Bayesian Robot Localization using Spatial Object Contexts

Transcription

1 The 29 IEEE/RSJ Inenaional Confeence on Inelligen Robo and Syem Ocobe 11-15, 29 S. Loui, USA Bayeian Robo Localizaion uing Spaial Objec Conex Chuho Yi, Il Hong Suh, Gi Hyun Lim, and Byung-Uk Choi Abac We popoe a emanic epeenaion and Bayeian model fo obo localizaion uing paial elaion among objec ha can be ceaed by a ingle conume-gade camea and odomey. We fi ugge a emanic epeenaion o be haed by human and obo. Thi epeenaion coni of peceived objec and hei paial elaionhip, and a qualiaively defined odomey-baed meic diance. We efe o hi a a opological-emanic diance map. To uppo ou emanic epeenaion, we develop a Bayeian model fo localizaion ha enable he locaion of a obo o be eimaed ufficienly well o navigae in an indoo envionmen. Exenive localizaion expeimen in an indoo envionmen how ha ou Bayeian localizaion echnique uing a opological-emanic diance map i valid in he ene ha localizaion accuacy impove wheneve objec and hei paial elaionhip ae deeced and inaniaed. I. INTRODUCTION Sevice obo pefom complex ak whoe execuion equie fequen ineacion wih human. Imagine viiing a univeiy and aking a ecuiy guad how o ge o Pofeo Suh office. The guad migh ay, Go o he IT building and poceed along he coido o you igh, whee you will ee eoom on you lef a well a claoom. When you each he end of he coido, un lef. A ha poin, you will find he office on you igh. You would find he coec oom wihou any difficuly following hee inucion. Human do no neceaily ue accuae quaniaive infomaion o peceive pace in he cuen locaion o fo aveling o anohe locaion. Inead, human emembe a few landmak ha define he pace. Baed on a pecific ucue o diinc objec, human eucue hei knowledge baed on paial conex and hen eue he knowledge [15]. Thi mehod may no define he exac locaion quaniaively, bu a many fagmen of paial conex accumulae, i enable a ufficienly high level of pace ecogniion and localizaion. I i neceay o build an accuae meic map and a emanic map fo ymbolic infeence fo he cuen obo localizaion poblem. Thi equie ymbol gounding beween he meic daa and he emanic epeenaion. An accuae eno i eenial fo he execuion of complicaed maneuve. In he emanic map popoed by Kuipe e al. [6], Thi wok wa uppoed by he Technology Innovaion Pogam of MKE/KEIT [28-F-38-1, Developmen of Conex Adapive Cogniion Technology] Chuho Yi i wih he Diviion of Elecical and Compue Engineeing, Hanyang Univeiy, Koea d1uck@hanyang.ac.k Il Hong Suh i wih he College of Infomaion and Communicaion, Hanyang Univeiy, Koea. All coepondence hould be addeed o I. H. Suh ihuh@hanyang.ac.k Gi Hyun Lim i wih he Diviion of Elecical and Compue Engineeing, Hanyang Univeiy, Koea hmeal@hanyang.ac.k Byung-Uk Choi i wih he Diviion of Elecical and Compue Engineeing, Hanyang Univeiy, Koea buchoi@hanyang.ac.k he emanic ucue of pace i infeed uing he model buil, baed on a hieachy of ucceive envionmenal daa accumulaed while he obo move aound. In addiion, a global map i deived by inegaing he global opology infomaion and he local meic daa. Many obo localizaion mehod have been udied ove he la decade, uing gid-baed map [5], feaue-baed map [8][2], opological map [7], emanic map [6], o elecing divee ype of map adapively [9]. In eeach on he pobabiliic Bayeian model, Thun e al. applied Baye ule o obo localizaion [3]. The opological map i epeened in emanic uing onology, and elie on emanic o infe navigaion [1][11]. In a ecen emanic appoach, Vaudevan e al. popoed a hybid map ha elie on numeical daa in an indoo envionmen and ue opological infomaion fo he inegaion of each pace [2][16]. Baed on hi appoach, we conuced a pobabiliic objec gaph uing he pace of ecognized objec. In addiion, we popoe place claificaion uing uch a epeenaion. Ranganahan e al. popoed a mehod of emanic locaion modeling fo pace ecogniion uing he emanic infomaion of he ecognized objec [1]. They uggeed a Bayeian infeence ucue ha conain pobabiliic model fo lowlevel feaue up o high-level infeence. Thee, he inceae in compuaional complexiy wa eolved wih a Makov chain Mone Calo appoach. Semanic epeenaion i neceay fo humanineacing evice obo o ake ode and complee ak. Unil now, we have concenaed on obo-ceneed knowledge ha enable human o ineac wih obo [13][17]. Mo eeach on high-level knowledge ha deeced objec in camea image and eimaed he diance fom he camea o he obeved objec o ceae elaionhip among he objec. Howeve, hee quaniie ae inaccuae due o len dioion and incoec feaue deecion and maching. Thi lead o eo when applied o eal envionmen. To cope wih hee iue, we ugge a emanic epeenaion and Bayeian model fo paial elaionhip among objec, and how ha ou epeenaion i ueful fo he localizaion of a mobile obo. The paial conex ued in he popoed emanic epeenaion include he obeved objec, he obo-o-objec (-o) diance conex ha epeen he diance fom he obo o a paicula objec, he -o beaing conex ha epeen he diecion fom he obo o a paicula objec, and he objec-oobjec (o-o) elaionhip conex ha denoe he elaionhip among objec. Fo hi, we ue viual paen ecogniion o /9/$ IEEE 3467

2 ecognize objec and eimae ough meic daa [18]. We nex popoe a opological-emanic diance map coniing of paial objec conex and paial obo conex. In a opological-emanic diance map, a node i one of he componen of a geneal opological map ha play he ole of a andad and conain infomaion on he paial objec conex. In addiion, paial obo conex ued in he popoed emanic epeenaion can explain an appoximae diance and beaing fom one aigned node o anohe. We decibe how an appoximae qualiaive diance i he nodeo-node (n-n) diance conex and qualiaive beaing i he n-n beaing conex. Global localizaion i heefoe made poible by he objec in he opological-emanic diance map and he paial objec conex. In addiion, he localizaion i poceed moe pecifically and locally baed on he infomaion gaheed fom he global localizaion, deiving he pobabiliy diibuion of a obo poibly locaing ielf aound he node. Theefoe, we popoe he pobabiliic Bayeian model by which localizaion accuacy can be impoved a many fagmen of paial objec conex ae coninually obained, and evenually veify he pacicaliy of he popoed mehod hough an expeimen o ha obo localizaion i ufficienly qualified fo navigaion in indoo envionmen. One of he mo impoan conibuion of he pape i he manne in which noie in he meauemen povided by he ERSP(Evoluionay Roboic Sofwae Package) yem wih a ingle camea i eaonably managed. II. TOPOLOGICAL-SEMANTIC DISTANCE MAP Almo all daa in ou mehod ae epeened emanically by mean of onology, which enue ha only ound and complee daa ae aeed and popagaed wih onology infeence. Noiy eno daa uch a fale-poiive and uenegaive can be fileed uing he elaionhip and ule of logical eaoning. In many cae of fale-poiive, he popeie ae illogical, uch a when a miclaified objec i floaing in he ai o peneaing wall o ohe objec. Thee cae can be evaluaed logically by axiomaic ule, and he obo will know wha o expec in he nex ep. Tha enable obo o pedic and pay aenion a pioi. A. Spaial Objec Conex Figue 1 how he change in eal elaionhip beween a obo and obeved objec accoding o he obo locaion aniion fom x o x in he eal wold. Paamee,, and ae he meic diance of he objec elaive o he obo, he meic beaing of he objec elaive o he obo, and he beaing among objec in eal-wold coodinae, epecively. The ubcip indicae he indexe of he objec obeved. Figue 2 how he change in meic elaionhip beween a obo and obeved objec accoding o he obo locaion aniion fom x o x. Paamee,, and denoe he eimaed meic diance of an objec elaive o he obo, he meic beaing of an objec elaive o he obo, and he beaing among objec in he obo coodinae, epecively, meaued by a ingle camea. In geneal, he x Fig. 1. Fig. 2. o 1 o 1 1, Objec ( o ) u, , 2 x 3, , 2 Objec ( o ) 3 3 Objec ( o ) 2 2 Spaial elaionhip beween obo and objec in he eal wold. Peviou fame ~ 13 ~ ~ ~ ~ 1, 1 2, 2 3, 3 ~ ~ x ~ ~ u ~, ~ 2 2 Cuen fame ~ ~ 3, 3 ~ Meic elaionhip beween obo and objec in image equence. 1, 1 Spaial Conex of Peviou Sae x u x 23 Semanic Repeenaion of Spaial Conex 2, , 3 x Spaial Conex of Cuen Sae, , 3 Fig. 3. Semanic epeenaion compiing a e of epecive paial objec conex among he obeved objec and he obo. diance and oienaion obained uing a commecial viion yem uch a he ERSP viion yem wih a ingle camea ae ofen inaccuae. In cae when a ingle camea i aached o a obo, he meic daa quaniie uch a diance and oienaion o an objec fom he obo become oo uneliable o be ued. Thu, fo emanic epeenaion, he meic daa ae linked o he ymbol of he paial elaionhip, accoding o he given condiion. Howeve, we ue he infomaion ha muual aociaion exploi he geomeic elaionhip beween objec (landmak) [19]. 3468

3 TABLE I SEMANTIC REPRESENTATION INCLUDING ALL OF THE SPATIAL OBJECT CONTEXTS IN FIGURE 3. Sae Semanic Repeenaion Peviou ae neaby(o 1, Robo), lef fon(o 1, Robo), igh nea(o 1, ), igh fa(o 1, ), fa(, Robo), fon(, Robo), lef nea(, o 1 ), igh nea(, ), fa(, Robo), igh fon(, Robo), lef fa(, o 1 ), lef nea(, ) Cuen ae nea(, Robo), lef fon(,robo), igh fa(, ) neaby(, Robo), igh fon(, Robo) lef fa(, ) Fig. 4. Example of opological node uing a genealized Voonoi gaph of he 4h floo of he IT building a Hanyang Univeiy, Seoul, Koea. u 1 x 1 u x u +1 x +1 Objec ecogniion i a fundamenal faco in emanic epeenaion. In geneal, an objec i ecognized viually by meauing he imilaiy beween feaue of obeved objec and hoe of coeponding objec model. In hi ecion, we ue cale-invaian feaue anfom (SIFT) feaue ha ae known o be invaian o cale and oaion change [4][12]. The diance fom he obo o he obeved objec i eimaed wih a ingle camea o deive a fagmen of paial conex. Afe an objec i ecognized, i heigh in he image pace i meaued uing a e of coeponding feaue, and hen a meic diance i eimaed by ERSP [18]. B. Semanic Repeenaion of Spaial Objec Conex Figue 3 how a emanic epeenaion coniing of obeved objec and hei epecive paial ymbol. Hee, he paial conex include diance, beaing, and elaionhip conex. The -o diance conex denoed by i he diance of he objec fom he obo. Each diance conex i epeened by one of a e of diance ymbol, ha i, = {neaby, nea, fa}. The -o beaing conex denoed by = { fon, lef fon, lef, lef ea, ea, igh ea, igh, igh fon} i he beaing of he objec elaive o he obo. The o-o elaionhip conex denoed by = { lef fa, lef nea, lef neaby, igh neaby, igh nea, igh fa} i he elaionhip among objec. Table 1 how a emanic epeenaion uing ymbol fo all of he paial conex in Figue 3. Ou obo localizaion applicaion find he poiion of he obo uing only hee ype of emanic epeenaion wih qualiaive meic daa. C. Topological-Semanic Diance Map Neihe he mehod we popoe hee no he expeimen we conduced focued on opological map-building, which can be conideed meely a placeholde cheme fom he oboic eeach poin of view. Figue 4 how an example of opological node obained uing a lae eno. In ou expeimen, we egieed opological node when he obo wa locaed a poiion defined by a genealized Voonoi gaph. Topological-emanic diance-map building ake place in hee ep: c 1 z 1 o 1 1 Fig. 6. c z m o c +1 Gaphical model of Bayeian localizaion. z +1 o ) Topological node fom a ange eno ae aigned o a diinc pace. 2) Obeved objec and eimaed paial objec conex ae egieed in he opological-emanic diance map a he obo oae 36 aound a fixed poiion. 3) The diance conex and he beaing conex eimaed duing obo movemen i added beween node. Thi i conideed a paial obo conex of n-n conex. The fi ep i idenical o he aignmen of diinc node. We ued a lae eno (SICK 5) o aign node ha wee ubequenly poceed o build a opological map [7]. In he econd ep, obeved objec and paial objec conex wee added o he opological-emanic diance map. Each paial objec conex individually deemine fo which cuen obo i locae a node. In he hid ep, he paial obo conex eimaed by odomey beween node ae added o he opological-emanic diance map. Thee conex coniue elaive infomaion beween node. Figue 5 give an example of a opological-emanic diance map. III. BAYESIAN MODEL FOR SEMANTIC LOCALIZATION In hi ecion, we decibe ou Bayeian model and hei mahemaical fomulaion. We ely paially on he pobabiliic Bayeian model popoed by Thun e al.[3] and Ranganahan e al. [1] o fomulae ou pobabiliic model. A. Pobabiliic Localizaion Model Figue 6 depic a gaphical model of Bayeian localizaion. The obo i given a map and i goal i o deemine i locaion elaive o hi map given obevaion of he envionmen and he obo movemen. 3469

4 on nea, fon nea, lef ea {, } o o ) (2), (4)) (2), (4) o (4) { ( v o, ( v o )} (2), (3)) (2), (3) Node 11 fa, lef ea fa, lef ea lef nea o (1) {, } o o ) (1), (1)) { } ) ( o (1), o(2) (1), (1) v (1) o (2) e (1,2) v (2) ρ θ { c, } c v v ) (1), (2)) { ( v o, ( v o )} (1), (2)) (a) (1), (2) (1), (2) o (3) ρ θ { c, } c v v ) (11), (12 ) ) (11), (12) 4 ep, fon Node 12 CoidoA { nea, ea, (12 ),BoadA ) (12 ),BoadA ) nea, lef ea (b) ih f { } ) lef nea nea, fon nea, fon (CoidoA,BoadA Fig. 5. Example of a opological-emanic diance map. (a) Semanic epeenaion compiing a e of epecive paial objec conex and paial obo conex. (b) Specific example of a opological-emanic diance map compiing a e of epecive paial objec conex and paial obo conex. (c) Relaionhip beween opological-emanic diance map and he eal envionmen. BoadA } BoadB BoadC DooA BoadC (c) SdooB BoadA CoidoA In he figue, obo locaion i denoed by x = [v (i) e (i,j) ρ (i,j) θ (i,j) ] T whee v (i) and e (i,j) ae he opological node and edge, epecively. In addiion ρ (i,j) and θ (i,j) ae meic diance and beaing on edge e (i,j), epecively. The objec i denoed by o. Thi meic-opological famewok enable localizaion o be bounded globally, he map ize o inceae monoonically in dimenionaliy, and he locaion o be calculaed locally beween wo node. A e of emanic fo paial objec conex conaining diance, beaing, and elaionhip i denoed by = {,, }.A e of emanic fo paial obo conex conaining diance and beaing i denoed by c = {c ρ,c θ }.Weepeenhe feaue exaced fom he image wih z andhemapi denoed by m. We have aleady dicued he ype of map, he opological-emanic diance map. Pobabiliic obo locaion epeen belief hough condiional pobabiliy diibuion. We will denoe belief ove a ae x by Bel(x ), which i an abbeviaion fo he following: Bel(x )=p(x : 1:,o 1:,z 1:,c 1:,u 1:,m) (1) Peio Bel(x ) i obained in a imila way a he deivaion uing Baye ule and he Makov aumpion. In paicula, we have Bel(x ) = η p(,o,z x,m) p(x x 1,c,u )Bel(x 1 )dx 1 (2) whee he pobabiliy p(,o,z x,m) i he emanic meauemen, he pobabiliy p(x x 1,c,u ) i he ae aniion, and Bel(x 1 ) i belief a ime 1, epecively. The pedicion model i p(x x 1,c,u )Bel(x 1 )dx 1. The pobabiliic localizaion model i divided ino wo pa, namely, he meauemen model and he pedicion model, which coepond o wo em on he igh ide of Eq. (2). The meauemen model ue he emanic epeenaion and i he main focu of hi wok. Conex ae unceain daa o hey hould be appoximaed wih a ochaic diibuion. The focu of hi ecion i he meauemen model and conex of objec. Hee, we aume ha eno ae unceain, and a lowe diibuion i appoximaed becaue fewe emanic ae available. The evee would alo be ue. We calculae he locaion poeio uing he emanic epeenaion decibed in Secion II. B. B. Semanic Meauemen Model The emanic meauemen model i baed on he join pobabiliy of obo locaion x, mapm, paial objec conex, objec o, and exaced feaue z, Fom he gaphical model in Figue 6, he join pobabiliy can be wien a p(,o,z x,m)= p(x )p(o x,m)p( x,o,m)p(z,o ) p(x,m) (3) Hee, we aume ha he pobabiliy p(x ) i he ame a p(x,m) becaue he obo poiion x i locaed on map m. Applying Baye law o he emanic meauemen model fom Eq. (3) give p(,o,z x,m) = p(o x,m)p( x,o,m) p(z)p(o z)p( o,z) (4) p(o )p( o ) We aume ha he pobabiliy p(z ) ha a unifom diibuion. Theefoe, we can obain he diibuion of he meauemen model a p(,o,z x,m) (5) = η p(o z )p( o,z )p(o x,m)p( x,o,m) whee η i he nomalizaion conan and p(o z ) i a em elaed o objec ecogniion. I i evaluaed baed on he 347

5 imilaiy beween exaced feaue of he obeved objec and he feaue of he coeponding objec model. We ue a upevied appoach o building he objec model. Taining daa coni of image, each of which conain only one objec. Thee image ae capued a evey known efeence diance. The objec model coni of feaue fo objec ecogniion, diance fom he camea o he coeponding objec, and he heigh of he objec in pixel. The p( o,z ) i he likelihood of imilaiy in paial conex beween he eimaed meic daa fo he obeved objec and he paial conex. I i evaluaed baed on he imilaiy beween exaced feaue of he obeved objec and he feaue of he coeponding objec model. Spaial conex of objec ae compued fom he eul of eimaed meic diance and beaing a p( o,z )={p( d o,z ),p( b o,z ),p( o,z )} (6) p( o,z )=N(, σ 2 ) (7) p( o,z )=N(, σ 2 ) (8) p( o,z )=N(,σ 2 ) (9) whee N(μ, σ 2 ) i nomally diibued wih mean μ and vaiance σ 2. In addiion,,,and ae ained a evey known efeence fo he paial conex of diance, beaing and elaionhip, epecively. The vaiance of σ, 2 σ,and 2 σ 2 ae inoduced o eflec he unceainy of viion eno. The fuhe an objec i fom he obo o fom ohe objec, he moe inaccuae he meic diance and beaing will be. Moe pecifically, dividing he paial conex moe finely will impove localizaion pefomance. The pobabiliy p(o x,m) i he likelihood of imilaiy beween he obeved objec in he cuen ae and hoe in he peviou ae, and fomulaed a p(o x,m)=exp( o o x 2 ) (1) whee o and o x epeen he obeved objec in he cuen and peviou ae, epecively. The pobabiliy p( x,o,m) i he likelihood of he paial conex. The likelihood of he paial conex i compued a p( x,o,m) = N a [f c( a x)f c ( a x ) N b f (11) c( ab x)] whee f c ( a, x ) = exp( a x 2 ), and and x epeen he eimaed paial conex of an objec in he cuen and peviou ae, epecively. Each paial conex of an objec belong o one of ome numbe of diffeen diibuion. Each conex in Eq. (11) i decibed by a componen pobabiliy deniy funcion, and i mixue of diibuion i he pobabiliy ha an obevaion come fom hi componen. Hee, a mixue of hee nomal diibuion wih diffeen mean may eul in a deniy wih hee paial conex of he objec, which i no modeled by andad paameic diibuion. C. Pedicion Model and Recuive Bayeian Model The la em on he igh ide of Eq. (2) i an updae em. The conol model u epeen imple moion daa elaed o he ae aniion, a follow: v (i) e (i,j) ρ (i,j) θ (i,j) = v (i) e (i,j) ρ (i,j) θ (i,j) + Δρ co(θ (i,j) +Δθ θ c (i,j) ) Δθ + ε ρ ε θ (12) whee Δρ and Δθ ae he obo moion of beaing and movemen, epecively, while ε ρ and ε θ ae a zeo-mean eo vaiable of moion and beaing. The localizaion poeio i calculaed uing Eq. (2), howeve, he modificaion i no applied diecly o ou Bayeian model becaue Eq. (2) i no a pedicable em. To olve hi poblem, we modify Eq. (2) in he localizaion model o a paicle file, a ype of ecuive Bayeian eimaion [3][14], o manage he complicaed compuaion. In paicle file, he ample of a poeio diibuion ae called paicle and ae denoed by χ := x [1],x [2],..., x [Y ] (13) Each paicle x [y] (wih 1 y Y ) i a concee inaniaion of he ae a ime. So-called impoance faco ae ued fo each paicle x [y] o incopoae he meauemen ino he paicle e. The impoance i hu he pobabiliy of he meauemen unde he paicle, given by w [y] = age diibuion popoal diibuion = p(o z )p( o,z )p(o x [y],m)p( x [y],o,m) (14) By eampling paicle wih a pobabiliy popoional o w [y], he euling paicle ae indeed diibued accoding o he poduc of he popoal and he impoance weigh w [y]. Bel(x )=η w [y] Y y p(x [y] x [y] 1,c,u,m)Bel(x 1 ) (15) Figue 7 how a lage paicle diibuion ha denoe he iniial obo locaion, and hen become malle wih ecuion, eflecing a numbe of paial objec conex. 3471

6 (Localizaion Eo) 6 5 Fig. 7. Example of a concenaed diibuion wih ecuion eflecing a numbe of conex. Eo (cm) A B C D Image equence Fig. 8. Example of ained objec(landmak) image. Fig. 11. Localizaion eo in ou expeimen uing paial objec conex. IV. EXPERIMENTAL RESULTS A Pionee 3 AT obo caying a ingle conume-gade camea wa diven aound an indoo envionmen ( m) o evaluae he pefomance of he popoed localizaion poce. Figue 8 how example of ained objec image eleced by he ue. The camea obeved 16 objec duing i avel. Diincive objec uch a dooplae, bullein boad, and panel boad wee ued fo objec ecogniion. Figue 9 illuae he opological-emanic diance map coniing of 15 node (yellow, ecangle) and 42 objec (geen, cicle). Solid line beween node ae he edge ha epeen n-n conex of diance and beaing. The doed blue line denoe he -o conex and he doed ed line epeen he o-o conex. The opological-emanic diance map i included a an onological epeenaion fo obo knowledge. Figue 1 how pa of he onology chema and map inance. The expeimen howed an 88.93% pobabiliy of exac accuacy in opological localizaion. Howeve, 13.7% of he pobable eo only occu aound he neighboing node. The eaon ha hi eo occu i ha a aniion i poceed beween a a node o a goal node in he aniion beween edge. Figue 11 how he localizaion eo fom node 11 o node 12 in he popoed emanic localizaion. Even hough he obo locaion wa iniially eimaed accuaely, he eo inceaed due o he lack of emanic obained in ubequen movemen. When objec fa fom he obo wee obeved, he localizaion eo and andad deviaion wee elaively high. Howeve, he diibuion of obo locaion became malle a he obo coninued o navigae and he obeved objec became cloe. If he peviouly obeved objec diappeaed due o he oaion of he obo, he andad deviaion inceaed. In hi expeimen, he maximum eo wa 68.4 cm, and he mean and andad deviaion wee 27.3 and cm, epecively. A poin A in Figue 11, he diibuion of obo locaion became malle due o obevaion of he objec nea he obo. A poin B, he objec diappeaed due o he oaion of he obo, heeby peading ou he diibuion. A poin C, he obeved objec wee fa fom he obo, leading o high unceainy in he paial conex. The eo in obo locaion a well a he diibuion became lage. Howeve, he iuaion a poin D how ha obo locaion eo and i coeponding diibuion deceaed a he obo moved cloe o objec. V. CONCLUSIONS Thi pape popoed a emanic epeenaion and Bayeian model fo obo localizaion uing paial conex among objec, and decibed hem uing ymbol. Ou popoed mehod enable obo o be localized uing paial objec conex and hei pobabiliic model. Expeimenal eul of ou popoed Bayeian obo localizaion cheme in an indoo envionmen demonaed ha a conexual evidence inceaed, locaion accuacy impoved depie uing an inaccuae eno uch a a conume-gade camea. REFERENCES [1] A. Ranganahan and F. Dellae, Semanic Modeling of Place uing Objec, Roboic: Science and Syem (RSS), 27. [2] S. Vaudevan, V. Nguyen, and R. Siegwa, Cogniive Map fo Mobile Robo - An Objec baed Appoach, Poceeding of he IROS, pp 7-12, 26. [3] S. Thun, W. Bugad, and D. Fox. Pobabiliic Roboic, MIT Pe, Cambidge, MA, 25. [4] D.G. Lowe, Diincive image feaue fom cale invaian keypoin, In l Jounal of Compue Viion, Vol. 6, n, pp , 24. [5] A. Elfe, Uing occupancy gid fo mobile obo pecepion and navigaion, IEEE Compue, Vol.22, pp.44-57, [6] B. Kuipe, The Spaial Semanic Hieachy, Aificial Inelligence, pp ,

7 Nod e 1 Bo ada SdooB Win dowb fa, igh Node 7 N ode 8 Node 9 Node 1 5 ep, fon 6 ep, igh ea 6 ep, lef fon 3 ep, ig h ne a, igh nea, e a ne a, igh ea nea, lef fa, fo n fa, l ef n ea, fon nea, lef ea ele v fa, lef ea Node 11 lef nea fa, lef e a CoidoA SdooA 8 ep, fon le f neaby ne a, lef fon DooC N ode 6 nea, igh fon nea, ig h fon Doo D lef neaby Window B DooA le f neaby lef nea SdooB lef ne a 4 e p, fon BoadA nea, ea nea, lef ea BoadB Node 12 nea, fon nea, fon BoadC ne a, igh fon DooA DooA 3 ep, fon 4 ep, fon PanelB Do ob ne a, ea Nod e 5 lef neab y nea, igh Node 13 nea, igh nea, lef Coid oc DooC ne a, igh ea igh neaby 3 ep, fon Co ido A fa, l ef Node 4 lef n eaby fa, lef fon PanelB Doo A DooB WindowA Window A CoidoA lef nea igh nea le f nea igh nea by DooA DooB PanelA Window A nea, igh ea neaby, fon 5 ep, fon nea by, igh No de 3 nea, igh ea Doo A nea, lef ea neaby, ig h ea Do ob PanelB nea, lef fon nea, e a neab y, e a nea, fon nea, igh fa, igh fon 2 ep, igh Node 2 4 ep, igh 2 ep, lef fon 2 ep, fon Node 15 nea, igh fo n Pane lb BoadD 7 ep, fon BoadB fa, lef No de 14 igh ne aby fa, lef fa, igh PanelB Windo wa Fig. 9. Reul of opological-emanic diance-map building. Fig. 1. Example of obo-ceneed onology chema and map inance. [7] P. Beeon, and B. Kuipe, Towad Auonomou Topological Place Deecion Uing he Exended Voonoi Gaph, Poceeding of IEEE ICRA, pp , 25. [8] V Mio, J., Zhou, and W. Dianayake, Towad viion baed navigaion in lage indoo envionmen, Poceeding of IEEE he IROS, pp , 26. [9] S. Tully, H. Moon, D. Moale, G. Kano, and H. Choe, Hybid Localizaion uing he Hieachical Ala, Poceeding of IEEE he IROS, pp , 27. [1] B. Kieg-Buckne, U. Fee, K. Luich, C. Mandel, T. Moakowki and R. J. Ro, Specificaion of an Onology fo Roue Gaph, Lecue noe in Compue Science, pp , 25. [11] J. Baeman and S. Faa, Modelling Model of Robo Navigaion Uing Fomal Spaial Onology, Lecue noe in Compue Science, pp , 24. [12] Y. Ke and R. Sukhanka, PCA-SIFT: A Moe Diincive Repeenaion fo Local Image Decipo, Poceeding of Compue Viion and Paen Recogniion, 24. [13] I.H. Suh, G.H. Lim, W. Hwang, H. Suh, J.H. Choi, and Y.T. Pak, Onology-baed Muli-layeed Robo Knowledge Famewok (OM- RKF) fo Robo Inelligence, Poceeding of he IEEE he IROS, 27. [14] F. Dellae, D. Fox, W. Bugad, and S. Thun, Mone Calo Localizaion fo Mobile Robo, Poceeding of he ICRA, [15] H. Choe, K.M. Lynch, S. Huchinon, G. Kano, W. Bugad, L.E. Kavakij, and S. Thun, Pinciple of Robo Moion - Theoy, Algoihm, and Implemenaion, MIT-Pe, 25. [16] S. Vaudevan and R. Siegwa, A Bayeian appoach o Concepualizaion and Place Claificaion: Incopoaing Spaial Relaionhip (diance) o Infe Concep, Poceeding of he IEEE IROS, wokhop Fom Seno o Human Spaial Concep (FS2HSC), 27. [17] G.H. Lim and I.H. Suh, weighed Acion-coupled Semanic Newok (wasn) fo Robo Inelligence, Poceeding of he IEEE/RSJ Inenaional Confeence on Inelligen Robo and Syem (IROS), 28. [18] M. E. Munich, P. Pijanian, E. D. Benado, L. Goncalve, N. Kalon, and D. Lowe, SIFT-ing Though Feaue wih ViPR, IEEE Roboic and Auomaion Magazine, pp , 26. [19] T. Bailey and H. Duan-Whye, Simulaneou Localizaion and Mapping (SLAM): Pa II, IEEE Roboic and Auomaion Magazine, pp , 26. [2] N. Kalon, E. D. Benado, J. Oowki, L. Goncalve, P. Pijanian, and M. E. Munich, The vslam Algoihm fo Robu Localizaion and Mapping, Poceeding of he ICRA,