基于自适应多特征融合的meanshift目标跟踪

整理文档很辛苦,赏杯茶钱您下走!

免费阅读已结束,点击下载阅读编辑剩下 ...

阅读已结束,您可以下载文档离线阅读编辑

资源描述

ISSN1000-1239CN11-1777TPJournalofComputerResearchandDevelopment47(9):1663-1671,2010:2008-07-01;:2010-01-27:(200801493;20080430223);(090412043)meanshift袁广林1,2薛模根1,3韩裕生3周浦城31(230009)2(230031)3(230031)(ygl6904@sina.com)MeanShiftObjectTrackingBasedonAdaptiveMult-iFeaturesFusionYuanGuanglin1,2,XueMogen1,3,HanYusheng3,andZhouPucheng31(SchoolofComputerandInformation,HefeiUniversityofTechnology,Hefei230009)2(SecondDepartment,ArtilleryAcademyofPLA,Hefei230031)3(ForthDepartrment,ArtilleryAcademyofPLA,Hefei230031)AbstractTraditionalmeanshifttrackingalgorithmhasachievedconsiderablesuccessinobjecttrackingduetoitssimplicityandrobustness.But,itmodelstheobjectstobetrackedwithsinglecolorfeature,whichleadstothatitismorepronetoambiguity,especiallyifthescenecontainsotherobjectscharacterizedbyacolordistributionsimilartothatoftheobjectofinterest.Toaddressthisproblem,aformulafortargetlocalizationwithmeanshiftbasedonmultiplefeaturesisdeduced.Inaddition,amethodthatevaluatesthediscriminabilityofeachfeatureswithrespecttoforegroundtobackgroundseparabilityandadaptivelycalculatesthefeaturesfusionweightbyprobabilityseparabilitycriterionisproposed.Withtheabovededucedformulaandproposedmethod,anovelmeanshifttargettrackingalgorithmbasedonadaptivemultiplefeaturesfusionispresented.Theproposedalgorithmisrunforeachfeatureindependentlyandtheoutputofthemeanshiftalgorithmforeachfeatureisweightedbasedonthefusionweight.Thestatesofthetargetinthecurrentframearecomputedthroughtheintegrationoftheoutputsofmeanshift.Experimentsareconductedwithcolorsequencesandgraysequences,anditsresultsshowthattheproposedalgorithmhasbetterperformancethantheclassicalmeanshifttrackingalgorithm.Keywordstargettracking;meanshift;featuresfusion;kernelfunctionhistogram;edgeorientationhistogram经典meanshift目标跟踪算法简单快速,具有较好的跟踪效果,但是它用单个特征描述目标,易受相似目标与背景的干扰,鲁棒性较差.针对此不足,推导出多特征融合meanshift目标定位公式;为了适应跟踪过程中目标与背景的变化,提出利用概率分布可分性判据动态评价特征对目标与背景的区分能力,并自适应地计算特征融合权重.在上述两个方面的基础上,对meanshift目标跟踪算法进行了改进,提出一种多特征融合meanshift目标跟踪算法.实验结果表明:提出的算法比经典meanshift目标跟踪算法具有更强的抗干扰性能和较高的跟踪精度.目标跟踪;均值漂移;特征融合;核函数直方图;边缘方向直方图TP391.410,,.20,.[1]meanshift,[2].;Bhattacharrya;,..[1]meanshift(MS),.[3]Lindeberg,meanshift.[4],.[5]Correlogram,,meanshift.[6]meanshift,.[7],meanshift.[8]meanshift.,,,.,.[2],.[9]RGB3,,,49,meanshift,,.[10],Fisher,.[11]AdaBoost,,meanshift,,.meanshift,.,,meanshift.(),;meanshift,;,.1(probabilitydensityfunction,PDF),.1PDF,1PDF.1.1q^={{q^ui}ui=1,,mi}i=1,,nk,{q^ui}iPDF,miibin,k,nk.{x*j}j=1,,n,k(x),bi(x*j):R2{1,,mi},x*jibin,q^ui:q^ui=Cnj=1kx*j2[bi(x*j)-ui],(1)Kroneckerdelta,C.1.2p^(y)={{p^ui(y)}ui=1,,mi}i=1,,nk,{p^ui(y)}iPDF,y,mi,k,bi()nk2.1.{xj}j=1,,nh,hk(x),p^ui(y):p^ui(y)=Chnhj=1ky-xjh2[bi(x*j)-u],(2),Ch.16642010,47(9)1.3q^i={q^Mi}ui=1,,mip^i(y)={p^ux(y)}ui=1,,mi:d(y)d(p^i(y),q^i)=1-^(p^i(y),q^i),(3),^i(y)^(p^i(y),q^i)=miui=1p^ux(y)q^uxBhattacharyya,0~1,.(3):dmf(y)=1-nki=1ki^i(y),(4),^i(y)=miui=1p^ui(y)q^uiBhattacharyya,ki,nki=1ki=1,2.2.2meanshift,[1-2],,meanshift,,.[1-2][3-8]meanshift.2.1meanshifty^0,y^0PDF{{p^ui(y^0)}ui=1,,m}i=1,,nk.(4),nki=1ki^i(y),nki=1ki^i(y)ki^i(y).ki^i(y)y^0Taylor:ki^i[p^i(y),q^i]12kimiui=1p^ui(y^0)q^ui+12kimiui=1p^ui(y)q^uip^ui(y^0).(5)(2)p^ui(y),(5):ki^i[p^i(y),q^i]12kimiui=1p^ui(y^0)q^ui+Ch2kinhj=1wijky-xjh2,(6):wij=miui=1q^uip^ui(y^0)[bi(xj)-ui].(7)(6)1y,(4),(6)2.meanshift,:kinhj=1wijxjgy^0-xjh2nhj=1wijgy^0-xjh2-kiy^1=0,(8)(8)g(x)=-k(x).nki=1ki^i(y)ki^i(y)nk(8),kinki=1ki=1,nk:y^1=nki=1kinhj=1wijxjgy^0-xjh2nhj=1wijgy^0-xjh2,(9)(9)meanshift.2.2.,,,,.,.,.,1(a),,1(c),.,[10]Fisher,.1665:meanshiftFig.1Featuresanditsdiscriminantcurve.(a)No109frameimage;(b)Gradsimageof(a);(c)No148frameimage;(d)Gradsimageof(c);and(e)Discriminantcurveofcolorfeature(dotted)andedgefeature(solid)inthisimagesequences.1.(a)109;(b)(a);(c)148;(d)(c);(e).AdaBoost[11],,.,,.,,,,.[7],1(a),,.(x0,y0),wh,:Step1.,{p^ti}i=1,,nk.Step2.{(x,y)|(x0-wb2)x(x0-w2)(x0+w2)x(x0+wb2),(y0-hb2)y(y0-h2)(y0+h2)y(y0+hb2)},wbhb,wb=hb=max(w,h).Step3.,{p^bi}i=1,,nk.Step4.ki=d(p^ti,p^bi)nkj=1d(p^tj,p^bj),i=1,,nk,,d(p^ti,p^bi)(3).1(e),().1(e),,,,,,.32meanshift,,meanshift.1.meanshift.:I1,I2,,In1y^0{{q^ui}ui=1,,mi}i=1,,nkki=1nk.16662010,47(9):y^2,,y^n.y^0,{{p^ui(y^0)}ui=1,,mi}i=1,,nknki=1kii[p^i(y^0),q^i]=nki=1kimiui=p^ui(y^0)q^ui.(7){wij}i=1,,nk,j=1,,nh.(9)y1.{{p^ui(y^1)}ui=1,,mi}i=1,,nknki=1kii[p^i(y^1),q^i]=nki=1kimiui=1p^ui(y1)q^ui.Whilenki=1kii[p^i(y^1),q^i]nki=1kii[p^i(y^0),q^i]Do(y^0+y^1)2y^1;nki=1kii[p^i(y^1),q^i]y^1-y^0,.y^1y^0.2.2,y^1,y^1y^0.4VisualC++6.0,P42.0GHzCPU,256MB.,meanshift.,.4.1,,,,[1-2,5-6].,.;,[11-13].,.,.,,.,.2.:RGB,b(x*i):R2{1,,m},x*iRGBbin;(2).:,xGx(x,y)yGy(x,y);(x,y)=arctan(Gy(x,y)Gx(x,y);b(x*i):R2{1,,m},x*ibin;(2)..4.2meanshift.meanshift,RGBbin161616,bin32,Epanechikov.500,12896,,3642.2.,100,303,422.meanshift,.3meanshift,,,1,38.2ms,meanshift,.1667:meanshiftTable1TrackingErrors(meanstd)andAverageTrackingTimeinColorHeadSequence1()TrackingAlgorithmXCoordinatorError(Pixels)YCoordinatorError(Pixels)TrackingTimemsClassicalMSAlgorithm12.79209.0011.038143.1721.8TheProposedAlgorithm4.7315.506.57620.0838.24.3meanshift.meanshif

1 / 9
下载文档,编辑使用

©2015-2020 m.777doc.com 三七文档.

备案号:鲁ICP备2024069028号-1 客服联系 QQ:2149211541

×
保存成功