《基于计算机视觉的三维测量技术文献翻译教学内容.doc》由会员分享,可在线阅读,更多相关《基于计算机视觉的三维测量技术文献翻译教学内容.doc(22页珍藏版)》请在taowenge.com淘文阁网|工程机械CAD图纸|机械工程制图|CAD装配图下载|SolidWorks_CaTia_CAD_UG_PROE_设计图分享下载上搜索。
1、Good is good, but better carries it.精益求精,善益求善。基于计算机视觉的三维测量技术文献翻译-重庆理工大学文献翻译二级学院专业班级学生姓名学号-译文:基于计算机视觉的三维测量技术摘要:本文根据计算机视觉原理,提出一种三维非接触测量技术。该技术根据人眼感知事物的原理,利用神经网络拟合图像坐标与空间坐标的映射关系;以光栅投影曲线为特征,采用小波边缘检测和搜索式无监督聚类,结合视觉几何不变性,实现亚像素级的立体精确匹配;并采用小波多尺度多分辨率的特性,拼接图像,融合数据,对物体进行全方位测量。实验表明,该技术设备简单,测量速度快,测量精度控制在0.5mm/m以内。
2、关键词:计算机视觉,立体匹配,几何不变性,神经网络,小波变换,聚类1 引言目前,三维测量仍以三维坐标测量机为主。但是它由于体积大、结构复杂而不能在线测量,是接触测量而不能测量柔软的物体。因此,研究快速无损、非接触在线测量在工业上十分重要。尽管现在有很多方法,如激光扫描法、结构光法、相位测量法,但是都不能同时满足测量精度、效率、成本、自动化和智能化等方面的要求。因此,在本文使用双摄像机融合光学轴抓拍物体。随着处理图像,立体匹配图像和数据集成,三维物体的信息就是从这个立体图像中获得。三维测量技术已应用于测量系统中的多点压成型机的测量,并取得了良好的效果。2 测量原理及系统设计本文介绍了基于计算机视
3、觉的三维非接触测量技术,三维对象的信息是从一对立体图像中获取。一般来说,有两个问题影响的三维物体获得确切的消息:一种是图像之间建立特殊点点和准确的映射关系,另一种是立体匹配问题。本文神经网络是用来映射关系接近的情况下摄像机标定。小波边缘检测,寻找非监督聚类和几何不变性适用于立体匹配。在多尺度,多分辨率的小波属性应用于图像拼接和数据集成。在实践中,这项技术包含了许多方法和技术,它可以测量任意大小和形状的对象。然而,有一些物体的表面很光滑。匹配功能不明显,因此用光栅对象预测。而扭曲的条纹上创建的对象被视为匹配功能。为了提高测量精度,用两个与融合光学轴相机,这两个相机和一小型自制的投影机就构成了一种
4、灵活的测量头。一个基于立体视觉的三维测量的原理草图如图1所示。3建立图像点和特殊点之间的映射关系实际上,获得从两个图像对三维物体的信息是获取图像点之间的映射和特殊点的关系,但是到现在为止没有任何方法可以完全描述非线性映射关系,因为有许多复杂的非线性影响因素,包括摄像的径向变形和横向变形。但是,神经网络可以模拟人类的视觉,建立了简单的非线性映射来处理复杂的单元,因此本文就从图像点的过程中当作黑箱特殊点。和BP网络的6个神经细胞中间层网络来设置点之间的形象和特殊点的映射关系。图片左边的点A和一个右边的点纳入BP网络,一个特殊的点被输出。换言之,这个BP网络的结构是4-6-3。利用神经网络,样本的选
5、择是很重要的。样本不仅在于衡量的范围,也显示测量系统的测量范围。虽然两个相机是用来抓拍对象,但是这部分对象只有在焊接处的视野内才能被获取。因此,物体三维信息的立体图像,镜头焦点的测量精度,测量范围和目标与摄像机之间的两个基准距离控制三维测量系统的测量范围。本文的结构和功能和两个相机是用来抓拍对象构成对称是相同的,相机的图像区域的是,如图2所示。该镜头的焦点是;两个图像之间的中心垂直线是。共同的部分被视为双摄像头的连接视野。而超出的部视为盲区。假设视野角度为2,基本的成像关系公式为:(1)这个内切圆是视野范围,如果两个相机光轴的夹角是,两个图像中心之间的距离是2,其比例为:(2)这样,一个2R2
6、R的示例模板由88的格子组成。这个示例模板固定在工作台上。分别获取三对立体图像,而示例模板沿垂直线方向移动到三个不同高度(0,R,2R)模拟三维测量范围。三对立体图像被视为训练样本,把它们输入网络。4 亚像素级的立体精确匹配对立体显示来说立体精确匹配要困难得多,所以申请采用立体显示在某种程度上受到限制。本文应用小波变换检测边缘点,寻找非主管聚类方法,提出以区分不同的边缘点群。在同一个点群的边缘点的二次曲线拟合,然后在立体精确匹配亚像素级的水平基础上取得几何不变性。41条纹边缘拟合中的非聚类搜索一般来说,图像往往含有随机噪声,小波变换能抑制噪声和检测移动,同时不同结构图像边缘的信息传播在所有决议
7、中。自从转化不变性是最重要的立体匹配的边缘特征。二次B-spine被用来处理一个多尺度的生成元素检测条纹边缘点。实际上,噪音仍然混合在这些离散边缘点中,因此,曲线拟合用于转化为连续曲线离散边缘点,并减少噪音。然而,在曲线拟合之前,至关重要的是,所有的离散边缘点根据图像中条纹边缘的实际情况分成不同的群。海明距离的聚类中心往往被视为约束条件群,换句话说,假设一个点的属性向量是,一个聚类中心的属性向量是,如果,n是聚类总数,这样的思想不符合的条纹边缘点的实际情况。在曲线拟合之前,不仅给定的群体,而且这组点属于已知,而群体数目与条纹边数相等。因此,在本文中提出了非主管聚类算法。如果D是一个集合点,n是
8、D点的数量,如果D分成组,划分方法如下所示。1)如果是属性向量,被称为初始群体,这里是,的组数等于n;2)假如=,结束;3)在覆盖下的基础上,两个群体之间的距离也就可以计算所有群体。假如,且(T代表转置矩阵),=min,最近的两组被选择;4)和是合并到,于是,所以群体总数减少;5)重复步骤(2)。42基于几何不变性的相应点搜索几何不变性的定义是几何图案和矢量保持精确不变。对于一个特殊的多边形,两种不同的成行将得到两种透视变换图像位面。以同样的方式,对于一个三维曲线,两种不同的二维曲线得到两个图像位面。因此,几何不变性应用于匹配直线和曲线。对于直线匹配,几何不变性由5个点在同一条直线或5条直线在
9、同一平面所代表。我们假设是特殊平面上的任意5条直线,直线方程为:(3)我们任意选择3直线,和在5条直线上(k1,k2,k3=1,2,3,4,5,k1k2,k1,k3,k2k3)。这三条直线方程给出为:(4)这些直线均按直线的角度转变成图像。直线的特征也转换相应的直线方程的参数。参数显示在上标处(例如)。它证明,尽管这连续的五条直线的形状可以有更多的变化,它们也服从几何不变性,如果M属于A,它们是:,(5)类似地,有一个组的二次曲线的一些几何不变量。如果这个特殊平面上的一条二次曲线,它的方程可以表现为如下的二次曲线:(6)如果是二次曲线的参数矩阵,它也表现为如下矩阵:(7)如果有两条二次曲线和,
10、它们的参数矩阵分别为和。运用几何投影将它们转化为和,其参数矩阵为和。它证明,如果是矩阵的轨道,有两个几何不变量不管几何投影模式是否变化。(8)(9)这样,直线和曲线就有效匹配了。本文光栅投影在垂直方向和水平方向被分别提出来,而两相机抓拍图像。随着小波边缘检测,搜索式无监督聚类,边缘点到二次曲线拟合。几何不变性,二次曲线匹配,垂直曲线和横向曲线交叉点的计算。因此,亚像素级的立体精确匹配得以实现。5 基于小波的图像拼接当大规模的测量表面时,许多对立体图象在不同的观点或者移动和旋转中被抓拍到。两个相邻图像需要镶嵌。图像镶嵌的重要问题是图像配准,也就是说,两个相邻图像之间的重叠部分,以便付诸表决,并且
11、两个相邻图像之间的相应匹配也是图像镶嵌的复杂工作。通讯匹配在相应的立体视觉匹配之后。在这之前,从相同的角度或者不同的角度沿着基本路线转换来抓住两个图像,并在这之后,这两张图片的角度不仅要是转换,而且要旋转。本文,一些随机黑点能容易的镶嵌,这些黑点被认为是重要的拼接点。同时,我们用线性和对称双正交分解两个图像来镶嵌,使粗糙的图像可以得到很好的匹配和拼接,最终得到一个大的图像。事实上,小波变换是一种带通滤波,小波向量的显示用不同尺度的频带宽度来衡量,所以每个小波的频率带宽是不相等的。两个图像用Mallat算法分解成不同频率波段的小波向量,然后不同规模选择不同的镶嵌宽度来满足和拼接,于是一个大的镶嵌
12、图便顺利且很好的完成了。6实验及结果分析在本次设计中,这项技术在MPF机的测量系统中得到了应用。在应用了该技术后,测量结果返回到CAD/CAE系统中显示闭环控制得到了实现。表面形状后测量,测量结果返回到CAD/CAE系统和闭环控制的实现。据测量条件、测量精度一旦成熟,我们选择两个摄像头(MTV1881CB),两个镜头和一个图像记录装置(METEOR)。这两个摄像头之间的距离为300毫米;物体表面和两部相机之间的距离为500毫米。A150150mm的曲面是该工艺的标准测量范围,测量结果在标签1上显示,测量步骤如下:1)建立与图像点和特殊点之间的映射关系;2)三维表面在工作台上进行,首先,二个摄像
13、机在没有干扰和光线的情况下同时抓拍一对立体图像。其次,在抓住两对立体图像对,一对在光栅的垂直方向上抓拍,另一对在光栅的横向上抓拍;3)进程映像,消除背景,减少噪音,如图3a,3b所示;4)功能检测,如图3c;5)搜索对应点,并镶嵌图像;6)计算三维坐标,重建三维表面,如图3d。实验表明,测量误差小于0.5mm,测量时间约2秒,包括图像抓拍、图像处理、建立图像点和特殊点的映射关系、搜索相应的坐标点和调整计算。图、3图像处理7结束语在本文中,提出了一种新的基于计算机视觉的三维测量技术,该技术设备简单、测量速度快、成本低。可以测量大型对象,测量精度低于0.5mm/m。它还提供了一个适用于工业计算机视
14、觉的新思路。实验结果表明,三维测量技术是非常完美的。原文:3DMeasurementTechnologyBasedonComputerVisionAbstract:Onthebasisofcomputervision,anoncontact3Dmeasurementtechnologywasproposedinthispaper.Usingneuralnetwork,themappingrelationbetweenimagepointandspecialpointwasestablished.Theprojectionofgratingonobjectisregardedasmatching
15、features,withwaveletedgedetection,searchingnon-supervisorclusteringandgeometricinvariance.Stereoprecisionmatchingisachievedatsubpixellevel.Furthermore,themulti-scaleandmulti-resolutionattributesofwaveletareappliedtoimagemosaicanddataintegration,soalargescaleobjectcanbemeasured.Experimentsshowthatthe
16、technologyhasmanyadvantages,suchassimpleequipment,fastspeedandlowcost,andthatthemeasuringerrorislessthan0.5mm/m.Keywords:Computervision;stereomatching;geometricinvariance;neuralnetwork;wavelettransform;clustering1IntroductionAtpresent,three-dimensional(3D)measuringmachineisstillamainrolein3Dmeasurem
17、ent.Butitcannotmeasureonlinebecauseofitsbulkanditscomplexconstruction,anditobtainsdatafrompointcontactsothatitcannotmeasuresoftobject.Therefore,itisimportantforindustrytoresearchnoncontactfastnondestructivemeasurementonline.Althoughtherehavebeenmanymethods,suchaslaserscanningmethod,structuredlightme
18、thod,phasemeasuringmethod,theycannotsimultaneouslysatisfythedemandsofmeasurementprecision,measurementspeed,automationandintellectualization,andlowcost.Consequently,inthispaper,usingtwo-camerawiththeconvergingoptical-axistograbimage.Withprocessingimage,stereomatchingimagemosaicanddataintegration,3Din
19、formationofobjectisobtainedfromapairofstereoimages.The3DmeasurementtechnologyhasbeenappliedtothemeasurementsystemoftheMulti-pointPress-formingMachine(MPFmachine)2,andgoodresultsareobtained.2MeasurementPrincipleandSystemDesignThispaperdescribesthe3Dnoncontactmeasurementtechnologybasedoncomputervision
20、,and3Dinformationofobjectisobtainedfromapairofstereoimages.Generally,therearetwoproblemsthatinfluenceobtaining3Dexactinformationofobject:theoneisestablishingtheexactmappingrelationbetweenimagepointandspecialpoint;theotherisstereomatchingproblem.Inthispaper,neuralnetworkisusedtoapproachingthemappingr
21、elationwithoutcameracalibration.Waveletedgedetection,searchingnon-supervisorclusteringandgeometricinvarianceareappliedtostereomatching.Themulti-scaleandmulti-resolutionattributeofwaveletisappliedtoimagemosaicanddataintegration.Inpractice,thetechnologyincludesmanymethodsandtechniques,itcanmeasurearbi
22、trarysizeandshapeobject.However,thesurfacesofsomeobjectsaresmooth.Matchingfeaturesareinconspicuous,sogratingisprojectedonobject.Andthedistortedstripesarecreatedonobject.Theyareregardedasmatchingfeatures.Forimprovingmeasurementprecision,two-camerawithconvergingoptical-axisischosen.Andthetwo-cameraand
23、thesmallself-madeprojectorconstituteaflexiblemeasuringhead.Asketchofthe3DmeasurementprinciplebasedonstereovisionisshowninFig.1.3EstablishmentoftheMappingRelationBetweenImagePointandSpecialPointActually,obtaining3Dinformationofobjectfromapairoftwoimagesisbymappingrelationbetweenimagepointandspecialpo
24、int,butuntilnownoapproachcancompletelydescribethenonlinearmappingrelationsincetherearemanycomplexnonlinearinfluencingfactorsincludingradialdistortionandlateraldistortionofcamera.However,neuralnetworkcansimulatehumanvisiontoestablishcomplexmappingbysimplenonlinearprocessingcells,sothispaperregardsthe
25、middleprocessfromimagepointtospecialpointasablackbox.AndBPnetworkwithamiddlelayerofsixneuralcellsisusedtosetupthemappingrelationbetweenimagepointandspecialpoint.PointAinleftimageandapointinrightimageareinputintotheBPnetwork,aspecialpointisoutput.Inotherwords,thestructureofBPnetworkis4-6-3.Usingneura
26、lnetwork,thechoosingoftrainingsamplesisimportantThetrainingsamplesnotonlylieinthemeasurablerange,butalsoshowmeasurementrangeofmeasurementsystem.Whiletwo-cameraisusedtograbobject,theobjectandthepartofobjectonlyinjointingviewingfieldcanbeabletobegrabbed.So3Dinformationofobjectfromapairofstereoimages,l
27、ensfocus,measurementprecision,oncemeasuringareaandthedistancebetweenobjectandbaselineoftwo-cameracontrol3Dmeasurementrangeofthesystemareobtained.Inthispaper,thestructureandfunctionofthetwocamerasthatareposedsymmetricallyareidentical,andtheimageareais,justasFig.2.Thelensfocusisf;thelinebetweentwoimag
28、ecentersisperpendicularto.Thecommonpartisregardedasjoiningviewingfieldoftwo-camera.Andthepartoutofisknownasblindarea.If2isviewingfieldangle,onthebasicofimagingrelation,theformulais(1)Aninscribedcircleisdoneinthejoiningviewingfield,ifisincludedangleoftwo-cameraopticalaxis,2isthedistancebetweentwoimag
29、ecenters,itsratiois(2)Inthisway,a2R2Rsampletemplatewith88gridsismade.Thesampletemplateisputworktable.Threepairsofstereoimagesaregrabbedrespectively,whilethesampletemplateismovedtothreedifferentheights(0,R,2R)alongtheverticaldirectiontosimulate3Dmeasurementrange.Thethreepairsofstereoimagesareregarded
30、astrainingsamples,andtheyareinputnetwork.4StereoPreciseMatchingatSubpixelLevelStereoprecisematchingismuchmoredifficultinstereovision,sotheapplyingofstereovisionisrestrictedinaway.Inthispaper,wavelettransformisappliedtodetectedgepoints,searchingnon-supervisorclusteringapproachisproposedtodistinguisht
31、hedifferentedgepointgroups.Theedgepointsinthesamepointgrouparefittedquadraticcurve,andthenstereoprecisematchingisachievedatsubpixellevelbasedongeometryinvariance.4.1StripeEdgesFittingBasedonSearchingNonsupervisorClusteringGenerally,imageoftencontainsrandomnoise,andwavelettransformcanrestrainnoiseand
32、detectedge,whiledifferentstructureimageedgesaredescribedbytheinformationspreadinginallresolutions.Sincetranslatinginvarianceisthemostimportantinstereomatchingbasedonedgefeature.QuadraticB-spineisselectedforamulti-scalegeneratingelementtodetectedgepointsofstripe.Actually,noiseisstillmixedinthesediscr
33、eteedgepoints,socurvefittingisusedtotranslatethediscreteedgepointsintoacontinuouscurve,andtoreducenoise.However,beforecurvesarefitted,itiscrucialthatalldiscreteedgepointsaredistinguishedintodifferentgroupsaccordingtothepracticalsituationofthestripeedgesinimages.Hammingdistancetoclusteringcenterisoft
34、enregardedasconstraintconditiontocluster,inotherwords,iftheattributevectorofapointisXl,andtheattributevectorofaclusteringcenteris,andif,nisthetotalnumberofgroups,thenXli,sotheideadoesntaccordwiththepracticalsituationofstripeedgepoints.Beforecurvesarefitted,notonlyisthenumberofgroupsgiven,butalsowhic
35、hgroupapointbelongstoisknown,andthenumberofgroupsisequaltothenumberofstripeedges.Therefore,asearchingnon-supervisorclusteringalgorithmisproposedinthispaper.IfDisanaggregateofpoints,nisnumberofpointsinD,andifDisdividedintogroups,dividingapproachisshownasfollows.1)Ifisattributevector,isknownasinitialg
36、roup,thatis,thenumberofgroupsisequalton;2)If=,end;3)Onthebasisofunderhood,thedistancebetweentwogroupsiscomputedforallgroups.If,and(Tstandsfortranspose),thatis=min,andtwonearestgroupsarechosen;4)andaremergedinto,thatis,sothetotalofgroupsdecrease1;5)Return(2).4.2SearchingCorrespondingPointsBasedonGeom
37、etricInvarianceGeometricinvarianceisdefinedthatgeometricalfigureandvectorkeepinvarianceinmathematicalmanipulation.Foraspecialpolygon,twodifferentshapepolygonswillbeobtainedintwoimageplanesbyperspectivetransform.Inthesameway,fora3Dcurve,twodifferent2Dcurvesareobtainedintwoimageplanes.Thereforegeometr
38、icinvarianceisappliedtomatchingstraightlinesandcurves.Forstraight-linematching,representationalgeometricinvarianceiscomposedoffivepointsinthesamestraightlineorfivestraightlinesinthesamesurface.Weassumethatisarbitraryfivestraightlinesonspecialplane,straight-lineequationis(3)Wearbitrarilychoosethreest
39、raightlines,andinthefivestraightlines(k1,k2,k3=1,2,3,4,5,k1k2,k1,k3,k2k3).Thesystemofequationsofthethreestraightlinesaregivenby(4)Andthesestraightlinesaretranslatedintoimagestraightlinesbyperspectivetransform.Theimagestraightlineshavealsocorrespondingstraight-lineequationparameters.Andtheparametersa
40、reshownwithsuperscript(forexample).Itistestified,thoughtheshapesoffivestraightlinescanmorechange,therearegeometricinvariants,ifMisdetA,theyare,(5)Analogously,therearesomegeometricinvariantsforagroupofquadraticcurves.Ifisaquadraticcurveonthespecialplane,itsequationcanbeshownasfollows(6)Andifisparamet
41、ermatrixofquadraticcurve,itisalsoshownbymatrixasfollows(7)Iftherearetwoquadraticcurvesand,theirparametermatrixesarerespectivelyand.Theyaretranslatedintoandbygeometricprojection,andtheirparametermatrixesareand.Itistestified,ifistrackofmatrix,therearetwogeometricinvariantswhethermodeofgeometricproject
42、ionischanged.(8)(9)Inthisway,straightlinesandcurvesarematchedeffectively.Inthispaper,gratingisprojectedonobjectinverticaldirectionandlateraldirectionrespectively,whiletwocamerasgrabimages.Withwaveletedgedetection,searchingnon-supervisor,edgepointsarefittedintoquadraticcurves.Withgeometricinvariance,
43、quadraticcurvesarematched,andcrosspointsofverticalcurvesandlateralcurvesarecomputed.Sostereoprecisematchingatsubpixellevelisachieved.5ImageMosaicBasedonWaveletWhenlarge-scalesurfaceismeasured,manypairsofstereoimagesaregrabbedfromdifferentviewpointsorwithmovingandrotatingobject.Andtwoadjacentimagesne
44、edmosaic.Theimportantquestionofimagemosaicisimageregistration,thatistosay,overlappedpartsbetweentwoadjacentimagesareputinorder,andcorrespondingmatchingbetweentwoadjacentimagesisalsoinvolvedinimagemosaic.Correspondingmatchinginregistrationisdeferredfromcorrespondingmatchinginstereovision.Intheformer,
45、twoimagesaregrabbedfromthesameviewpointorfromthedifferentviewpointsthataretranslatedalongthebasicline,andinthelatter,theviewpointsoftwoimagesarenotonlytranslatedbutalsorevolved.Inthispaper,someblackpointsarepastedatrandomonobjectinordertomosaiceasily,andtheblackpointsareregardedasregistrationfeature
46、points.Meanwhile,weusebiorthogonalwaveletwithlinearityandsymmetrytodecomposetwoimagesthatarewillbemosaic,sotheimagescanbematchedandregisteredfromcoarsetofineonmulti-scale,andlastlyabigimageisbecome.Infact,wavelettransformisaband-passfilter,waveletvectorondifferentscalesshowsthestatedwidthoffrequency
47、band,andsofrequencybandwidthofeachwaveletvectorisunequal.TwoimagesaredecomposedintowaveletvectorsondifferentfrequencybandsbasedonMallatalgorithm,andthenthedifferentmosaicwidthsareselectedondifferentscalestomatchandregister,soabigmosaicimageissmoothandfine.6ExperimentsandResultsAnalysisInthispaper,th
48、etechnologyisappliedtothemeasurementsystemofMPFmachine.Aftertheshapedsurfaceismeasured,themeasuringresultsarereturnedtoCAD/CAEsystem,andclosed-loopcontrolisachieved.Accordingtomeasurementcondition,measurementprecisionandonceshapedarea,weselecttwocameras(MTV1881CB),twocameralensesandaimage-recorddevice(METEOR).Thedistancebetweentwocamerasis300mm;thedistancebetweenthesurfaceandthetwocamerasis500mm.A150150mmcurvesurfaceismeasuredwiththetechnology,measurementresultsareshowninTab.1,andmeasurementstepsaregivenasf