外文翻譯--基于計算機視覺的三維測量技術(shù)

1、<p>　　文 獻(xiàn) 翻 譯</p><p>　　二級學(xué)院 </p><p>　　專 業(yè) </p><p>　　班 級 </p><p&g

2、t;　　學(xué)生姓名 學(xué) 號 </p><p><b>　　譯文：</b></p><p>　　基于計算機視覺的三維測量技術(shù)</p><p>　　摘 要:本文根據(jù)計算機視覺原理，提出一種三維非接觸測量技術(shù)。該技術(shù)根據(jù)人眼感知事物的原理，利用神經(jīng)網(wǎng)絡(luò)擬合圖像坐標(biāo)與空間坐標(biāo)的映射關(guān)系；以光柵投影曲線為

3、特征，采用小波邊緣檢測和搜索式無監(jiān)督聚類，結(jié)合視覺幾何不變性，實現(xiàn)亞像素級的立體精確匹配；并采用小波多尺度多分辨率的特性，拼接圖像，融合數(shù)據(jù)，對物體進(jìn)行全方位測量。實驗表明，該技術(shù)設(shè)備簡單，測量速度快，測量精度控制在0.5 mm/m以內(nèi)。</p><p>　　關(guān)鍵詞:計算機視覺，立體匹配，幾何不變性，神經(jīng)網(wǎng)絡(luò)，小波變換，聚類</p><p><b>　　引言</b>&

4、lt;/p><p>　　目前，三維測量仍以三維坐標(biāo)測量機為主。但是它由于體積大、結(jié)構(gòu)復(fù)雜而不能在線測量，是接觸測量而不能測量柔軟的物體。因此，研究快速無損、非接觸在線測量在工業(yè)上十分重要。盡管現(xiàn)在有很多方法，如激光掃描法、結(jié)構(gòu)光法、相位測量法，但是都不能同時滿足測量精度、效率、成本、自動化和智能化等方面的要求。</p><p>　　因此，在本文使用雙攝像機融合光學(xué)軸抓拍物體。隨著處理圖像，立體

5、匹配圖像和數(shù)據(jù)集成，三維物體的信息就是從這個立體圖像中獲得。三維測量技術(shù)已應(yīng)用于測量系統(tǒng)中的多點壓成型機的測量，并取得了良好的效果。</p><p><b>　　測量原理及系統(tǒng)設(shè)計</b></p><p>　　本文介紹了基于計算機視覺的三維非接觸測量技術(shù)，三維對象的信息是從一對立體圖像中獲取。一般來說，有兩個問題影響的三維物體獲得確切的消息：一種是圖像之間建立特殊點點

6、和準(zhǔn)確的映射關(guān)系，另一種是立體匹配問題。本文神經(jīng)網(wǎng)絡(luò)是用來映射關(guān)系接近的情況下攝像機標(biāo)定。小波邊緣檢測，尋找非監(jiān)督聚類和幾何不變性適用于立體匹配。在多尺度，多分辨率的小波屬性應(yīng)用于圖像拼接和數(shù)據(jù)集成。在實踐中，這項技術(shù)包含了許多方法和技術(shù)，它可以測量任意大小和形狀的對象。</p><p>　　然而，有一些物體的表面很光滑。匹配功能不明顯，因此用光柵對象預(yù)測。而扭曲的條紋上創(chuàng)建的對象被視為匹配功能。為了提高測量精度

7、，用兩個與融合光學(xué)軸相機，這兩個相機和一小型自制的投影機就構(gòu)成了一種靈活的測量頭。一個基于立體視覺的三維測量的原理草圖如圖1所示。</p><p>　　3 建立圖像點和特殊點之間的映射關(guān)系</p><p>　　實際上，獲得從兩個圖像對三維物體的信息是獲取圖像點之間的映射和特殊點的關(guān)系，但是到現(xiàn)在為止沒有任何方法可以完全描述非線性映射關(guān)系，因為有許多復(fù)雜的非線性影響因素，包括攝像的徑向變形

8、和橫向變形。但是，神經(jīng)網(wǎng)絡(luò)可以模擬人類的視覺，建立了簡單的非線性映射來處理復(fù)雜的單元，因此本文就從圖像點的過程中當(dāng)作黑箱特殊點。和BP網(wǎng)絡(luò)的6個神經(jīng)細(xì)胞中間層網(wǎng)絡(luò)來設(shè)置點之間的形象和特殊點的映射關(guān)系。圖片左邊的點A和一個右邊的點納入BP網(wǎng)絡(luò)，一個特殊的點被輸出。換言之，這個BP網(wǎng)絡(luò)的結(jié)構(gòu)是4-6-3。</p><p>　　利用神經(jīng)網(wǎng)絡(luò)，樣本的選擇是很重要的。樣本不僅在于衡量的范圍，也顯示測量系統(tǒng)的測量范圍。<

9、;/p><p>　　雖然兩個相機是用來抓拍對象，但是這部分對象只有在焊接處的視野內(nèi)才能被獲取。因此，物體三維信息的立體圖像，鏡頭焦點的測量精度，測量范圍和目標(biāo)與攝像機之間的兩個基準(zhǔn)距離控制三維測量系統(tǒng)的測量范圍。</p><p>　　本文的結(jié)構(gòu)和功能和兩個相機是用來抓拍對象構(gòu)成對稱是相同的，相機的圖像區(qū)域的是，如圖2所示。該鏡頭的焦點是;兩個圖像之間的中心垂直線是。共同的部分被視為雙攝像頭的連

10、接視野。而超出的部視為盲區(qū)。假設(shè)視野角度為2，基本的成像關(guān)系公式為：</p><p><b>　　(1)</b></p><p>　　這個內(nèi)切圓是視野范圍，如果兩個相機光軸的夾角是β，兩個圖像中心之間的距離是2，其比例為：</p><p><b>　　(2)</b></p><p>　　這樣，一個2

11、R×2R的示例模板由8×8的格子組成。這個示例模板固定在工作臺上。分別獲取三對立體圖像，而示例模板沿垂直線方向移動到三個不同高度（0，R，2R）模擬三維測量范圍。三對立體圖像被視為訓(xùn)練樣本，把它們輸入網(wǎng)絡(luò)。</p><p>　　亞像素級的立體精確匹配</p><p>　　對立體顯示來說立體精確匹配要困難得多，所以申請采用立體顯示在某種程度上受到限制。本文應(yīng)用小波變換檢測

12、邊緣點，尋找非主管聚類方法，提出以區(qū)分不同的邊緣點群。在同一個點群的邊緣點的二次曲線擬合，然后在立體精確匹配亞像素級的水平基礎(chǔ)上取得幾何不變性。</p><p>　　4．1 條紋邊緣擬合中的非聚類搜索</p><p>　　一般來說，圖像往往含有隨機噪聲，小波變換能抑制噪聲和檢測移動，同時不同結(jié)構(gòu)圖像邊緣的信息傳播在所有決議中。自從轉(zhuǎn)化不變性是最重要的立體匹配的邊緣特征。二次B-spine

13、被用來處理一個多尺度的生成元素檢測條紋邊緣點。</p><p>　　實際上，噪音仍然混合在這些離散邊緣點中，因此，曲線擬合用于轉(zhuǎn)化為連續(xù)曲線離散邊緣點，并減少噪音。然而，在曲線擬合之前，至關(guān)重要的是，所有的離散邊緣點根據(jù)圖像中條紋邊緣的實際情況分成不同的群。海明距離的聚類中心往往被視為約束條件群，換句話說，假設(shè)一個點的屬性向量是，一個聚類中心的屬性向量是，如果，n是聚類總數(shù)，，這樣的思想不符合的條紋邊緣點的實際情

14、況。在曲線擬合之前，不僅給定的群體，而且這組點屬于已知，而群體數(shù)目與條紋邊數(shù)相等。因此，在本文中提出了非主管聚類算法。</p><p>　　如果D是一個集合點，n是D點的數(shù)量，如果D分成組，劃分方法如下所示。</p><p>　　1) 如果是屬性向量，被稱為初始群體，這里是，的組數(shù)等于n；</p><p><b>　　2）假如=，結(jié)束；</b>

15、</p><p>　　3）在覆蓋下的基礎(chǔ)上，兩個群體之間的距離也就可以計算所有群體。假如，，且（T代表轉(zhuǎn)置矩陣）， = min{}，最近的兩組被選擇；</p><p>　　4）和是合并到，于是，所以群體總數(shù)減少；</p><p>　　5）重復(fù)步驟（2）。</p><p>　　4．2 基于幾何不變性的相應(yīng)點搜索</p><

16、p>　　幾何不變性的定義是幾何圖案和矢量保持精確不變。</p><p>　　對于一個特殊的多邊形，兩種不同的成行將得到兩種透視變換圖像位面。以同樣的方式，對于一個三維曲線，兩種不同的二維曲線得到兩個圖像位面。因此，幾何不變性應(yīng)用于匹配直線和曲線。</p><p>　　對于直線匹配，幾何不變性由5個點在同一條直線或5條直線在同一平面所代表。</p><p>　　

17、我們假設(shè)是特殊平面上的任意5條直線，直線方程為：</p><p><b>　　(3)</b></p><p>　　我們?nèi)我膺x擇3直線，和在5條直線上(k1,k2,k3=1,2,3,4,5,k1≠k2,k1,≠k3,k2≠k3)。這三條直線方程給出為：</p><p><b>　　(4)</b></p><

18、;p>　　這些直線均按直線的角度轉(zhuǎn)變成圖像。直線的特征也轉(zhuǎn)換相應(yīng)的直線方程的參數(shù)。參數(shù)顯示在上標(biāo)處（例如）。它證明，盡管這連續(xù)的五條直線的形狀可以有更多的變化，它們也服從幾何不變性，如果M′屬于A，它們是：</p><p>　　, (5)</p><p>　　類似地，有一個組的二次曲線的一些幾何不變量。如果這個特殊平面上的一條二次曲線

19、，它的方程可以表現(xiàn)為如下的二次曲線：</p><p><b>　　(6)</b></p><p>　　如果是二次曲線的參數(shù)矩陣，它也表現(xiàn)為如下矩陣：</p><p><b>　　(7)</b></p><p>　　如果有兩條二次曲線和，它們的參數(shù)矩陣分別為和。運用幾何投影將它們轉(zhuǎn)化為和，其參數(shù)矩陣為

20、和。它證明，如果是矩陣的軌道，有兩個幾何不變量不管幾何投影模式是否變化。</p><p><b>　　(8)</b></p><p><b>　　(9)</b></p><p>　　這樣，直線和曲線就有效匹配了。</p><p>　　本文光柵投影在垂直方向和水平方向被分別提出來，而兩相機抓拍圖像。隨

21、著小波邊緣檢測，搜索式無監(jiān)督聚類，邊緣點到二次曲線擬合。幾何不變性，二次曲線匹配，垂直曲線和橫向曲線交叉點的計算。因此，亞像素級的立體精確匹配得以實現(xiàn)。</p><p><b>　　基于小波的圖像拼接</b></p><p>　　當(dāng)大規(guī)模的測量表面時，許多對立體圖象在不同的觀點或者移動和旋轉(zhuǎn)中被抓拍到。兩個相鄰圖像需要鑲嵌。圖像鑲嵌的重要問題是圖像配準(zhǔn)，也就是說，兩個

22、相鄰圖像之間的重疊部分，以便付諸表決，并且兩個相鄰圖像之間的相應(yīng)匹配也是圖像鑲嵌的復(fù)雜工作。通訊匹配在相應(yīng)的立體視覺匹配之后。在這之前，從相同的角度或者不同的角度沿著基本路線轉(zhuǎn)換來抓住兩個圖像，并在這之后，這兩張圖片的角度不僅要是轉(zhuǎn)換，而且要旋轉(zhuǎn)。</p><p>　　本文，一些隨機黑點能容易的鑲嵌，這些黑點被認(rèn)為是重要的拼接點。同時，我們用線性和對稱雙正交分解兩個圖像來鑲嵌，使粗糙的圖像可以得到很好的匹配和拼接

23、，最終得到一個大的圖像。</p><p>　　事實上，小波變換是一種帶通濾波，小波向量的顯示用不同尺度的頻帶寬度來衡量，所以每個小波的頻率帶寬是不相等的。兩個圖像用Mallat算法分解成不同頻率波段的小波向量，然后不同規(guī)模選擇不同的鑲嵌寬度來滿足和拼接，于是一個大的鑲嵌圖便順利且很好的完成了。</p><p>　　6 實驗及結(jié)果分析</p><p>　　在本次設(shè)計

24、中，這項技術(shù)在MPF機的測量系統(tǒng)中得到了應(yīng)用。在應(yīng)用了該技術(shù)后，測量結(jié)果返回到CAD / CAE系統(tǒng)中顯示閉環(huán)控制得到了實現(xiàn)。</p><p>　　表面形狀后測量，測量結(jié)果返回到CAD / CAE系統(tǒng)和閉環(huán)控制的實現(xiàn)。據(jù)測量條件、測量精度一旦成熟，我們選擇兩個攝像頭（MTV1881CB），兩個鏡頭和一個圖像記錄裝置（METEOR）。這兩個攝像頭之間的距離為300毫米；物體表面和兩部相機之間的距離為500毫米。A

25、150×150 mm的曲面是該工藝的標(biāo)準(zhǔn)測量范圍，測量結(jié)果在標(biāo)簽 1上顯示，測量步驟如下：</p><p>　　1） 建立與圖像點和特殊點之間的映射關(guān)系；</p><p>　　2）三維表面在工作臺上進(jìn)行，首先，二個攝像機在沒有干擾和光線的情況下同時抓拍一對立體圖像。其次，在抓住兩對立體圖像對，一對在光柵的垂直方向上抓拍，另一對在光柵的橫向上抓拍；</p><p

26、>　　3）進(jìn)程映像，消除背景，減少噪音，如圖3a，3b所示；</p><p>　　4）功能檢測，如圖3c；</p><p>　　5）搜索對應(yīng)點，并鑲嵌圖像；</p><p>　　6）計算三維坐標(biāo)，重建三維表面，如圖3d。</p><p>　　實驗表明，測量誤差小于0.5mm，測量時間約2秒，包括圖像抓拍、圖像處理、建立圖像點和特殊點的映

27、射關(guān)系、搜索相應(yīng)的坐標(biāo)點和調(diào)整計算。</p><p>　　圖、3 圖像處理</p><p><b>　　7 結(jié)束語</b></p><p>　　在本文中，提出了一種新的基于計算機視覺的三維測量技術(shù)，該技術(shù)設(shè)備簡單、測量速度快、成本低?？梢詼y量大型對象，測量精度低于0.5 mm/m。它還提供了一個適用于工業(yè)計算機視覺的新思路。實驗結(jié)果表明，

28、三維測量技術(shù)是非常完美的。</p><p><b>　　原文：</b></p><p>　　3D Measurement Technology Based</p><p>　　on Computer Vision </p><p>　　Abstract: On the basis of computer visi

29、on, a noncontact 3D measurement technology was proposed in this paper. Using neural network, the mapping relation between image point and special point was established. The projection of grating on object is regarded as

30、matching features, with wavelet edge detection, searching non-supervisor clustering and geometric invariance. Stereo precision matching is achieved at subpixel level. Furthermore, the multi-scale and multi-resolution att

31、ributes of wavelet ar</p><p>　　Key words: Computer vision; stereo matching; geometric invariance; neural network; wavelet transform; clustering</p><p>　　1 Introduction</p><p>　　At

32、present, three-dimensional(3D) measuring machine is still a main role in 3D measurement. But it cannot measure on line because of its bulk and its complex construction, and it obtains data from point contact so that it c

33、annot measure soft object. Therefore, it is important for industry to research noncontact fast nondestructive measurement on line. Although there have been many methods, such as laserscanning method, structured light met

34、hod, phase measuring method, they cannot simultaneously s</p><p>　　Consequently, in this paper, using two-camera with the converging optical-axis to grab image. With processing image, stereo matching image m

35、osaic and data integration, 3D information of object is obtained from a pair of stereo images. The 3D measurement technology has been applied to the measurement system of the Multi-point Press-forming Machine (MPF machin

36、e)[2], and good results are obtained.</p><p>　　2 Measurement Principle and System Design</p><p>　　This paper describes the 3D noncontact measurement technology based on computer vision, and 3D

37、information of object is obtained from a pair of stereo images. Generally, there are two problems that influence obtaining 3D exact information of object: the one is establishing the exact mapping relation between image

38、point and special point; the other is stereo matching problem. In this paper, neural network is used to approaching the mapping relation without camera calibration. Wavelet edge detecti</p><p>　　However, the

39、 surfaces of some objects are smooth. Matching features are inconspicuous, so grating is projected on object. And the distorted stripes are created on object. They are regarded as matching features. For improving measure

40、ment precision, two-camera with converging optical-axis is chosen. And the two-camera and the small self-made projector constitute a flexible measuring head. A sketch of the 3D measurement principle based on stereo visio

41、n is shown in Fig.1.</p><p>　　3 Establishment of the Mapping Relation Between Image Point and Special Point</p><p>　　Actually, obtaining 3D information of object from a pair of two images is by

42、 mapping relation between image point and special point, but until now no approach can completely describe the nonlinear mapping relation since there are many complex nonlinear influencing factors including radial distor

43、tion and lateral distortion of camera. However, neural network can simulate human vision to establish complex mapping by simple nonlinear processing cells, so this paper regards the middle process from im</p><

44、p>　　Using neural network, the choosing of training samples is important The training samples not only lie in the measurable range, but also show measurement range of measurement system.</p><p>　　While two

45、-camera is used to grab object, the object and the part of object only in jointing viewing field can be able to be grabbed. So 3D information of object from a pair of stereo images, lens focus, measurement precision, onc

46、e measuring area and the distance between object and baseline of two-camera control 3D measurement range of the system are obtained.</p><p>　　In this paper, the structure and function of the two cameras that

47、 are posed symmetrically are identical, and the image area is , just as Fig.2. The lens focus is f; the line between two image centers is perpendicular to . The common part is regarded as joining viewing field of two-c

48、amera. And the part out of is known as blind area. If 2 is viewing field angle, on the basic of imaging relation, the formula is</p><p><b>　　(1)</b></p><p>　　An inscribed circle is

49、done in the joining viewing field, if β is included angle of two-camera optical axis, 2 is the distance between two image centers, its ratio is</p><p><b>　　(2)</b></p><p>　　In this w

50、ay, a 2R×2R sample template with 8×8 grids is made. The sample template is put worktable. Three pairs of stereo images are grabbed respectively, while the sample template is moved to three different heights (0,

51、 R, 2R) along the vertical direction to simulate 3D measurement range. The three pairs of stereo images are regarded as training samples, and they are input network.</p><p>　　4 Stereo Precise Matching at Su

52、bpixel Level</p><p>　　Stereo precise matching is much more difficult in stereo vision, so the applying of stereo vision is restricted in a way. In this paper, wavelet transform is applied to detect edge poin

53、ts, searching non-supervisor clustering approach is proposed to distinguish the different edge point groups. The edge points in the same point group are fitted quadratic curve, and then stereo precise matching is achieve

54、d at subpixel level based on geometry invariance.</p><p>　　4.1 Stripe Edges Fitting Based on Searching Nonsupervisor Clustering</p><p>　　Generally, image often contains random noise, and wavelet

55、 transform can restrain noise and detect edge, while different structure image edges are described by the information spreading in all resolutions. Since translating invariance is the most important in stereo matching ba

56、sed on edge feature. Quadratic B-spine is selected for a multi-scale generating element to detect edge points of stripe.</p><p>　　Actually, noise is still mixed in these discrete edge points, so curve fittin

57、g is used to translate the discrete edge points into a continuous curve, and to reduce noise. However, before curves are fitted, it is crucial that all discrete edge points are distinguished into different groups accordi

58、ng to the practical situation of the stripe edges in images. Hamming distance to clustering center is often regarded as constraint condition to cluster, in other words, if the attribute vector of a point</p><p

59、>　　If D is an aggregate of points, n is number of points in D, and if D is divided into groups, dividing approach is shown as follows.</p><p>　　1) If is attribute vector, is known as initial group , th

60、at is ,the number of groups is equal to n; </p><p>　　2) If =, end;</p><p>　　3) On the basis of under hood, the distance between two groups is computed for all groups. If ,, and (T stands for

61、 transpose), that is = min{}, and two nearest groups are chosen; </p><p>　　4) and are merged into , that is , so the total of groups decrease 1;</p><p>　　5) Return (2).</p><p>

62、　　4.2 Searching Corresponding Points Based on Geometric Invariance</p><p>　　Geometric invariance is defined that geometrical figure and vector keep invariance in mathematical manipulation.</p><p&g

63、t;　　For a special polygon, two different shape polygons will be obtained in two image planes by perspective transform. In the same way, for a 3D curve, two different 2D curves are obtained in two image planes. Therefore

64、geometric invariance is applied to matching straight lines and curves.</p><p>　　For straight-line matching, representational geometric invariance is composed of five points in the same straight line or five

65、straight lines in the same surface.</p><p>　　We assume that is arbitrary five straight lines on special plane, straight-line equation is </p><p><b>　　(3)</b></p><p>　　W

66、e arbitrarily choose three straight lines ,and in the five straight lines (k1,k2,k3=1,2,3,4,5,k1≠k2,k1,≠k3,k2≠k3). The system of equations of the three straight lines are given by</p><p><b>　　(4)</

67、b></p><p>　　And these straight lines are translated into image straight lines by perspective transform. The image straight lines have also corresponding straight-line equation parameters. And the paramet

68、ers are shown with superscript (for example ). It is testified, though the shapes of five straight lines can more change, there are geometric invariants, if M′is det A, they are</p><p>　　,

69、 (5)</p><p>　　Analogously, there are some geometric invariants for a group of quadratic curves. If is a quadratic curve on the special plane, its equation can be shown as follows</p><p><

70、b>　　(6)</b></p><p>　　And if is parameter matrix of quadratic curve, it is also shown by matrix as follows</p><p><b>　　(7)</b></p><p>　　If there are two quadrat

71、ic curves and ,their parameter matrixes are respectively and . They are translated into and by geometric projection, and their parameter matrixes are and . It is testified, if is track of matrix, there are two geom

72、etric invariants whether mode of geometric projection is changed.</p><p><b>　　(8)</b></p><p><b>　　(9)</b></p><p>　　In this way, straight lines and curves are

73、 matched effectively.</p><p>　　In this paper, grating is projected on object in vertical direction and lateral direction respectively, while two cameras grab images. With wavelet edge detection, searching no

74、n-supervisor, edge points are fitted into quadratic curves. With geometric invariance, quadratic curves are matched, and cross points of vertical curves and lateral curves are computed. So stereo precise matching at subp

75、ixel level is achieved.</p><p>　　5 Image Mosaic Based on Wavelet</p><p>　　When large-scale surface is measured, many pairs of stereo images are grabbed from different viewpoints or with moving

76、and rotating object. And two adjacent images need mosaic. The important question of image mosaic is image registration, that is to say, overlapped parts between two adjacent images are put in order, and corresponding mat

77、ching between two adjacent images is also involved in image mosaic. Corresponding matching in registration is deferred from corresponding matching in stereo visi</p><p>　　In this paper, some black points are

78、 pasted at random on object in order to mosaic easily, and the black points are regarded as registration feature points. Meanwhile, we use biorthogonal wavelet with linearity and symmetry to decompose two images that are

79、 will be mosaic, so the images can be matched and registered from coarse to fine on multi-scale, and lastly a big image is become.</p><p>　　In fact, wavelet transform is a band-pass filter, wavelet vector on

80、 different scales shows the stated width of frequency band, and so frequency bandwidth of each wavelet vector is unequal. Two images are decomposed into wavelet vectors on different frequency bands based on Mallat algori

81、thm, and then the different mosaic widths are selected on different scales to match and register, so a big mosaic image is smooth and fine.</p><p>　　6 Experiments and Results Analysis</p><p>　　

82、In this paper, the technology is applied to the measurement system of MPF machine. After the shaped surface is measured, the measuring results are returned to CAD/CAE system, and closed-loop control is achieved. Accordin

83、g to measurement condition, measurement precision and once shaped area, we select two cameras (MTV1881CB), two camera lenses and a image-record device (METEOR). The distance between two cameras is 300mm; the distance bet

84、ween the surface and the two cameras is 500mm. A 150×150 mm </p><p>　　1) Establish the mapping relation between image points and special points;</p><p>　　2) 3D surface is carried on the wor

85、ktable, firstly, two cameras grab synchronously a pair of stereo images without projecting and with light. Secondly, grab two pairs of stereo images, the one is grabbed while grating is put on vertical orientation, the o

86、ther is grabbed while grating is put in lateral direction;</p><p>　　3) Process image, remove background, and reduce noise, just as Fig.3a, 3b;</p><p>　　4) Detect feature, just as Fig.3c;</p&g

87、t;<p>　　5) Search corresponding points, and mosaic image;</p><p>　　6) Calculate 3D coordinate, reconstruct 3D surface, as shown in Fig.3d. </p><p>　　Experiment shows measurem

88、ent error is less than 0.5mm, measured time is about 2s including grabbing image, processing image, establishing the mapping relation of image point and special point, searching corresponding points and computing coordin

89、ate.</p><p>　　7 Conclusions</p><p>　　In this paper, a new 3D measurement technology based on computer vision is proposed with simple equipment, fast measurement speed and low cost. Large object

90、 can be measured, and measurement precision is less than 0.5 mm/m. It also provides a new idea for computer vision applied to industry. Experiments show that the 3D measurement technology has good robustness.</p>