基于高斯混合背景模型运动检测与DSP实现

中南民族大学硕士学位论文基于高斯混合背景模型运动检测与DSP实现姓名：王飞申请学位级别：硕士专业：通信与信息系统指导教师：谢勤岚20100501IDSPIIAbstractWiththedevelopmentofcomputerscience,mathematicandotherdisciplines,machinevisioniswidelyusedindailylife,industries,nationaldefense,ect.Intelligentsurveillance,forwhichthemostimportantthingistofindaquick,accurateandrobustmethodformotionobjectsdetection,isoneapplicationofmachinevision.Thestudiesofthepastseveraldecadeshavegainedfruitfulresults,whichoffersatheoreticallysupportfortheimplementofintelligentmonitoringsystem.Atthesametime,theDSPandothershigh-speedprocessors'provideasupportfortheembeddedrealizingofcomplexalgorithmsfromhardware.Inthispaperthetraditionalmethodsformotiondetectionareintroduced,andthealgorithmaboutGaussianbackgroundModelisstudiedintensively.Andthemaincontentsofthispaperareasbelow:PaperintrudesthemethodofmotionobjectsdetectionbasedonsingleGaussianBackgroundModel.AnddeterminestheGaussiandistributionofeachpixelwithstatisticalmethodduringthebackgroundmodelingphase;thenseparatesthebackgroundandforegroundpixelswithmaximumsimilaritythresholddecision;atlastupdatestheparametersofthemodelofbackgroundwithK-meansmethod.Inthispaper,westudytheGaussianMixedbackgroundModel(GMM)formotionobjectsdetection.AndeachpixelisexpressedbyKindependentGaussianmodels.EachpixelofthenewimageiscomparedwiththeKbackgroundmodels,thematchedonesarerecognizedasbackgroundpoints,ortheyarerecognizedasforegroundpoints;Parametersofthebackgroundmodelareupdatedbyadaptively;allthepointswhicharejudgedastheforegroundconstitutethemovingobject.WeproposesomeimprovementsfortheGMM.AndestablishtheGMMwithfirstframeimageasaGaussianmodelwiththecurrentmeanvalueandtheinitialvariance,weight1foreachpixel,theremainingK-1GaussianModelsareinitializedbyemptyimages,withthemean-0,variance-theinitialassignmentandweight-0;Andtheparametersareupdatedbydynamiclearningconstant.Experimentsdemonstratethesignificantincreaseofthereal-timeandaccuracy.AnexperimentisachievedbasedonTi'sTMS320DM642processorplatform.Manytimesunderdifferentenvironmentalexperimentsprovedthatthesystem’sreal-time,accuracyandstability.Keywords:motiondetection,GMM,match,OnScreenDisplay(OSD)1______211.11.2[1][2][3-5]DSP21.3DSP7720PDm1.43DSP42.15ChiarellaDSP62.22.1()⎩⎨⎧Τ≥--=elsetyxItyxIyxd0)1,,(),,(1,),(yxd()tyxI,,tT72.22.3基于高斯混合背景模型运动对象检测及DSP实现8图2.4三帧差分法仿真结果图时间差法以其计算简单容易硬件实现，在简单背景下的运动检测得到广泛应用。但其受光照变化、相机抖动和阈值选取等条件的限制，当对检测效果要求较高时不宜采用。2.3背景减除法背景减除法是运动检测与跟踪领域研究的热点，它对简单背景与复杂背景下运动对象检测均能起到较好的效果，较有较好的鲁棒性，得到广泛的应用和研究。背景减除法的基本思想是，通过统计图像序列在一段时间内各像素点的像素值的平均分布，从而建立昀接近真实环境的背景模型，将当前帧图像和背景参考模型进行逐像素比较，按昀大可信度估计方法，估计出背景点与前景点，提取运动对象；同时，为了保证检测结果的准确性，还要对模型进行更新，不断调整背景各参数，使得背景模型尽可能地模拟真实背景情况。背景减除法是目前运动对象检测比较常用的方法，特别是在背景相对静止的场合应用较多。背景减除法主要的几个步骤分为：背景模型的建立，背景模型的更新，背景差和后处理等步骤。一种比较简单和常用的背景建模运动检测方法是，将当前帧与背景图像相减，然后对差分图像进行阈值处理，从而分离出前景点。对于像素值的昀合理模拟需要背景是完全静止的而且模型可自适应更新。一个好的背景模型的建立，要能效地解决运动检测中常常存在的一些问题：1）背景的扰动、光照变化、摄像机抖动、噪声干扰等：如树枝、树叶的摆动，水面92.4DSP10113.13.1.1()()∞∞-=--xexfx,21222smsp()0,ssm()2,~smNX()xf3.13.1a.m=x5=mb.m=x()spm21)(max==fxfxm()xfmXsm±=xOxsmOx3.2DSP123.2m3.2()xfy=mms()()spm21=fs3.3Xm3.3s3.1.2()YX,()()()()()()∞-∞∞∞-⎪⎭⎪⎬⎫⎥⎦⎤-+⎪⎩⎪⎨⎧⎢⎣⎡-------=yxyyxxyxf,,2121exp121,2222212121212221smssmmrsmrrspsrssmm,,,,212111,0,021-rss()()rssmm,,,,~,222121NYX()()∞∞-=--xexfxX,21212121smsp()()∞∞-=--yeyfyY,21222222smsp13r2121,,,ssmmrXYXY()yxF,()()yFxFYX,()YX,yx,{}{}{}yYPxXPyYxXP≤≤=≤≤,()()()yFxFyxFYX=,()YX,,XY0=r()nXXX,...,,21()nxxxF,...,21()nXXX,...,,21()nkk≤1nxxx,...,,21()()()()nXXXnxFxFxFxxxFnL212121,...,=nXXX,...,,21()'21,...,,nXXXX=)}()'(21exp{||)2(1)(12/12/mmp---=-xCxCxfn()()'21'21,...,,,,...,,nnxxxxmmmm==()CX,~m()niNXiii,...,2,1,,~2=smnnXCXCXC+++L2211nCCC,...,210⎟⎠⎞⎜⎝⎛++∑∑==niiiniiinnCCNXCXCXC12212211,~smL[45]3.1.3X{}()Θ∈==qq,,xpxXPqqΘnXXX,...,21XnXXX,...,21()∏=niixp1;q{}nnxXxXxX===,...,2211DSP14()()()Θ∈==∏=qqqq,;;,...,,121niinxpxxxLLqqnxxx,...,,21qΘ()q;,...,21nxxxL∧qq∧q()qqq;,...,,max;,...,2121nnxxxLxxxLΘ∈∧=⎟⎠⎞⎜⎝⎛3.23.2.1()iyxI,,,{}(){}tiiyxIXXt≤≤=1:;,,,1Lm2s∑==tiiXt11m()∑=-=tiiXt1221ms3.2.215()()⎩⎨⎧Τ-,,,,yxyxIm()yxI,()yx,Tlsl3.2.32sm()()()()()yxIttt,111+++-=amam()()()()()()()21221,1ttttyxImasas-+-=++a()()()()()21,|,tttyxIsmbha+=DSP16b()()()()()21,|,tttyxIsmh+()tm()2ts3.33.3.13.4()∑,,mhtXm∑{}tXXX,...,,21()KiXtitit,...,2,1,,,,=∑mhtX()()∑∑==KititittitXXP1,,,,,*mhv17∑titi,,,m⎟⎠⎞⎜⎝⎛=∑=11,,Kititivvh()()()()∑=----Τ∑∑ttttXXnteXmmpmh12121221,,Iktk2,s=∑I()2,,,,tititXsmhtX()()∑=×=KititittitXXP12,,,,|smhv2,tisti3.3.23.51titi,,ls=ΤKiXXtitittitit,...,2,1,,,,,,=⎪⎩⎪⎨⎧Τ≥-Τ-mmlDSP18k()bvbv+-=-1,,1tktkb[]1,0bbtk,m2,tks()()()()yxItktktk,11,,1,+++-=amam()()()()()2,1,2,21,,1tktktktkyxImasas-+-=++()()()2,,1,,|,tktktkyxIsmbha+=tXtX3.28()1,,1--=tktkvbvtX3KiKktktiti,...,2,11,,,==∑=vvv3.3.3ti,v2,tis2svsv2,,titisvtiti,,sv2,,titisvtiti,,svKBB⎟⎠⎞⎜⎝⎛Τ=∑=bkkbB1minargvΤ19ΤΤBtX3.4bbbbbKKK1K11K-102b3.31Nnnn,...,2,111=+==bbDSP203.5a1a(2)a(3)b(1)b(2)b(3)c(1)c(2)c(3)d(1)d(2)d(3)e(1)e(2)e(3)f(1)f(2)f(3)g(1)g(2)g(3)3.6Kaa(1)a(2)a(3)bdfK=135cegK=13521a(1)a(2)a(3)b(1)b(2)b(3)c(1)c(2)c(3)d(1)d(2)d(3)e(1)e(2)e(3)f(1)f(2)f(3)3.7Tace3.6K=5,T0.70.90.95bdfDSP22a(1)a(2)a(3)b(1)