《基于计算机视觉的三维测量技术(文献翻译).pdf》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).pdf（21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、精选 doc 最新版重 庆 理 工 大 学文献翻译二级学院专业班级学生姓名学号.精选 doc 最新版译文：基于计算机视觉的三维测量技术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 0.5 mm/m 以内。关键词:计算机视觉，立体匹配，几何不变性，神
2、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觉的三维非接触测量技术，三维对象的信息是从一对立
3、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。.精选 doc 最新版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灵活的测量头。一个基于立体
4、视觉的三维测量的原理草图如图 1 所示。3 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(xl,yl)和一个右边的点(xr,yr)纳入 BP 网络，一个特殊的点(x,y,z)被输出。换言之，这个 BP 网络的
5、结构是 4-6-3。利用神经网络，样本的选择是很重要的。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.精选 doc 最新版像区域的是2Tx2Ty，如图2 所示。该镜头的焦点是f;两个图像之间o1o2的中心垂直线是AC。共同的部分ABCD被视为双摄像头的连接视野。而超出ABCD的部视为盲区。假设视野角度为 2，基本的成像关系公式为：ar
6、ctg(Tx)f(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 2M，其比例为：R (M csc f)sin(M csc f)TxT2x f2(2)这样，一个 2R2R 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.精选 doc 最新版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
7、。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
8、是Xl，一个聚类中心的属性向量是Ci，如果minXlCi，n 是聚类总数，Xli，这样的思想i1n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 D 点的数量，如果 D 分成M组，划分方法如下所示。1)如果Xi是属性向量，iNn1,2,3,n被称为初始群体，这里是Xi，M的组数等于 n；2）假如M=M，结束；3）在覆盖下的基础上，两个群体之间的距离dmin(有群体。假如X 矩阵），dmin(4）i,)也就可以计算所kl，X k TX X(X X)(X X)（T 代表转
9、置，且i=min，最近的两组(i,l)被选择；X X,)kl和j是合并到，于是iiji，所以群体总数减少；.精选 doc 最新版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 条直线在同一平面所代表。我们假设li(i 1,2,3,4,5)是特殊平面上的任意 5 条直线，直线方程为：fi(x,y)aixbiy ci 0(3)我
10、们任意选择 3 直线lk1，lk2和lk3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：Ak1k2k3X 0(4)这些直线均按直线(l,landl)k1k2k3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ak1）。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：M134M125I1M124M135(5),M124M235I2M234M125类似地，有一个组的二次曲线的一些几何不变量。如果C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
11、次曲线：Q(x,y)Ax2 Bxy Cy2 Dx Ey f 0(6)如果P是二次曲线的参数矩阵，它也表现为如下矩阵：TXLPXL 0(7).精选 doc 最新版如果有两条二次曲线C1和C2，它们的参数矩阵分别为P1和P2。运用几何投影将它们转化为C3和C4，其参数矩阵为P3和P4。它证明，如果tr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。1I1 tr(P31P4)tr(P1P2)(8)I1 tr(P41P3)tr(P21P1)(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
12、。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
13、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.精选 doc 最新版术后，测量结果返回到 CAD/CAECAD/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
14、量精度一旦成熟，我们选择两个摄像头（MTV1881CBMTV1881CB），两个镜头和一个图像记录装置（METEORMETEOR）。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。A 150150 mm2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
15、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。.精选 doc 最新版图、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 0.5 mm/m。它还提供了一个适用于工业计算机视觉的新思路。实验结果表明，三维测量技术是非常完美的。.精选 doc 最新版原文：3D Measurement Technology Based3D Measurem
16、ent Technology Basedon Computer Visionon Computer VisionAbstract:Abstract:On the basis of computer vision,a noncontact 3D measurementtechnology was proposed in this paper.Using neural network,the mappingrelation between image point and special point was established.Theprojection of grating on object
17、 is regarded as matching features,withwaveletedgedetection,searchingnon-supervisorclusteringandgeometric invariance.Stereo precision matching is achieved at subpixellevel.Furthermore,the multi-scale and multi-resolution attributes ofwavelet are applied to image mosaic and data integration,so a large
18、 scaleobject can be measured.Experiments show that the technology has manyadvantages,such as simple equipment,fast speed and low cost,and thatthe measuring error is less than 0.5 mm/m.KeyKey words:words:Computer vision;stereo matching;geometric invariance;neuralnetwork;wavelet transform;clustering1
19、Introduction1 IntroductionAt present,three-dimensional(3D)measuring machine is still a mainrole in 3D measurement.But it cannot measure on line because of its bulkand its complex construction,and it obtains data from point contact sothat it cannot measure soft object.Therefore,it is important for.精选
20、 doc 最新版industry to research noncontact fast nondestructive measurement on line.Although there have been many methods,such as laserscanning method,structuredlightmethod,phasemeasuringmethod1,theycannotsimultaneously satisfy the demands of measurement precision,measurementspeed,automation and intelle
21、ctualization,and low cost.Consequently,in this paper,using two-camera with the convergingoptical-axis to grab image.With processing image,stereo matching imagemosaic and data integration,3D information of object is obtained froma pair of stereo images.The 3D measurement technology has been appliedto
22、 the measurement system of the Multi-point Press-forming Machine(MPFmachine)2,and good results are obtained.2 Measurement Principle and System Design2 Measurement Principle and System DesignThis paper describes the 3D noncontact measurement technology basedon computer vision,and3D information of obj
23、ect isobtained from apairof stereo images.Generally,there are two problems that influenceobtaining 3D exact information of object:the one is establishing theexact mapping relation between image point and special point;the otheris stereo matching problem.In this paper,neural network is used toapproac
24、hing the mapping relation without camera calibration.Wavelet edgedetection,searching non-supervisor clustering and geometric invarianceare applied to stereo matching.The multi-scale and multi-resolutionattribute of wavelet is applied to image mosaic and data integration.Inpractice,the technology inc
25、ludes many methods and techniques,it canmeasure arbitrary size and shape object.However,the surfaces of some objects are smooth.Matching featuresare inconspicuous,so grating is projected on object.And the distortedstripes are created on object.They are regarded as matching features.Forimprovingmeasu
26、rementprecision,two-camerawithconvergingoptical-axis is chosen.And the two-camera and the small self-made.精选 doc 最新版projector constitute a flexible measuring head.A sketch of the 3Dmeasurement principle based on stereo vision is shown in Fig.1.3 Establishment3 Establishment ofof thethe MappingMappin
27、g RelationRelation BetweenBetween ImageImage PointPoint andand SpecialSpecialPointPoint Actually,obtaining 3D information of object from a pair of twoimages is by mapping relation between image point and special point,butuntil now no approach can completely describe the nonlinear mappingrelation sin
28、ce there are many complex nonlinear influencing factorsincluding radial distortion and lateral distortion of camera.However,neural network can simulate human vision to establish complex mapping bysimple nonlinear processing cells,sothis paper regardsthe middleprocess from image point to special poin
29、t as a black box.And BP networkwith a middle layer of six neural cells is used to set up the mappingrelation between image point and special point.Point A(xl,yl)in leftimage and a point(xr,yr)in right image are input into the BP network,a special point(x,y,z)is output.In other words,the structure of
30、 BP.精选 doc 最新版network is 4-6-3.Using neural network,the choosing of training samples is importantThe training samples not only lie in the measurable range,but alsoshowmeasurement range of measurement system.While two-camera is used to grab object,the object and the part ofobject only in jointing vie
31、wing field can be able to be grabbed.So 3Dinformationofobjectfromapairofstereoimages,lensfocus,measurement precision,once measuring area and the distance betweenobject and baseline of two-camera control 3D measurement range of thesystem are obtained.In this paper,the structure and function of the tw
32、o cameras that areposed symmetrically are identical,and the image area is2Tx2Ty,justas Fig.2.The lens focus is f;the lineo1o2 between two image centersis perpendicular toAC.The common partABCD is regarded as joiningviewing field of two-camera.And the part out ofABCD is known as blindarea.If 2 is vie
33、wing field angle,on the basic of imaging relation,the formula is arctg(Tx)(1)f.精选 doc 最新版An inscribed circle is done in the joining viewing field,if isincluded angle of two-camera optical axis,2M is the distance betweentwo image centers,its ratio isR (M csc f)sin(2)(M csc f)TxT f2x2In this way,a 2R2
34、R sample template with 88 grids is made.Thesample template is put worktable.Three pairs of stereo images are grabbedrespectively,while the sample template is moved to three differentheights(0,R,2R)along the vertical direction to simulate 3D measurementrange.The three pairs of stereo images are regar
35、ded as training samples,and they are input network.4 Stereo Precise Matching at Subpixel Level4 Stereo Precise Matching at Subpixel Level Stereo precise matching is much more difficult in stereo vision,sothe applying of stereo vision is restricted in a way.In this paper,wavelettransformisappliedtode
36、tectedgepoints,searchingnon-supervisor clustering approach is proposed to distinguish the.精选 doc 最新版different edge point groups.The edge points in the same point group arefitted quadratic curve,and then stereo precise matching is achieved atsubpixel level based on geometry invariance.4.1 Stripe Edge
37、s Fitting Based on Searching Nonsupervisor Clustering Generally,image often contains random noise,and wavelet transformcan restrain noise and detect edge,while different structure image edgesare described by the information spreading in all resolutions.Sincetranslating invariance is the most importa
38、nt in stereo matching based onedge feature.Quadratic B-spine3,4 is selected for a multi-scalegenerating element to detect edge points of stripe.Actually,noise is still mixed in these discrete edge points,socurve fitting is used to translate the discrete edge points into acontinuous curve,and to redu
39、ce noise.However,before curves are fitted,it is crucial that all discrete edge points are distinguished intodifferent groups according to the practical situation of the stripe edgesin images.Hamming distance to clustering center is often regarded asconstraint condition to cluster,in other words,if t
40、he attribute vectorof a point is Xl,andthe attribute vector of a clustering center isCi,and ifmini1nXCli,n is the total number of groups,then Xli,sothe idea doesnt accord with the practical situation of stripe edge points.Before curves are fitted,not only is the number of groups given,but alsowhich
41、group a point belongs to is known,and the number of groups is equalto the number of stripe edges.Therefore,a searching non-supervisorclustering algorithm is proposed in this paper.If D is an aggregate of points,n is number of points in D,and ifD is divided intoM groups,dividing approach is shown as
42、follows.1)IfXi is attribute vector,iNn1,2,3,n is known as initial.精选 doc 最新版group,that isXi,the numberM of groups is equal to n;2)IfM=M,end;3)On the basis of under hood,the distancedmin(groupsiscomputedforallgroups.IfX,)between twoklk,Xi,andX X(X X)T(X X)(Tstandsfortranspose),thatis=min,and two near
43、est groups(i,l)are chosen;X X(,)dminkl4)andij are merged into,that is iiji,so thetotal of groups decrease 1;5)Return(2).4.2 Searching Corresponding Points Based on Geometric Invariance4.2 Searching Corresponding Points Based on Geometric InvarianceGeometric invariance is defined that geometrical fig
44、ure and vectorkeep invariance in mathematical manipulation5.For a special polygon,two different shape polygons will be obtainedin two image planes by perspective transform.In the same way,for a 3Dcurve,two different 2D curves are obtained in two image planes.Thereforegeometric invariance is applied
45、to matching straight lines and curves.For straight-line matching,representational geometric invarianceis composed of five points in the same straight line or five straight linesin the same surface.We assume thatli(i 1,2,3,4,5)is arbitrary five straight lines onspecial plane,straight-line equation is
46、fi(x,y)aixbiy ci 0(3)(3)We arbitrarily choose three straight lineslk,lkandlk3 in the12five straight lines(k1,k2,k3=1,2,3,4,5,k1k2,k1,k3,k2k3).Thesystem of equations of the three straight lines are given byAk1k2k3X 0.精选 doc 最新版(4)And these straight lines are translated into image straight lines(l,lan
47、dl)by perspective transform.The image straight lines havek1k2k3also corresponding straight-line equation parameters.And the parametersare shown with superscript(for exampleak1).It is testified,though theshapes of five straight lines can more change,there are geometricinvariants,if Mis det A,they are
48、M134M125I1M124M135(5),M124M235I2M234M125Analogously,there are some geometric invariants for a group ofquadratic curves.IfC is a quadratic curve on the special plane,itsequation can be shown as followsQ(x,y)Ax2 Bxy Cy2 Dx Ey f 0 (6)And ifP is parameter matrix of quadratic curve,it is also shownby mat
49、rix as followsTPXL 0XL(7)If there are two quadratic curvesC1 andC2,their parametermatrixes are respectivelyP1 andP2.They are translated intoC3 andC4 by geometric projection,and their parameter matrixes areP3 andP4.It is testified,iftr is track of matrix,there are two geometricinvariants whether mode
50、 of geometric projection is changed.I111tr(P3P4)tr(P1P2)(8)11I2tr(P P)tr(P432P1)(9)In this way,straight lines and curves are matched effectively.精选 doc 最新版In this paper,grating is projected on object in vertical directionand lateral direction respectively,while two cameras grab images.Withwavelet ed