg m en tatio n.  C o m m onl y ,  t h ese m e t hods c a be a p p r o x i m at ely  cl assi fi ed i n t o  t w o t y pes  [1 5,  1 6 ] :  back g r o u n d  st r u ct u r base d vi de o ob ject   se gm ent a t i on   an d f o re gr o u n d   e x t r act i o n base d vi de o ob ject   se gm ent a t i on .   In  recent tim es , a num ber of t r acki ng m e thods bas e on s p arse re prese n ta tions ha ve  bee n  propose d   [1 7- 2 1 ] .  To ex pan d  t r ac ki n g  ro b u st ness , a l o cal  spars e  pre s ence m odel  was pr op ose d  i n  [22]  usi ng t h e   m ean  sh ift algo rith m  to  d e tect obj ects. By th e p r esu m p t u o u s rep r esen tatio n   o f   particles as to g e th er,  H. Lu  et al [17 ]   ex pressed   o b j e ct track ing  as  m u lti-task  sparse learn i ng  prob lem .  J. Fiscu s , J. Garo fo lo, T. Rose, an d M.  M i chel , [2 3]  p r o v i d es a n  e x p e ri m e nt al  evaluat i on  o f  seve r a l  t r acki n g al g o ri t h m s  on t h e  adva nce d  vi d e o a n d   sig n a b a sed  surv eillan ce, in  a con f eren ce on adv a n c ed   v i d e o  an d sign al  b a sed  su rv eillan c e, using   d a taset for  fo llowing   v a ri o u s   p e op le. Th e effort of su ch   rev i ews  i s  o f   b e tter-quality st ill, as i n  [2 4 ]   wh ich   d i scu ss  tracking speci fic targets only, for exam pl e, sports players.  The survey of J.  C. McCall  a n d  M. M. Tr ived i [ 2 5 ]   is on  track i ng   p a th s fo r driv er aid.  In this pa pe r,  initially, a  m o ving object is  se lected by fram e  differenc ing m e thod and extracte d   ob ject  by  seg m ent  t h resh ol d i ng. T h e b u bbl e sort  al go ri t h m  (B SA) ar ra nge s t h e regi o n s (l ar ge t o  s m al l )  t o   make sure that  there is at leas t one  big re gion (object ).    To   track  the obj ect, a m o tio n  m o d e l is con s tru c t e d  t o   set th e syste m   m o d e ls o f  th α - β  filter, LKF an d  EKF. Man y  exp e rim e n t s h a v e   b e en  con d u c ted   on  b a l l s with  diffe re nt sizes  in im age sequences a n d com p are d  thei t r a c ki n g   per f o r m a nce i n   n o rm al  l i ght  an bad  l i ght   co nd itio ns. The p a ram e ters o b t ain e d  are RMSE, AE, OTE, TDR ,  and p eak  si g n a l - to -no i se ratio  (PSNR ) Ex peri m e nt al  resul t s  i ndi cat e  t h at  t h e pr op ose d  al go ri t h m s  per f o r m  wel l  for t h e det ect i on an d t r ac ki ng  of   di ffe re nt  si zes  of  bal l s  i n   vi de o se q u ences  ca pt u r ed  by   usi n g a  st at i c  i p h o n e6  cam era. T h or ga ni zat i on  of  t h e   pape r i s  as fol l ows .  The m ovi ng  ob ject  det e ct i on an d t r ac k i ng m e t hods ar e pr op ose d  i n   Sect i on  2. Sect i on 3  prese n t s  e x peri m e nt al  resul t s   and  di sc ussi on s. Fi n a l l y , t h e c oncl u si o n  i s   gi ven  i n   Sect i o 4.       2.   THE PROPOSED  METHOD  2. 1.    Sim p le Back ground  Su btr a ction   Sim p l e  B ackg r o u nd  Su bt ract i on  (SB S ) i s  t h base f o r se veral   devel o p m ent s  i n  o b j e c t  det ect i on    [2 6- 2 7 ]  w h i c h  det ect s m ovi ng  o b j ect s thro ugh  calcu latin g th e ab so lu t e   differe n ce  between t h e re ference   im age   ,  [ 2 8]  a n d i n com i ng  vi d e fram e  , . Notice that t h refe rence  im age   ,  mu s t   h a v e   o n l y static b a ck gro und ; th e in co m i n g   f r a me f eatu r es po ssib l e m o v i ng   o b j ects al on g w ith b a ckg r ound  inf o rm ation. A f ter the  refe re nce im age   ,  an d t h e i n c o m i ng vi deo  fram e   ,  are taken  from a   vi de o se q u ence , t h det ect ed  b i nary  m o t i on  d e t ect i on m a sk  , is calcu lated  as  fo llows:     ,  1; if| ,  , | 0; if| ,  , | τ                                                                          (1)     whe r e,  τ  i s  an e m pi ri call y  a sel ect i on o f  t h res hol d,  whi c h i s   use d  t o   di st i n g u i s pi xel s   of  m ovi ng o b j ect s fr om   th o s e of   th e back gro und   in   an im age frame. The  existe nce of m oving   objects is i ndi cated if the  absolute   diffe re nce bet w een t h e re fe rence im age   ,  and t h e i n c o m i ng  vi de fram e   ,  go  b e y o nd   τ  these  pi xel s  of   t h e de t ect i on  m a sk  ,  are th en  lab e led   with   1  an d tho s e th at  are  non-active are la beled  with  0.      2. 2.   Bubble S o rt Algorithm   Th bu bb le sort algo rith m  is b a sed   on  a co m p ariso n  of  consecutive  ele m ents of a li st. It  uses  a n   ob ject   det ect i on p r ocess. B u b b l e  so rt   alg o rit h m  arran g es th e regi on s (lar g e  to  sm all) to  mak e  sure th at  th ere is  at  l east  one bi g re gi o n  (t ar g e t  ob ject ) i n  o b ject   det ect i o n  pr ocess.  It  pe rf orm s  a seri es of  passes o v e r t h e   sequence:   a.   Com p are each ele m ent (not  includi ng the l a st one with  i t s neighbour to the  ri ght-if they are  out of  order, swap t h e m . Once t h largest  elem ent is reache d , it  kee p s on  swa ppi ng until it goe s to t h e last   p o s ition .   b.   Com p are each ele m ent (not includi ng t h e last two)  with  its neighbour to the ri ght-i f  they are out of  order, swap t h e m . Once the  second  largest e l e m ent is reached, it ke e p on swappi ng  until it goes to the  secon d  last po sitio n .  Th is  p r o c ess con tinu e un til th ere are  no   u n s orted  elemen ts are left.          Evaluation Warning : The document was created with Spire.PDF for Python.
In d onesi a n  J  E l ec En g &  C o m p  Sci    ISS N :  2 5 0 2 - 47 52       n e w method  fo r b a ll  tra cki ng  b a s ed  on   α - β , l i n ea r K a l m a n   an Ext e n d e d  ( H at hi r a N e nav at h)   99 1 2. 3   Mathematica l  Mo delling  Th e m a in  o b j e ctiv e o f  obj ect track ing  is t o   esti m a te  the state trajectories  of  an  o b ject  a  m ovi ng  o r   m ovabl e o b jec t . It  i s  wel l - kn o w n t h at  a  poi nt  m ovi ng i n  o u r  2D  space  can  be  descri bed  b y  i t s  2D  posi t i on  a n d   v e lo city v ect o r s.    X k x , y , x , y                                                                                  (2)                                                       For instance, “ E quation (2)” can be us ed as  a state vector of suc h  a poi nt  in  th e Cartesian co ord i n a te syste m ,   w h er  de not e s  m a t r i x  t r a n s pos e,   x  and  y  a r e t h e   posi t i o n   coo r di nat e s  al on g   x, y  axes res p ectively, a n x and  y  are the  ve locity vectors ,   and   ω  is th e ang u l ar  v e lo city.  Th e system  u s ed  in th is  work is m o d e lled   u s ing  th e “constan t  v e lo city” (CV) m o d e l, o r  m o re p r ecisely “n early-con stan t-v e lo city   m o d e l”. Th Process  state an d d y n a mic  m easu r emen t m o d e ls are  assu m e d  to   b e   lin ear in d i screte ti m e . It can   b e   written  as,    x  A x w                                                                                        (3)     z H x v                                                                                              (4)     whe r x  is th N 1  vector state at discrete tim k  ,  z  is th M 1  vector m easure m ent at discrete-time   k H  is the consta nt m easurem en m a trix of size   M N A  is th e con s tan t  state tran sitio n  m a trix  o f  size   M N w  is an   N 1   zero  m ean Gaussian distri bute d   proces sing(or s y ste m ) noise vector at tim e    h a vi ng  constant c o vari ance m a trix  of    , and  is an      1   z e ro m ean ra ndom   m easurem e n noise  vect or at tim e    h a v i ng   a co nstan t  po sitiv defin ite co v a rian ce m a trix  of  ,    are  ass u m e d to  be  indepe ndent  of eac ot he r f o r al l     and  It  i s  ass u m e d t h at  t h peri od  bet w een ea ch  observation  is a constant   s econds .     2. 4     Trac king Al gori t hms   2. 4. α - β  filtering  The  α - β  t r ac ki ng  [ 2 9]  al go ri t h m  can be e x pr essed a s  f o l l o w s . P r edi c t i o n st ep  (t im e up dat e ):     x  x ∆ s                                                                 (5)     s   s                                                                                                            (6)     whe r e t h ve ctor x and s x  are th p r ed icted   p o s ition  m a trix   an pred icted  v e lo city matrix respectively. Correction st ep (measurem ent update ):    x  x  α z x                                                                                               (7)     s s  z  x                                                                                               (8)     whe r e the  vect or   ,  ̂   are th e esti m a ted  p o s itio n  m a trix  and  es ti m a ted  v e lo city  matrix , resp ectiv ely,  α  and  β   are tun i ng   co nstan t s b e t w een  nu m b er  0  and   n u m b e r 1  to  sm o o t h  t h e po sition  and  v e l o city esti mates,   respectively.     2. 4. 2   Linea r  Ka lman Filter  Th e LKF  is  a ti m e -d o m ain   recu rsi v e filter with   th e cap a bilit y to  estim a t e th e state  o f   a d y n a m i sy st em  by  usi ng a se ri es o f   m easurem ents. C o nsi d e r i n g  t h e sy st em   model  e quat i o ns  “Equat i on  (3 )  and   “Equ atio n (4 )”; th e Lin e ar  Kal m an  filter equatio n s  are  g i v e n   b y  [30 ]   P   A P  A Q ,                                                                                                 (9)     K P  H HP  H R                                                                                   (10)     x  A x                                                                                                       (11)     x x  K z H x                                                                                        (12)     P                                                                                                 (13)     Evaluation Warning : The document was created with Spire.PDF for Python.
                        I S SN 2 502 -47 52  I ndo n e sian  J Elec Eng  & Com p  Sci, V o l. 10 No 3 ,  Jun e   2 018  :   98 9 – 99 99 2 fo  1 , 2 , , wh er  is th e i d en tity m a t r ix   is th a prio ri  estim a t e of the  state   gi v e n m easurem ent s   u p  to and  i n clud ing  tim e    1 .    is  th a post e ri or i  estim a te of the state   specifi ed m easurem e n ts  up t o   and  i n cl udi ng   t i m e   .    is t h Kalm an  g a in   is t h e c o varia n ce  of t h a prio ri  estim ation e r ror     ,  and   is th co v a rian ce  o f  th a po steri o ri  estim ation error    .  Th Kalm an  filter  is   in itialized  b y   x  E x                                                                                                                (14)     P  E  0  0   0  0                                                                                       (15)     whe r  .  is the e xpectation  ope r ator.    2. 4. 3   Extended Kal m an filter   The EKF 31 is t he n onlinea r v ersion o f t h e Ka lma n f ilt e r. T he n o n l i n e a r m o d e l s f o r E K F  are   gi ve n by ,     Process M odel :    x  f x w                                                                               (16)     Me asurement Model:                                                                      (17)        Th e al g o rith m for t h e ex tend ed Kalm an  fi lterin g  is esse n tially si milar to  th at  of Linear  Kalm an  filterin g except  f o l l o wi ng   m odi fi cat i ons,     F     | x                                                                                   (18)     H    | x                                                                                         (19)       3.   EX P E R I M E NT S   3. 1.   Impleme ntation de tails  In  th is  wo rk we i m p l e m en ted  o u r track e r i n  MATLAB R 2 0 1 3 a  PC m ach in e with   In tel i7-377 0  CPU  (3.4GHz)  with 2GB m e m o ry, wh ich   ru n s  29fp s  on  t h is  p l atform . In  add itio n, th e self-m a d v i d e o  con s istin of J P EG im age sequence  wi th  720 x128 0  pi xel s   per  fram e  i s  t a k e n usi ng t h e i P h one 6 cam era. Th e   envi ro nm ent  for t h e ba d l i g ht  con d i t i on i s  creat ed by  swi t c hi n g  o ff a l l  t h e li ght s i n  t h e room  (im p r o per  illu m i n a tio n ) wh ile th e env i ron m en t is co n s id ered  as a  n o rm al  lig h t  co nd itio n   wh en th e roo m   is p r op erly   illu m i n a ted .     3. 1. 1   Experimental Setup     The dista n ce c onsi d ere d  for tracki ng t h e target obj ect is a p proxim a tely  6m eters with target objects  bei n g bal l s  of  di ffe rent  si zes (sm a l l ,   m e dium  and bi g)  wh ose ra di i  are  3. 32 5c ms, 4. 9cms an 8 . 5cms respect i v el y .  T h vi de o ca pt u r ed  usi n g i P ho ne6  as  f o llo wi n g  sp ecification s : Im ag e Ty pe: RGB co lou r  sp ace,  Im age Dim e nsi on:   720x1280   pi xel s  ( H ei g h t =  1 2 8 0   and  Wi dt h= 7 2 0 ) , F r am e R a t e :  29 f p s,  Nu m b er of   fram e s: Fo r No rm al Lig h t  co n d ition  (For small s i ze b a ll  =5 4   fram e s, Fo r m e d i u m  siz e  b a ll= 6 0  frames, For  bi g si ze bal l = 5 8 f r am es) and F o r B a d Li ght  c o n d i t i on  (O nl y  for sm al l  si ze  bal l = 70  fram e s). Tabl 1 sh o w s t h e   in itial v a lu es  of th e b a sic  p a rameters for th e t r ack i n g algo rit h m s   3. 2   Perfor mance Measures   a.   Ab so lu te erro r (AE):   AE is th e m a g n itu d e   o f  t h e d i fference b e tween  th e tru e   v a lu e and th e track ed   v a lue  of the  object.     ϵ | x  x   |                                                                                                            (20)                                          whe r x  t h t r ue val u e o f  ob ject   pa ram e t e rs  and  x    is th e track e d   v a lu of th o b j ect  p a ram e t e rs.  b.   Ro o t  Mean   Squ a re Error(RM S E): RMSE is  o n e   of th m o st wid e ly u s ed  fu ll-referen ce  qu ality assess men t   m e t r i c , whi c h i s  com put ed by  squ a re r oot  o f  t h e avera g e o f   squ a re d i n t e nsi t y  di fferen ces bet w ee n t r acke d   ( x  ) an d t r ue i m age  pi xel s   ( x  ).   Evaluation Warning : The document was created with Spire.PDF for Python.
In d onesi a n  J  E l ec En g &  C o m p  Sci    ISS N :  2 5 0 2 - 47 52       n e w method  fo r b a ll  tra cki ng  b a s ed  on   α - β , l i n ea r K a l m a n   an Ext e n d e d  ( H at hi r a N e nav at h)   99 3   RMSE  ∑∑ x   x                                                                            (21)                                                       whe r N a n M are the  im age dim e nsions.    c.   Peak  Si g n al - t o N o i s e R a t i o  (PS N R ) :   The   val u o f  Pea k   si gnal - t o - n oi se rat i o  ca n easi l y  be  obt ai ne by   usi n g t h e m ean sq uare d e r r o r.     PSNR 1 0 l og                                                                                            (22)     M ean s qua re e r r o r:     MSE  ∑∑ x   x                                                                        (23)                                                                                        whe r e,  L is  the dynam i c range  of the im age (8   bits pe pixel gray scale i m ages L =  255).    d.   Tracki n g  d e t ect i on  rat e  (T DR ):  Trac ki n g   de t ect i on rat e  is  th e ratio of a  n u m b e o f  frames in  wh ich  th o b j ect is  d e tected  to th e t o tal  n u m b e o f  frames in  wh ich th e object pres e n t.     TDR                   100                                                                         (24)                                                  e.   Object trac king error  (OTE): Object  track i n g  error is the no rm al in co nsisten c y in  th e cen tro i d   of t h t r acke d   ob ject   fr om  i t s  t r ue va l u e.  It  i s  gi ve by ,     OTE                                                                                        (25)     whe r e,   x   and  y   are the act ual 2D coordinates  of the  object,  x    and  y    are t h e trac ked  2D  coordinates of the  object.      Tabl e 1. Si m u lat i on param e t e rs  f o T r acke r s   α - β  LKF   EKF   Para m e ters  Valu es  Para m e ters   Valu es  Para m e ters   Valu es  α  0. dt dt β  0. 005     1 0 0 0 0 1 0 0 0 1 0 0 0 1 dt dt     1 0 0 0 0 1 0 0 0 1 0 0 0 1 dt dt       0 0 1 0 0 0 0 1     0 0 1 0 0 0 0 1   --  --     0055 . 0 055 . 0 0143 . 0 1524 . 0     0045 . 0 0045 . 0 0045 . 0 2845 . 0   --  --     ) 4 (   01 . 0 eye   ) 4 (   01 . 0 eye   --  --     ) 4 (   100 eye   ) 4 (   100 eye       3. 3     Experimental results   3. 3. 1   Tra c k i ng  res ult s  of  s m a ll s i ze   ba ll in no rma l  lig ht co nditio The RMSE of the  LKF,  α - β  filter an d EKF  for sm all size b a ll d a taset are  g i v e n  in Tab l 2 .  As  obs er ved i n  Ta bl e 2, t h e a v er age val u e of t h e R M SE was r e duce d  aft e r appl y i n g  t h e L K F t o  sm al l bal l  dat a On a n  a v erage ,  the m i nim u m   RMSE reducti on  was  0. 03 30 an d th e m a x i m u m  w a s 0 . 447 8   fo r  sm all si ze b a ll   d a ta. Th is  resu lt ind i cates th at th e pred icted   ou tpu t  of  t h e L K F is cl oser t o  the   desired output. The bes t   avera g e values  are em phasized in Ta ble  2.  Sim ilarly, the  PSNRs  of t h e LKF,  α - β  filter and  EKF fo r s m all  si ze bal l  dat a s e t  are  gi ve n i n   Tabl 2.  As  o b s erve d i n  Ta bl e 2,  t h e  ave r ag e val u e   of t h PSNR   was  i n c r ease d   after app l yin g  th e LKF to  sm a ll b a ll  d a ta. On   a n  average, the m a xim u PSN R w a s 48 .4 4d an d  the  m i nim u m  was 33 .1 0 d B  f o r s m al l  si ze bal l   dat a . Fi g u r e1  s h ows t h e error of ce ntroid , er ro r o f  ra diu s RM SE,  cent r oi d a n d  r a di us  of  x - y  c o o r di nat e s,  AE  and  OT E i n  e ach f r am e for  sm al l  si ze bal l  base d o n  L K F u nde r   n o rm al lig h t  co nd itio ns. Fi g u re 1 shows the erro r of cen t ro id , error  o f   rad i u s , RMSE, cen t ro id  and  rad i u s   of  x-y  c o or di nat e s, A E  a n d  O T E  i n  eac fram e  fo r sm al l  si ze bal l  based  o n   L K u nde n o r m al  l i ght  co ndi t i ons.   Evaluation Warning : The document was created with Spire.PDF for Python.
                        I S SN 2 502 -47 52  I ndo n e sian  J Elec Eng  & Com p  Sci, V o l. 10 No 3 ,  Jun e   2 018  :   98 9 – 99 99 4 The OTEs of  t h L K F,  α - β  fi l t e r and E K f o r sm al l  si ze bal l  dat a set  are gi ve n i n  Ta bl e  2.  As o b se rv e d  i n   Table  2, the  average  val u e of  the OT was reduce d  a f ter  ap p l ying  th e LKF t o  sm all b a ll d a ta. On  an   av erag e,  th e m i n i m u m   OTE  red u c tion was  0 . 04 76  an d th e m a x i mu m  was 1 . 35 62  for sm all siz e  b a ll d a ta. This resu l t   in d i cates th at th e pred icted  ou tpu t  of th e LKF is clo s er  to th e d e sired   o u tp u t . Sim ilarly,  th e AEs  of the LKF,  α - β  filter and   EKF fo r sm all  size b a ll  d a taset are  g i v e n  i n   Tab l 2 .   As observ e d  i n  Tab l e 2 ,  th e av erage v a l u of t h e A E  w a s  red u ce d aft e r  ap pl y i ng t h e   LKF t o  sm all  ball data.  On  an a v era g e, t h e minim u m  AE was   0. 12 7 6  a n d  t h e  m a xim u m  was 0. 9 8 9 8   f o sm al l  si ze bal l  dat a         Fi gu re  1.  The  e xpe ri m e nt al  resul t s  o f  sm al l  bal l  t r acki n usi n g  L K F       3. 3. 2   Trackin g  resu l ts of me dium  siz e  ba ll in norma l lig ht co nditio The OT Es of the LKF,  α - β  filter an d  EKF for m e d i u m  si ze b a ll d a taset are g i v e n  in   Tab l e 2 .   As  obs erved in Ta ble  2, t h e a v erage  value  of t h e OT wa red u ced  aft e r a ppl y i ng t h e L K F  t o  m e di um  si ze bal l   dat a O n  a n  a v erage ,  t h e  m i nim u m  OTE re d u ct i o n   was  0. 0 5 9 1  a n d t h e m a xi m u m  was 0 . 69 7 2   fo r m e di um  si ze  ball data. The  best avera g e values are em phasized  in  Tab l e 2 .  Similarly,  th e AEs of th e LKF,  α - β  filter an EKF  f o r  m e di um  si ze bal l  dat a set  are  gi ve i n  Ta bl e 2 .   As  obs er ved  i n  Ta bl e 2 ,  t h e a v er age  val u e  o f  t h e A E   was redu ced  after app l yin g  the LKF to  sm all b a ll d a ta.  On   a n  a v er ag e ,  th min i mu m A E   w a s  0. 214 6  and  th max i m u m  was 1 . 25 19   fo small size b a ll d a ta. Th is resu lt  in d i cates th at t h p r ed icted   ou tpu t  of t h e LKF i s   cl oser  t o   t h e  de si red o u t p ut .       Tabl 2.  A v era g param e t e rs com p ari s on  f o r  di f f ere n t  si ze  of  bal l s   usi n f i ve t r ac kers  u n d er   n o rm al an d   b a d  ligh t  cond ition s   In Norm a l  light co ndition   In bad light conditi on   Para m e te rs  For s m all  size ball   For m e diu m   size  ball  For big size ball   For s m all  size ball   α - β    LKF   EKF   α - β    LKF   EKF   α - β    LKF   EKF   α - β    LKF   EKF   AE   0. 9 0. 12   0. 76   0. 7 0. 21   1. 25  1. 31   0. 63   1. 34  0. 45   0. 11   1. 05   RMSE  0. 4 0. 03   0. 17   0. 2 0. 04   0. 28  0. 45   0. 21   0. 30  0. 09   0. 02   0. 17   TDR (% )   96. 100  100   91. 100   91. 6  93. 1   100   93. 1  85. 7   100   85. 7   OTE   1. 3 0. 04   0. 16   0. 6 0. 05   0. 29  1. 01   0. 48   0. 29   0. 22   0. 02   0. 14     0 10 20 30 40 50 60 -1 -0 . 8 -0 . 6 -0 . 4 -0 . 2 0 0. 2 0. 4 F r am e N u m ber D i s t anc e ( i n  m e t e r s ) E r ror  of  k a l m an f i l t er c ent r o i d   V s  F r am N u m b er 0 10 20 30 40 50 60 -0 . 4 -0 . 3 -0 . 2 -0 . 1 0 0. 1 0. 2 0. 3 0. 4 0. 5 F r am e N u m b er D i sta n ce ( i n  m e t e r s ) E r r o r  of   k a lm an   f i l t er   r a d i u s  V s  F r am e N u m b er 0 10 20 30 40 50 60 0 0. 05 0. 1 0. 15 0. 2 0. 25 0. 3 0. 35 F r am e N u m b er er r o r R o ot   M e an S q u a re err o r     RM S E 0 10 20 30 40 50 60 0 100 200 300 400 500 600 700 800 900 F r am e N u m b e r c e nt r o i d   an d rad i u s  x - y  c o adi na t e s A c t ual  c oordi nat es     ce n t r o i d r a di us 0 10 20 30 40 50 60 0 0. 1 0. 2 0. 3 0. 4 0. 5 0. 6 0. 7 0. 8 0. 9 F r am e N u m b e r Er r o r A b s o l u t e  e rro r (A E )     AE 0 10 20 30 40 50 60 0 0. 0 2 0. 0 4 0. 0 6 0. 0 8 0. 1 0. 1 2 0. 1 4 F r am e  N u m b er Er r o r O b j e ct T r a c ki n g  E r r o r ( O T E )     OT E Evaluation Warning : The document was created with Spire.PDF for Python.
In d onesi a n  J  E l ec En g &  C o m p  Sci    ISS N :  2 5 0 2 - 47 52       n e w method  fo r b a ll  tra cki ng  b a s ed  on   α - β , l i n ea r K a l m a n   an Ext e n d e d  ( H at hi r a N e nav at h)   99 5 The RMSEs  of the LKF,  α - β  filter an d  EKF  fo r m e d i u m  siz e  b a ll d a taset are g i v e n  i n  Table 2 .   On  an   avera g e, t h e m i nim u m  R M SE red u ct i o n was  0. 04 8 9  an d t h e   m a xim u m  was 0. 28 6 9  f o r m e di um  si ze bal l   dat a .   Si m ilarly,  th e PSNRs  o f  th LKF,  α - β  filter an d  EKF for med i u m  s i ze b a ll d a taset are g i v e n  in  Tab l e 2 .  As  obs erved i n  Ta ble 2, the a v erage va lue of  t h PSNR was increase d   afte r  appl y i n g  t h LKF t o  m e di um  bal l   data. On an average ,  the m a x i m u m  PSNR  was 45 .1 6 d B  an d t h e m i nim u m  was 28. 8 2 d B  for m e di um   si ze bal l   d a ta.  Fi gu r e  2 sh ow th e err o r  o f  cen tro i d,  er ro of  r a d i u s , RMSE,  cen t ro id   an r a d i us o f   x- coor d i n a tes, A E   an d OTE in each   fram e  fo r med i u m  size b a ll b a sed   on  LKF un d e no rm al l i g h t  co nd ition s         Fi gu re  2.  The  e xpe ri m e nt al  resul t s  o f  m e di u m  si ze bal l  t r acki n usi n g L K F             3. 3. 3   Trackin g  resu l ts of bi g siz e  ball in n o rm al  light condition   The OTE s  of the  L K F,  α - β   filter and   EKF fo r b i g  size  ball d a taset are g i v e n in   Tab l e 2 .   On an  avera g e, t h m i nim u m  OTE red u ct i o w a s 0. 2 9 9 7  a n d  t h e m a xim u m  was 1. 0 1 0 2  fo r bi g si ze  bal l  dat a Si m ilarly, th AEs of th e LKF,  α - β  filter and  EKF fo b i g   size b a ll d a taset are g i v e n  in   Tab l e 2 .   As  o b serv ed  in  Tab l e 2 ,  the av erag e v a lue o f  th e AE  was red u c ed  after app l yin g  th e LKF to  sm all b a ll d a ta.  On  an  avera g e,  t h e m i ni m u m  AE wa s 0. 6 3 9 9 9  a n d t h e m a xim u m  was  1. 34 1 2   fo r  bi g  si ze  bal l  d a t a .   The RMSEs  of the L K F,  α - β  filter an d   EKF for  b i g  size  b a ll d a taset are g i v e n  in  Table 2 .  On  an   avera g e, t h e m i nim u m  R M SE red u ct i o was  0. 2 1 1 2  a n d  t h e m a xim u m  w a s 0. 4 5 4 5   fo b i g si ze bal l   dat a . Th e   best avera g values are  em phasized in  Tabl e 2 .  Sim i lar l y, t h e PSN R of  t h e LK F,  α - β  fi lter  an E K F f o r big   si ze bal l  dat a s e t  are  gi ve n i n   Tabl 2.  As  o b s erve d i n  Ta bl e 2,  t h e  ave r ag e val u e   of t h PSNR   was  i n c r ease d   aft e r ap pl y i ng  t h e LKF t o  bi g  bal l  dat a . O n   an ave r ag e, t h e   m a xim u m  PSNR  was  45 .0 0 d B  an d t h e m i ni m u m   was 2 7 . 9 1dB   f o bi g si ze bal l  dat a . Fi g u re  sho w s t h e err o r o f  cent r oi d ,  e r r o of  radi us,  R M SE, cent r oi d an radi us of x-y coordinates,  AE and  OTE in each fram e  for big size  ball based  on  LKF unde r norm al light  co nd itio ns.    3. 3. 4   Trackin g  resu l ts of sm all siz e  ball in  bad light c o nditi on   The OT Es of t h e LKF ,   α - β  filter an d  EKF fo r sm all  size b a ll d a taset in  b a d  lig h t  con d itio n  are  g i ven  in Table  2. As  observe d  in  T a ble 2, the a v e r age  va lue  of  th e   O T E   w a s   re d u c ed  af te r  ap p l ying  th e LK F  to   sm al l  bal l  dat a . O n  an a v e r ag e, t h e m i nim u m  OTE re duct i on  was  0. 02 2 7  an d t h e  m a xim u m  was 0. 2 2 75  f o r   sm a ll size b a ll d a ta. Th is result in d i cates th at th e pred ic t e out put   of  t h e L K F i s  cl oser  t o  t h desi re d o u t put .   The best  a v e r a g e values   are  e m phasized  in  Tab l e 2.  Sim i larly,  th e AEs o f   th e LKF,  α - β  filter and  EKF  fo sm a ll size  b a ll  d a taset are g i ven  in  Tab l e 2 .   As ob serv ed  in Table 2, the a v era g e val u e of the AE  was reduc e d   after app l yin g   th e LKF to  sm all b a ll d a ta. On  av erag e, th min i m u m  AE  was 0.119 4  and  th e m a x i m u m was  1 . 0 528   for sm all size b a ll d a ta.     The RMSEs of the LKF,  α - β  filter an d  EKF fo r sm all s i ze b a ll d a taset in  b a d  lig h t   co nd itio n  are  gi ve n i n  Tabl 2. As  ob ser v ed  i n  Tabl e 2. O n  an avera g e, t h m i nim u m  RM SE red u ct i o n  was 0. 02 1 8  an d t h e   0 10 20 30 40 50 60 -1 -0 . 5 0 0. 5 1 1. 5 2 F r am e N u m ber D i st a n ce ( i n  m e te r s ) E r ror  of  k a l m an  f i l t er  c ent roi d  V s  f r am e i ndex 0 10 20 30 40 50 60 -0 . 4 -0 . 3 -0 . 2 -0 . 1 0 0. 1 0. 2 0. 3 0. 4 F r a m e N u m b er Di s t a n c e ( i n  m e t e rs ) E r ror of   k a l m an f i l t er  rad i u s  V s  f r a m e i n dex 0 10 20 30 40 50 60 0 0. 05 0. 1 0. 15 0. 2 0. 25 0. 3 0. 35 F r am e N u m ber err o r R oot  M e an S qua r e  err o r     RM S E 0 10 20 30 40 50 60 0 100 200 300 400 500 600 700 800 F r am e N u m b e r c e nt roi d  an d radi us   x - y   c oad i n a t es A c t u a l  c oor di nat es     c ent r o i d ra d i u s 0 10 20 30 40 50 60 0 0. 2 0. 4 0. 6 0. 8 1 1. 2 1. 4 1. 6 1. 8 F r am e N u m b e r E rror A bs ol u t e  error   ( A E )     AE 0 10 20 30 40 50 60 0 0. 02 0. 04 0. 06 0. 08 0. 1 0. 12 F r am e  N u m b er E rror Obj e c t  T r ac k i ng   E r r o r ( OT E )     OT E Evaluation Warning : The document was created with Spire.PDF for Python.
                        I S SN 2 502 -47 52  I ndo n e sian  J Elec Eng  & Com p  Sci, V o l. 10 No 3 ,  Jun e   2 018  :   98 9 – 99 99 6 max i m u m  was 0 . 1 767  fo r sm all size b a ll d a ta. Similarly, th e PSNRs  o f  t h LKF,  α - β  filter an d  EKF  for sma l l   si ze bal l  dat a s e t  are  gi ve n i n   Tabl 2.  As  o b s erve d i n  Ta bl e 2,  t h e  ave r ag e val u e   of t h PSNR   was  i n c r ease d   after app l yin g  th e LKF to  sm a ll b a ll  d a ta. On   a n  average, the m a xim u PSN R w a s 45 .1 7d an d  the  m i nim u m  was 27 .8 9 d B  f o r s m al l  si ze bal l  dat a               Fi gu re  3.  The  e xpe ri m e nt al  resul t s  o f   bi g  si ze bal l  t r ac ki n g   usi n g L K F       3. 3. 5   Objec t   trac king  fr ame res u l ts o f  sm a ll size ba ll in  no rma l  lig ht co nditio n   Fig u re 4  sho w s so m e  o f  th e test resu lts for s m all s i ze b a ll u n d e n o rm al  lig h t  con d ition .   We h a v e   evaluate our  approach using sel f-m ade Database,  with   th e f r a m e   size  b e ing  72 x 12 80 . W e  h a v e  selected  so m e  fram e s fro m  th e v i d e o   wh ile track i ng   th e ob j ect co n t in uo usly d u ring  its  m o v e m e n t . Th e targ et ob j ect is  m a rked  by  t h bo u ndi ng  b o x  ( r ed ci r c l e ). Fi g u re  4 s h o w f r am es out  of t h e t r acki n resu l t s  based  o n  L K F i n   Figure 4(a),  α - β  filter in  Figure 4(b) and  EKF in  Figu re  4 ( c).  In  fram es 2 0  and  30  sh own  in  Figure 4(a), th target object tracked  using L K F is  d o m i nated by  t h e re d b o u n d i n g b o x  as  the cam era is  static and the size of  t h e ob ject  i s  reduce d  wi t h   di st ance.  The sa m e  phen o m e na i s  obser ved i n  fram e s 30, 45  and  20 , 2 6  sh o w n i n   Fi gu re  4( b) a n d Fi g u r e 4 ( c) usi n α - β   filter an d   EKF,  respectiv ely. Ev en  th ou gh  the target o b j ect is small, th pr o pose d  t r ack i ng al go ri t h m s  can  d e tect  th e o b j ect (sm a ll  b a ll).                                       (a)                                                                         ( b                                                                  (c )     Fig u re  4 .    Obj e ct track ing   resu lt of th e sm all  size b a ll: (a)  based   on  LKF  (7 9 ,  20  and   3 0  fram e s),  (b)  based   on   α - β   filter (8, 15 , 30  and   4 5  fram e s) and   (c)  b a sed on   EKF (7 10 20  and   2 6  fram es)      3. 3. 6   Trackin g  fr ame results  of  medium si ze ba ll in no rma l  lig ht condition                     T h Figure  s h ows  each  fram e out  of the tracking re sults ba s e d on L K F in  Figure 5(a),  α - β  filter  in  Figu re  5(b)  an d   EKF in  Fi g u re  5 ( c).  In  spite o f  th reflectio n s  fro m  th e targ et obj ect on  th e tiles of the floo wh ich  is con s id er as  n o i se, the LKF,  α - β  filter, and  EKF al g o rith m s  are  effectiv e in trackin g  t h o b j ect.  0 10 20 30 40 50 60 -4 -2 0 2 4 6 8 10 F r am e  N u m b er Di s t a n c e ( i n  m e t e rs ) E r ror of   k a l m an   f i l t e r  c e nt roi d  V s  f r a m e i nde x 0 10 20 30 40 50 60 -3 -2. 5 -2 -1. 5 -1 -0. 5 0 0. 5 F r am e N u m b e r D i s t anc e ( i n  m e t e r s ) E r r o r  of  k a l m a n   f i l t er  r a d i us  V s  f r am e   i ndex 0 10 20 30 40 50 60 0 0. 5 1 1. 5 2 2. 5 3 3. 5 F r am e  N u m b er e r ror) R o ot  M ean  S qua r e   err o r     RM S E 0 10 20 30 40 50 60 0 20 0 40 0 60 0 80 0 10 00 12 00 F r a m e N u m b er c e n t r o i d   an d r adi us  x - y  c o a d i n at e s A c t ua l  c o o r di na t e s     ce n t r o i d r a di us 0 10 20 30 40 50 60 0 1 2 3 4 5 6 7 8 9 10 F r a m e N u m b er E rror A b s o l u t e  e r r o r (A E )     AE 0 10 20 30 40 50 60 0 0. 2 0. 4 0. 6 0. 8 1 1. 2 1. 4 F r am e N u m ber Er r o r O b j e c t  T r a c k i n g  E rro r(O T E )     OT E Evaluation Warning : The document was created with Spire.PDF for Python.
In d onesi a n  J  E l ec En g &  C o m p  Sci    ISS N :  2 5 0 2 - 47 52       n e w method  fo r b a ll  tra cki ng  b a s ed  on   α - β , l i n ea r K a l m a n   an Ext e n d e d  ( H at hi r a N e nav at h)   99 7                                  (a)                                                                         ( b                                                                  (c )     Fig u re  5 .  Obj e ct track ing   resu lt of th e m e d i u m  size b a ll:  (a)  b a sed on   LKF (1 0,  1 5 35  an d 50   fram e s),  (b)  base d on   α - β  filter  (9, 1 2 , 3 4 a n d   52  fram e s) an (c) b a sed  on   EKF (8 , 18 , 3 8 a nd  4 8   frames)      3. 3. 7   Trackin g  fr ame results  of  big siz e  ba ll in  no rma l  lig ht  co ndi tio n   The t e st  re sul t s f o bi g  si ze  bal l  u nde no r m al  l i ght  con d i t i on are  sh o w n i n   Fi g u re  6 .   W e   ha ve  sel ect ed som e  fram e s from  t h e vi de w h i l e  t r acki ng t h ob j ect  cont i n u o u s l y  du ri n g  i t s  m ovem e nt , w h e r e t h e   target object is  m a rked by a bou ndi ng  box (re d circle). Fi gure 6 shows  each four fram e s out of the tracking  r e su lts b a sed  on   LK i n  Figu re  6( a) α - β   filter in Figu re 6(b )  and  EKF in   Fig u re  6 ( c).  The values   of cova riance   m a trices   and   p l ays an  im p o r tan t  ro le in th esti m a t i o n   o f   LKF  and  EKF p a ram e te rs.  In  th is p a per, t h v a lu es  o f    and   were d e term in ed   usin g trial error wh ich fo rm s th delicacy of the  propose d  m e thod. C h oosing incorrect values of Q a n d R  may  lead to incr ease in RMSE and  decrease  in t h e  accuracy.                                          (a)                                                                         ( b                                                                        (c )     Fig u re  6 .  Obj e ct track ing   resu lt of th e b i g  si ze b a ll: (a b a sed   o n  LKF  (15, 25 , 45 an d 54fram e s), (b)  b a sed   o n   α - β  filter  (16 ,   2 8 4 0 a nd   5 2  fra m e s) and   (c)  b a sed   o n  EKF  (18 ,  26 , 48  and 53 fram es)       3. 3. 8   Trackin g  fr ame results  of s m all siz e  ball in b a d light  condition   The res u lts for s m all size ball unde r ba d light conditio n are shown i n  Figure  7. The target object is  m a rked  by  t h bo u ndi ng  b ox  ( r ed ci rcl e ). Fi g u re  7 sh o w s 4 f ram e s out  of t h e t r acki n g res u l t s  based o n  L K F i n   Figure 7(a),  α - β  filter in  Figure 7(b) and  EKF in  Figu re  7 ( c).  In  fram es 4 3  and  50  sh own  in  Figure 7(a), th target object tracked  using L K F is  d o m i nated by  t h e re d b o u n d i n g b o x  as  the cam era is  static and the size of  t h e ob ject  i s  reduce d  wi t h   di st ance.  The sa m e  phen o m e na i s  obser ved i n  fram e s 34, 48  and  45 , 5 0  sh o w n i n   Fi gu re 7( b )   a n Fi gu re 7(c ) ,  usi n α - β  filter and  EKF, respectiv ely.                                               (a)                                                                         ( b                                                                        (c )     Fi gu re  7.  O b je ct  t r acki n res u l t  o f  t h e  sm al l si ze bal l   un de r  ba d l i g ht  c o n d i t i on:  (a)  b a sed  o n  L K ( 2 5 ,   3 5 4 3 an d 50   f r a m e s) , (b ) b a sed on   α - β   filter (19 ,  2 4 , 3 4   and  48  fram es)  an d (c)  b a sed   o n  EKF  (26 ,  36 , 45 , and     50 fram es)      4.   CO NCL USI O N   In t h i s   pape r, a  new m e t hod i s  devel ope d f o r t r acki ng  of  di ffe rent  si zes o f  bal l s  usi n α - β , LKF a n EKF.  In  pr op o s ed t r acki ng  fr am ewor k, t h re e di ffe rent  t r ac ki n g  al go ri t h m s  were us ed  to  track the target object   u n d e r th n o rmal lig h t  co nditio n  fo d i fferen t sizes of b a l l s an d  also   u nder b a d  ligh t  con d ition s  for the sm a l l   si ze of t h bal l .  Ex peri m e nt al resul t s  d e m o nstrate th at LKF alg o r ith m  is  m o re effective and e fficient than  α - β   and E K F al gorith m s  for linea r system   m odel applications.  In sp ite  o f  th reflection s   from  th e targ et obj ect  o n   Evaluation Warning : The document was created with Spire.PDF for Python.
                        I S SN 2 502 -47 52  I ndo n e sian  J Elec Eng  & Com p  Sci, V o l. 10 No 3 ,  Jun e   2 018  :   98 9 – 99 99 8 th e tiles o f  th e flo o wh ich  is co n s i d ered  as n o i se, t h e LKF,  α - β  filter, an d  EKF al g o ri th m s  are effectiv e in  t r acki n g t h e o b ject . F o r fut u re w o rk , an  opt i m i zati on m e t hod co ul d  be consi d ere d  i n  det e rm i n i ng t h e   cova riance  m a trices Q and R ,   to obtain a n  optim a l  LKF a n EKF  pa ram e ter  estim a tor m o re accurately.      REFERE NC ES  [1]   K. Cannons., “A review of v i sual tr acking”, in Tech.  R e p.  CSE-20 08-07,  York Univ., Ontario ,  C a n a da, 2008.  [2]   A. Yilmaz, O. Javed,  and M. Shah , “Object tr acking: A survey ”, in  ACM Comput. Surv.,  vol. 38 , no. 4, pp. 1–45,  2006.  [3]   B. Babenko, M.-H. Yang, a nd S. Belongie, “Robust object tr acking with onli n e m u ltiple inst ance l earn i ng”,  in   IEEE  T r ans. Pat t ern Ana l .  Mach .  Inte ll. , vo l.     33 , no . 7 ,  pp . 1619 –1632, Aug 201 1.  [4]   D. Com a niciu,  V. Ram e s h , and  P .  M eer , “ K ernel-based ob jec t  tracking in  I EEE T r ans. Pat t ern Anal. Ma c h Intell . , vol. 25 , n o . 5 ,  pp . 564–57 7, May   2003.  [5]   S.  Ha re ,  A.  Sa ffa r i,   a nd P.  H.  S.  T o rr,  Struc k S tructured outpu t tracking with kernels”,  in P r oc.  IEEE Int .  Conf.   Comput. Vis., p p . 263–270 , 201 1.  [6]   X. Mei and H. Ling, “Robust visual  tracking an d vehicle classif i cation via spars e  representation , in  IEEE Tran s.  Pattern  Anal . M a ch. In te ll. , vo l.  33, no . 11 , pp . 2 259–2272, Nov  2011.  [7]   D. Ross, J. Lim, R.-S. Lin and   M.-H.  Yang, “I ncremental learn i ng for robust vi sual tracking ”,  in Int. J. Comput.  Vis., vol. 77 , no 1, pp. 125–141,  2008.  [8]   X.  Li,  W.  Hu,  C.  Shen,  Z.  Zhang,  A.  Dick,  and A.  He ngel, “A  survey  of appearan ce models in visual object  track ing”,  in   AC M Trans. Intell.  Syst.  Technol. , v o l. 4 ,  no . 4 ,  pp A: 1-A: 42 , 201 3.  [9]   B. D. Lucas  and  T. Kanade “An iterative imag e registration technique  with an  application to s t ereo vision”,  in   Proc. 7th  Int .  Joi n t Conf.  Artif . In tell . , pp. 674–67 9, 1981 [10]   N.  Alt,  S.  Hinterst oisser, and N. Navab,  “Rapid selection of relia b le  templa tes fo r visual tracking , in P r oc. IEE E   Conf. Comput.  Vis. Pattern R e cognit., pp . 1355– 1362, 2010 [11]     S. Baker and I.  Matthews,  Lu cas-Kanade 20  y ears on:  “A unif y ing framework”,  in  Int.  J. Comput. Vis. , vol. 56, n o 3, pp . 221–255 2004.  [12]   I. Matthews, T. Ishikawa,  and  S. Baker, “The  temp late update problem”, in  I EEE T r ans. Pat t ern Anal. Ma c h .   Intell . , vol. 26 , n o . 6 ,  pp . 810–81 5, Jun 2004 [13]   M. J. Black  and   A. D. Jepson , “Eigen Track ing:  Robust ma tchin g  and  tracking o f  arti culated objects using  a view- based rep r esentation”, in   Int.  J. C o mput. Vis. , vol.  26, no . 1 ,  pp . 63 –84, 1998 [14]   G.  D.  Hager  and P.  N.  Belhumeur,  “Efficient region  track i ng with par a metric models of  geometr y  and   illum i nat i on” , in   IEEE Trans. Pa ttern  Anal. Mach . Intell. , vol. 20 no. 10 , pp . 1025 –1039, Oct 1998 [15]   G. Zhang, W. Zhu,  “Automatic video object  segmentation by in tegrating object registration  and background   constructing technology”,   in Pr oceed ings of th e  2006 Int e rnat io nal Conf erenc e   on Com m unicati ons, Circu its an d   S y stems, pp. 43 7–441, 2006 [16]     W.-C. Hu, “R eal- time on-lin e video object  segmenta tion b a sed on motion detection without backgroun d   construction”, in   Int. J. Inno vat.  Comput.,  Inform. Control,  7  (4),  pp. 1845–1860 2011.  [17]     Jing Cheng, Sucheng Kang R obust Visual Track ing  with  Im proved Subspace Repr esent a tion Mode l”,  i n   TE LKOMNIKA  Te le c o mmunic a tion,  Com puting, Electronics  and  Control,  Vol 15, no. 1, March  20 17.  [18]   Indah Agustien Siradjuddin ,  M. Rahmat  Wid y anto, T. B a sarud d in, “Particle  Filter with Gaussian Weighting fo Hum a n Tracking ,  in Indonesian  Journal of Electrical E ngin eerin g and Computer  Scien ce ( I JEECS ) , Vol 10, no. 6,  Oct 2012.  [19]     D. Wang, H.  Lu,  and M.-H.  Yang, “ O nline   objec t tr acking   with sparse pro t ot ypes” ,  in   IEEE Trans. Ima g Proc e ss. , vol. 22 , no . 1 ,  pp . 314– 325, Jan  2013.  [20]   B. Liu, J. Huan g, L. Yang, and  C. Kulikowsk,  “Robust trackin g  using local sp arse appearance model and K- selec tion” in Pr oc. I E EE Conf Comput. Vis.  Pattern  Recogn it.,  pp. 1313–1320 2011.  [21]   B. Ris t ic  and M .  L.  Hernand e z ,  “ T rack ing s y s t em s , in   P r oc.  IE EE  RA DA R , Rome, I t aly ,  pp. 1-2,  2008.  [22]   W. Zhong, H. Lu, and M.-H . Yang,   “Robust object  tracking via  sparse coll aborative  app ear ance model”, in  IE EE  Trans. Image Pr ocess. , vol. 23, n o . 5 ,  pp . 2356–2 368, May   2014.    [23]   J.  Fisc us,  J.  Garofolo,   T.  Rose, a nd M.   M i ch el , “ AVSS mul tipl e  camer a  p e rson tracking  challenge  evalua tion   over vi ew” ,  in  Pr oc . 6 t h I EEE A V SS, Genova, I t aly ,  2009.  [24]     C .  B .  S a n t i a g o ,   A .  S o u s a ,  M .  L .  E s t r i g a ,  L .  P .  R e i s ,   and M. Lames, “Survey  on team  tracking tech niques applied to  sports”, in   Proc. AIS, Povoa   de  Varzim , Portugal,  pp. 1–6 , 2010 [25]     J .  C. M c Call and M .  M .  Trivedi, “ V ideo-bas ed lane es tim ati on and tracking for driver as s i s t anc e :  S u rve y , s y s t em ,   and ev alu a tion ,   in  IEEE Trans. Inte ll. Transp. S y st ., vo l. 7, no. 1,  pp. 20–37 , Mar   2006.  [26]     C. R. Wren, A. Azarbay e h a ni, T. Da rrell, and A. P. Pentland,  Pfinder: “Real-time tracki ng of the  human body ”, in   IEEE  T r ans. Pat t ern Ana l .  Machi n e Int e ll ., vol. 1 9 , no . 7 ,  pp . 780 –785, Jul 1997 [27]   L. Maddalena and A. Petrosi no, “A  self-organizing approach to bac kground subtraction for visual surveillance  applications”, in  IEEE Trans. Image Process. , vol. 17, no. 7, pp. 11 68–1177, Jul 20 08.  [28]   M. Oral and U.  Deniz, “Cent r of  m a ss  m odel—A novel appr oach to b ackgro und m odelling f o r segm entation  of  moving objects”,   in  Image Vis.  Comput .,   vol. 2 5 , no . 8 ,  pp . 136 5–1376, Aug 20 07.  [29]   Paul R. Kal a ta ,  “ T he track ing  i ndex: A gen e ralized par a meter for  α - β  and   α - β - γ   target tr ackers " , in  IEEE  Transactions on  Aerospace and   Electronic Systems , AES Vol. 20 , no . 2 ,  pp .174–1 81, mar 1984 Evaluation Warning : The document was created with Spire.PDF for Python.