matlab中英文翻译文献

附录英文原文Sceｎe rｅcognitｉon for minｅrescue robotｌocalizａtｉon bａsed ｏn visionAbstract：Ａneｗsｃeneｒecognitiｏn system waｓpｒeseｎｔed bａｓe ｄoｎｆuzzy ｌoｇiｃand hiddｅn Markov mｏｄeｌ（HMM) tｈat caｎb ｅaｐpｌiｅd in minｅresｃｕe robot loｃalizａtion ｄuｒing ｅmergｅｎc ｉｅs. Thｅsyｓtem usｅs monocuｌar ｃamera ｔo aｃquire oｍni－ｄireｃtional images oｆｔhe mine eｎvironment wｈeｒe the rｏbｏt locａteｓ. Ｂy ａdoｐｔｉｎｇｃｅｎter-surround ｄifferencｅmethｏd, tｈe ｓａliｅnt locaｌimage regioｎs ａre extractｅd from the imａges as nａtural landｍａｒｋs．Ｔｈeｓe landｍarｋs are oｒｇａｎｉzed by uｓｉng HＭM to repｒｅsｅnｔｔhe scene wheｒe ｔhe roｂｏt is,and fuzzｙlogiｃｓtrateｇｙｉs used to mａtch the sｃene aｎdｌandmark．By thｉs way, the lｏcaｌizaｔiｏｎproｂlem，which ｉs the sceneｒeｃｏｇnitiｏn probｌem in t ｈｅsysｔem，can be conｖerted into the evａｌuationｐｒｏbｌｅｍｏf HMM. Ｔｈe contriｂｕtｉonｓoｆtｈeｓe skiｌls make the system ｈａｖｅｔhｅａbｉlityｔｏdｅal with changes iｎscale, 2D rotａｔion and viｅwpoｉnt. Tｈe reｓulｔs oｆｅxperiｍｅｎｔsａｌsｏproｖe thａt the system haｓhigher rａtio ｏf reｃｏgnition aｎd locａｌizatｉoｎin botｈｓtatiｃanｄdynamiｃmｉne environments．Keｙwords: roboｔlocatｉｏn;scene recognｉtiｏn；sａlｉenｔｉmａｇe; mａtching sｔrａtegｙ；fuzzy ｌogic; hidden Ｍarkov mｏdｅｌ1 IntｒｏdｕctioｎSｅarch and ｒesｃｕｅｉn disasteｒarea iｎｔhe domain oｆｒobot i ｓ a burｇeoning and chaｌｌengｉng suｂjecｔ[1]. Minｅrｅscue robｏｔwas devｅloped tｏｅntｅr miｎeｓdｕring emerｇｅnciｅs to loｃaｔe possｉblｅeｓｃａpe ｒoｕtｅs for ｔhose tｒapped ｉnｓiｄe and ｄeｔermine whｅｔheｒｉｔiｓsaｆｅfｏr humａｎｔo enteｒorｎoｔ. Loｃalizaｔion is a fundamentａl ｐｒｏblｅm in thｉs ｆielｄ．Lｏcａlizaｔion metｈoｄｓbaｓed ｏn caｍera caｎbe mａinly classiｆied intｏgｅoｍetr ｉc, topoloｇiｃal ｏr ｈybrid onｅs[2］．With ｉｔs feasibility ａnｄeffectiv ｅneｓs, sceｎe rｅcogniｔｉon becｏmes oｎｅof ｔhe imｐortａnt tｅcｈnologies of topｏlｏgiｃａl localｉzａtiｏn．Cｕrrentｌｙmｏst scene rｅｃognｉtiｏn methods are baseｄoｎgloba ｌimａge fｅatｕｒes aｎｄｈａve tｗｏｄistiｎct stages:ｔraininｇｏｆf ｌine and ｍaｔｃhinｇoｎｌｉnｅ.During tｈe training sｔage,roboｔcollｅctｓtｈｅiｍages of ｔhｅeｎviｒonmenｔwhere itｗorｋs aｎd prｏcｅssｅs the iｍagｅs to eｘtrac ｔgｌoｂaｌfeatuｒeｓｔhat ｒepｒｅｓｅnt the ｓｃene. Somｅappｒoacｈeｓwｅre used to ａnalｙｚe thｅdａtａ-set of ｉmage dｉrectｌy ａｎd ｓｏｍe ｐriｍａry feａｔures were found, suｃｈas tｈe PCA ｍｅｔhod [3］. Howevｅr，tｈｅPCA ｍethoｄiｓnｏt efｆeｃtive in dｉｓtin ｇｕiｓhiｎg tｈe clａsses of ｆｅａtures．Anotherｔyｐｅofａｐproach ｕseｓａｐpeａrａnｃｅfｅatures including coｌor, texｔｕre anｄｅdｇe deｎsity toｒepresent ｔｈe imａｇｅ．For eｘample, ＺＨOＵe ｔal[４］usｅd ｍultiｄimensionａl histｏgrams ｔo descｒibｅgｌｏｂａl ａpｐeaｒance features. This method iｓsｉmple bｕｔsｅnsｉtiｖｅｔo ｓcａｌｅand iｌluｍinａtｉon chanｇｅs.Ｉnｆａcｔ, alｌkiｎdｓoｆglobal iｍage features aｒe sufｆereｄfrｏm the cｈange of envirｏnｍent.LOWE [5] ｐresenteｄa SIFＴmethoｄthａt uses ｓimilａrity invarｉant desｃripｔors fｏrmed ｂy ｃharacteristic sｃａlｅａｎｄorｉｅntatｉoｎat intereｓt pointｓto obｔain tｈｅfeaｔureｓ.The ｆｅaturesａre iｎｖariａntｔｏimagｅscaling, ｔraｎslaｔｉｏn,rotａｔｉoｎanｄpaｒｔiaｌly invａriant tｏilluｍinａtion cｈanges．But SＩFTｍay generate 1000 or mｏre intereｓt ｐoinｔs, ｗhich may sloｗdown tｈe pｒocesｓoｒdramaｔically.Duｒing thｅmatching stage, ｎeareｓt neiｇhboｒstrａtｅｇy(ＮN）is wideｌy adopted for its faciｌｉty and ｉｎtｅlｌigｉbilｉty[6]. But it caｎｎoｔcaptuｒe the cｏnｔriｂｕtiｏn ｏfｉndiviｄｕaｌfｅature for ｓcenｅreｃognitioｎ．Inｅxｐeｒiｍｅntｓ, theＮN ｉs not goｏd enough to ｅxpreｓｓtｈe sｉmiｌarity ｂeｔween twｏpaｔternｓ. Ｆuｒｔhｅrmo ｒe，ｔhe seleｃteｄfeaｔｕresｃan not reｐreseｎt the scene thoroughlｙａc ｃｏｒding ｔｏtｈe sｔａte-ｏf-ａｒt ｐaｔterｎrecｏgｎition, ｗhicｈmakes recognitioｎnot reliａbｌｅ［7].Ｓo ｉnｔhis worｋ a new recogｎｉtiｏn sｙstem is pｒesｅnｔed，wh ｉch is ｍore reliaｂle and ｅffectｉve if ｉtｉsｕsed ｉn a complexｍin ｅenviroｎmenｔ. In ｔｈis syｓtｅm，wｅimprove ｔhe invariａｎce by extrａctinｇsalieｎt lｏｃal image rｅgions as ｌａndｍａrkｓｔo replace ｔｈｅwhole iｍage to dｅal ｗitｈlarge changｅs iｎscａle,2D rotatio ｎand viewpoint. Ａｎd ｔhe nｕｍber ｏf ｉnteresｔｐｏints is rｅduceｄｅffecｔiｖely,ｗhiｃh makes ｔhe processing easier．Fuzzy rｅcognｉtion ｓtrａｔegy is designed to ｒｅｃｏgnizｅｔhe laｎdmａｒks iｎplaｃeｏf NＮ, wｈich can streｎｇthｅn the contriｂuｔion of indiｖｉdual fｅature for sｃene ｒecogｎitiｏn．Ｂecauｓe of iｔｓparｔial iｎｆormatｉon ｒｅsuｍing ability, ｈidden Markov mｏｄel is aｄopted to ｏrｇanizｅthｏｓe ｌａｎｄmａｒｋｓ, whｉｃh cａn ｃapｔure tｈe ｓtructure orｒelatiｏnshｉp aｍong them．Sｏsｃeｎe ｒｅcｏgniｔiｏｎｃaｎbｅｔransformed ｔo tｈｅevaluａｔiｏn probｌem of HＭＭ,whｉcｈmａkes recｏgnition robust.2 Salienｔｌocａl iｍagｅregions detecｔｉonRｅseａrcheｓｏｎbｉoｌogｉcal vｉsｉon system inｄiｃａｔe that ｏrganism(likｅdroｓophiｌa) oftｅn ｐaｙs attenｔｉoｎto certaｉn ｓｐeci ａｌregions in the scenｅfoｒtｈｅｉｒbeｈａvioral reｌevance or local i ｍage ｃues whｉｌｅｏbserｖｉnｇsurrｏundings [8]. These ｒegions ｃan be taｋeｎas naｔｕｒal lａndmaｒks to effeｃtiｖelｙrepresenｔａnd ｄisｔｉngｕｉsh difｆerenｔenｖironmeｎtｓ．Inspireｄby tｈose，we ｕse cｅnteｒ-ｓurrounｄdiffｅrence ｍethod to deｔeｃｔsalient regｉons in mｕｌtｉ－scale ｉmａge spaces．Ｔhe opponencieｓof color and textｕre a ｒe compuｔeｄtｏｃreａteｔhe saliency ｍap.Follｏw-up,sｕｂ-ｉmａｇeｃｅnteredａt the ｓalient position in S ｉs ｔaken aｓthe lａnｄmark region. The ｓizｅof ｔhe ｌandmark rｅgio ｎｃan ｂe ｄecided adaptｉvｅlｙaｃｃoｒding ｔo ｔhｅcｈaｎges of ｇraｄient orientatiｏn ofｔhe local iｍage[11].Moｂiｌe robotｎaviｇation rｅquirｅs thaｔnａtｕrａl landmarｋs ｓｈould be deteｃｔed sｔａbly wheｎenvironments cｈaｎge ｔo somｅextｅnt. To ｖalｉdａte theｒepeaｔａbilｉty on laｎｄmark ｄetection ｏｆｏuｒaｐprｏａcｈ，we have done some expeｒiments ｏｎthe cａses oｆscale, ２D rotation and viｅwpoinｔｃhaｎges eｔc. Fig.1 ｓhoｗs tｈaｔtｈｅdｏoｒis ｄeｔectｅd for iｔs saliency wｈen viewｐｏint changeｓ. More dｅtaｉlｅｄａnalyｓiｓand rｅsults ａbouｔscａle and rotation caｎｂe ｆｏｕnd iｎour prｅvｉoｕｓworks［１2].3 Scenｅｒeｃogｎｉtionａnd lｏcalizaｔionDifferenｔfｒｏm other scｅnｅreｃoｇｎｉtｉoｎｓysｔems, our sｙstem doｅｓn’t need ｔｒainｉng ｏｆｆlｉｎe.In otｈerｗordｓ,ｏur sｃenｅs are not classifieｄiｎaｄvanｃe.Wheｎrｏbotｗａnders, scｅnes ca ｐtｕｒeｄaｔinｔerｖals of fixed tｉmｅaｒe usｅd to build the vｅｒｔeｘof ａｔopological maｐ,which repｒｅｓｅnts the plａce whｅre robｏt lｏcａtes. Althｏugh ｔhe mａp’ｓgeｏmetrｉｃlayｏut iｓｉgnorｅｄｂｙｔhe ｌocaｌiｚaｔion ｓysteｍ, iｔis useful fｏr visualｉzaｔion anｄdeb ｕgging[13]and benefiｃｉａｌtｏpaｔh planｎｉng．Ｓo localｉｚation m ｅａnｓsearchingｔhe bｅst match of curｒeｎｔscene oｎｔhｅｍａp.Ｉn ｔhiｓpaper hidｄeｎＭarkov model ｉs useｄto organize the eｘtｒａcｔed landmａrks from currｅnt scｅne anｄcreaｔe thｅｖeｒtｅx o ｆtｏpｏｌｏgicalｍap foｒｉts paｒtｉａl iｎformaｔioｎresuming ａbi ｌiｔy.Rｅsｅmｂled ｂy pａnoramic vision ｓysｔem, rｏbｏt loｏks arｏｕnｄto get omｎi-ｉmaｇｅｓ. FrｏmFiｇ.1 Eｘpｅriｍenｔon vieｗpｏint changeｓeａch imａgｅ, salient locａl rｅgioｎs are detecｔeｄaｎd ｆoｒmed to be a sequencｅ, ｎａmed aｓlａndmark sequeｎcｅｗhose order is tｈe sａmｅas thｅiｍage sequencｅ. Then a hｉdden Maｒｋov modeｌｉs ｃreat ｅd baseｄon the landmａｒk ｓeｑueｎｃe ｉnvolviｎｇk ｓａlient local ｉｍage regiｏｎs，ｗhｉｃｈｉs taken as the descｒiptｉon of tｈｅplaｃｅwhｅｒe tｈe rｏbot loｃateｓ. In oｕｒsｙsteｍEVI-D70 caｍeraｈas a view fiｅｌd ｏf ±１70°．Coｎsidｅｒｉng the oｖeｒｌap effect,we sampleenｖｉrｏnｍent every 45° to gｅt８iｍages.Let ｔｈe ８ｉmａgｅs ａs ｈiddｅｎsｔａtｅSi(1≤i≤８）,the cｒｅated HMＭcan be illustrａted by Fｉｇ.2．The paｒameｔers ｏf HＭM，aij anｄｂｊk, arｅａchｉeｖed by learning, using Baulm-Wｅlch aｌgｏｒiｔh ｍ［14］.The threshold of ｃｏnｖergence is sｅt as 0.0０1.Aｓfor tｈe eｄge of ｔopolｏgiｃａl map,we assign it witｈdｉｓｔａｎcｅinfｏrｍaｔｉoｎbetwｅen two vertices. The disｔances can be ｃo ｍputed accorｄiｎg tｏｏdｏmetry readiｎgｓ.Fig.2 HＭM ｏf enｖｉronmｅnｔTo locate iｔself on ｔｈｅtｏpoｌogicａl map,roｂoｔｍuｓt ru ｎｉtｓ‘eye’ onｅnviｒoｎment aｎｄextrａｃｔ a lanｄｍaｒk sequence Ｌ1′ −Ｌk′ , theｎｓearｃh the map for the beｓtｍatｃｈed verteｘ（sｃｅne). Difｆerｅnt ｆｒom traditional ｐrobａｂilｉstic lｏcalizatｉon[15], in ｏuｒsyｓｔem loｃａlizationｐrｏblem ｃａｎbｅｃoｎvertｅd tｏthe eｖaｌuatｉoｎprｏｂlｅmｏf HＭM. The vｅrtｅx wｉｔｈthe g ｒeateｓt ｅｖaｌuation vaｌue, which musｔalsｏbe ｇｒｅatｅｒthan a thrｅshｏld，is taken as the best matchｅｄvertex，whichｉｎdicaｔes the most ｐossible placｅｗhere thｅｒｏbot iｓ.4 Ｍatch ｓtｒatｅgy ｂaｓed on fuzzyｌogicOne oｆthe kｅy issｕes in iｍａge ｍatcｈpｒｏｂｌeｍｉｓtｏｃhoosｅthe most ｅffectivｅfeaｔures oｒdeｓcｒｉptors ｔo repreｓｅｎt thｅoriginal ｉmage. Dｕe tｏrobot movement,ｔhose extｒactｅd landmaｒｋregｉｏns ｗｉll changｅaｔｐixel lｅveｌ. Sｏ, tｈe deｓｃｒiptoｒs oｒfeaｔuｒeｓchｏsen shｏuld be ｉnvａrｉanｔto soｍe exteｎt accｏrdiｎg tｏｔhe changes ｏf ｓcaｌｅ，rotaｔioｎaｎd vｉewpointｅtc. Ｉn ｔhisｐaper, wｅuｓe 4 feaｔures comｍonly adoptｅd in thｅｃommunｉty tha ｔａｒe brieflｙdesｃriｂed as ｆoｌlows.ＧO:Gｒadient orｉentation.It hａｓbeeｎproved ｔhａｔiｌluminati ｏｎand ｒotaｔion changｅs are likelｙto hａve ｌess influence oｎit ［5].ASMａnd ENT：Angular sｅcoｎｄmomｅnｔａnd entropy，whｉchａre twｏtexture descriｐtors.Ｈ: Hue,which ｉs used to ｄesｃrｉbe the fundaｍｅｎtａl ｉnformation of the iｍaｇe．Anotheｒｋey issue in match pｒｏblem iｓto choｏse a good match str ａteｇy or aｌgorithｍ.Usually neaｒest ｎeighｂor ｓtrａｔegy （NＮ) isｕs ｅd to measurｅtｈe similariｔy betｗｅen twｏpatterｎs. But we ｈａvｅfound ｉn thｅｅｘperimenｔs that NN can’t ａdｅｑuaｔｅｌy exhibit thｅindｉvidual descriptor orｆeatｕre’s cｏntｒibuｔioｎto similａrｉty measuremeｎt．Aｓinｄiｃａｔeｄin Fig．4, ｔhｅinput imａｇe Fig．4(a) cｏmes froｍｄｉfferent vieｗof Fig.４（b）. But tｈe dｉstaｎcｅbetw ｅen Ｆiｇs.4（a）aｎｄ(b）compuｔed by Ｊeffereｙｄｉvｅrgenｃe ｉs larger ｔhan Ｆig.4(c)．To soｌve ｔhe problem,we design a new matcｈalｇoｒiｔhｍbased on fuzzy loｇｉc ｆoｒｅxhibiting ｔｈe subtｌｅchａngeｓof eaｃｈｆe ａtures. The alｇorithｍis described asｂｅlow.And the landmａrk in thｅdaｔaｂase ｗhｏｓeｆｕsｅｄｓｉｍilarit ｙdegree iｓｈighｅrｔhaｎany othｅｒs is taｋen ａs tｈe best maｔc ｈ．Ｔhe mａｔchｒesuｌtｓof Ｆｉgｓ.２(b)and(ｃ) are dｅmonsｔｒa ｔed by Fig．３.As ｉndicatｅｄ，this ｍeｔhｏd caｎmeａｓure tｈe similar ｉｔy eｆfectiｖｅｌｙbeｔwｅeｎtwo patterｎs.Fig.3Simｉlarity compｕtｅｄusiｎｇfuｚｚy stｒａtegy５Ｅxｐeriments and ａｎalｙsｉsThｅｌocaliｚａtion syｓｔｅm has been iｍｐleｍeｎted oｎａmｏbile robot, whｉch is buiｌｔby ｏuｒlabｏrａtｏｒy.Ｔhｅvｉsion syｓｔem is composeｄof ａＣCD ｃａmerａaｎdａfraｍe-grabｂeｒIVC-420０. Ｔhe resoｌutiｏnｏｆimage iｓsｅt to bｅ4０0×32０and thｅｓample fｒeqｕｅnｃy iｓseｔto ｂe 10 frames/ｓ. Thｅｃoｍputeｒs ｙｓtem is composｅd oｆ１GHz processor and51２M ｍemory，which i ｓcarriedｂyｔhｅｒｏbｏt. Preｓenｔly the robot works in ｉｎｄoor ｅｎviｒonmｅｎts.Becauｓe HＭM is adopｔed ｔo rｅｐresent aｎd recogｎize ｔhe ｓcene, oｕr ｓyｓtｅm has the aｂiliｔｙto capture tｈe ｄｉsｃriｍiｎatioｎaboｕt ｄistributiｏn of ｓaliｅnt local ｉmagｅregiｏns and dｉstinguish siｍｉlarｓｃenes effecｔively. Tａble 1shows tｈｅｒecｏgnｉｔｉon ｒｅsult of statｉc enviｒｏnmｅnｔs ｉnclｕdｉng ５ｌaneｗａys aｎd a silo. 10 ｓcｅnｅs are selectｅd ｆｒoｍeaｃh enviｒonｍent and HMMs arｅｃｒe ａted for each sｃene．Theｎ２0 sｃｅｎes arｅｃolｌｅctedｗｈen the robｏt enters eacｈenｖironｍentｓuｂseqｕｅnｔｌy toｍatch the 60HMMs aｂｏｖe.Ｉn tｈe table, “truth” ｍeans that the scene to be localiｚed matcｈｅs ｗitｈtｈe right scｅne （ｔhe evaluａtioｎvaluｅｏfＨMM is ３0% greａter tｈaｎthe second hｉgh ｅvalｕatｉoｎ). “Ｕncｅrtaｉnｔｙ” mｅanｓthat ｔhｅevaluatiｏn valｕｅｏf HMＭiｓｇreaｔeｒｔhaｎtｈeｓeconｄｈｉgh eｖａluatiｏn ｕnder 10%.“Error match” mｅaｎｓｔhaｔthe sｃene to ｂｅloｃａlized matches witｈthe ｗｒong scene. In the tａｂle, the ｒａtio of error mａtch is 0. But it is possiｂle ｔhat tｈe sｃenｅto be loｃａliｚed can’t match any sｃeneｓａndｎｅｗverｔeｘes ａre created. Fuｒthｅrmore, ｔｈｅ“ratｉo of trｕtｈ” abouｔsiｌo is lｏｗer bec ａｕseｓaliｅｎt cues ａrｅfewer in tｈis kiｎd of eｎviｒonment.Iｎtｈe perｉoｄof aｕtomatｉｃexploriｎg, sｉmiｌar sｃenes can be coｍbineｄ. Ｔｈe prｏcesｓcａｎｂe summaｒizｅd aｓ: whｅn loｃaｌization succeｅdｓ, ｔhｅｃurｒent laｎdmarｋsｅquｅncｅis addｅd to theａccｏmpａnying obseｒvation sｅqueｎcｅｏf the matched vｅrteｘun-repeatedly according to theiｒorｉentaｔion （iｎclｕdｉng ｔhe angle oｆtheｉmaｇｅfrom wｈｉcｈthe salieｎｔlocal regｉon and ｔhe hｅadi ｎg of the roboｔcoｍｅ)．The paramｅters ofＨMＭarｅlｅarnｅd aｇain.Cｏmpaｒｅd wｉth tｈｅapproaches using appeａrａncｅｆeatuｒｅs of the whoｌｅimａge(Methｏd2, Ｍ2),ouｒsｙstem (Ｍ1）uses loｃａl salｉｅnt regｉonｓto localizｅanｄmap, whiｃｈmａkes it ｈａve moｒe ｔolerance of scale，viｅwpoiｎt chａnges cａuｓed by ｒｏbot’ｓm ｏvemeｎt aｎｄhｉｇher ratiｏoｆｒecｏgnｉtion andｆｅwer aｍouｎt of veｒticesｏn theｔoｐｏｌｏgicａl ｍap. So, ouｒｓyｓｔeｍhaｓｂｅtter pｅrfoｒmaｎcｅiｎdyｎamｉcｅnｖiｒoｎｍent.Ｔheｓe can be seen in Table 2. Lａｎewaｙs 1, 2, ４, 5 ａｒe ｉn operatioｎｗhere sｏme miners are ｗorｋinｇ, whiｃh puzzle ｔheｒobｏt.6Ｃonclusions1) Salｉeｎt locａｌimagｅfeatuｒes are ｅxtrａcted to ｒeplaｃe the wholｅｉmagｅto partiｃipａtｅin ｒecogｎition, ｗｈicｈimprｏvｅthｅｔolｅｒance of chanｇeｓin ｓｃalｅ，2Ｄrotaｔｉon and vieｗｐｏiｎt of envirｏnｍｅnｔimage.２) Ｆｕzzy logic iｓusｅdｔｏrｅcognize the local ｉmage, ａnd eｍphａsize thｅindiｖiｄｕal feａture’s ｃontriｂｕtｉoｎｔｏrｅｃognition, whichｉmproveｓｔhe ｒｅｌiabiliｔｙof lanｄmａrks.3）ＨMＭｉs usｅｄto ｃaptuｒｅthe ｓtructｕre or ｒeｌationsｈip of tｈose loｃaｌｉｍaｇeｓ，ｗhich cｏnverｔs the scene reｃoｇniｔion proｂlem ｉnto ｔhｅｅvaluation problem of HMＭ.4）The rｅsuｌts ｆrｏm the abｏve exｐeriｍentｓdｅmoｎstｒatｅt ｈaｔｔhe ｍｉnｅrｅscue roboｔsｃene recognition sysｔｅm ｈas highｅr ratio of recｏｇnition and localization．Future wｏrk ｗill ｂｅｆoｃused on usiｎg HMM toｄeal witｈｔheｕnｃｅrtainｔy ｏf locａlizａtｉｏｎ.中文翻译基于视觉的矿井救援机器人场景识别摘要：基于模糊逻辑和隐马尔可夫模型（HMM)，论文提出了一个新的场景识别系统，可应用于紧急情况下矿山救援机器人的定位。