CN100557379C - The binocular stereo vision measurement method of geometric parameters of spatial circle - Google Patents

Abstract

The invention discloses a kind of binocular stereo vision measurement method that is applicable to the geometric parameters of spatial circle of the geometric parameters of spatial circle non-contact type on-line measurement on the industrial products.Be intended to overcome the problem that measuring accuracy is not high, measuring speed is slow and automaticity is low of existence.Method is divided into calibration phase, Flame Image Process stage and space circle match stage.At first adopt nerual network technique to carry out camera calibration, utilize twin camera that space circle is carried out the sub-pix Edge extraction, develop simple algorithm based on the gradient of image and gray scale distribution characteristics then and realize the marginal point coupling, obtain the actual three dimensional space coordinate of rounded edge, at last by relevant mathematics geometric knowledge match space circle, try to achieve the geometric parameter of space circle, comprise the orientation on its radius, home position and plane, space circle place.Image processing speed of the present invention is fast, the automaticity height, and when angled about 50 ° of plane, space circle place and the plane of delineation, the relative error of measurement space circle is better than ± and 0.6%.

Description

The binocular stereo vision measurement method of geometric parameters of spatial circle

Technical field

The present invention relates to a kind of method that is applicable to the geometric parameters of spatial circle non-contact type on-line measurement on the industrial products, more particularly, it relates to a kind of binocular stereo vision measurement method of geometric parameters of spatial circle.

Background technology

Circle is a fundamental element of forming geometry of objects, and its precision characteristic often has influence on the final performance of entire product, thereby the geometric parameter of space circle has accurately been measured crucial meaning.Traditional geometric parameters of spatial circle measuring method adopts three coordinate measuring machine to carry out, but three coordinate measuring machine is bulky, and inconvenience is carried, and is contact type measurement, can not satisfy the needs that on-line measurement and large-scale circular hole are measured.Continuous development along with Flame Image Process, computer technology and industrial camera manufacture level, computer vision technique has also obtained development at full speed, can realize the quick measurement of object attitude and size in three dimensions, have that speed is fast, automaticity is high, good, the non-contacting advantage of precision, become important means and developing direction that product manufacturing dimension is measured just gradually.

More existing scholars did research for the vision measuring method of geometric parameters of spatial circle:

The method used of comrade Zhang Jianxin of University Of Tianjin be at first on the plane of delineation marginal point match circle with oval image draw centre coordinate, and then come (Zhang Jianxin by the three-dimensional coordinate that the inverse process of demarcating is obtained the space circle center of circle, the applied research of technique of binocular stereoscopic vision in commercial measurement, University Of Tianjin's doctorate paper, 1996), this method exists the error of principle, because it has ignored the change of shape that perspective projection brings, mistake is just big more more when the plane of delineation and plane included angle, space circle place, and can not the measurement space radius of a circle and the plane equation at circle place.

2. be published in Chinese Control in 2004 by Hu Chunhua, the cleer and peaceful Zhu Ji flood of scholar Liu three comrades and concentrate the paper of exercise question for " circle center image deformation error model research in the vision measurement " with the academic nd Annual Meeting of decision-making, careful analysis the difference between oval image center and the space circle center of circle subpoint, but do not provide the method for error compensation.

3. Chinese patent patent No. ZL 03142659.X, publication number CN1566900, open day 2005.01.19, applicant BJ University of Aeronautics ﹠ Astronautics, the vision measuring method of a kind of geometric parameters of spatial circle of invention and created name, this application case has proposed to utilize the polar curve constraint to obtain the actual three dimensional space coordinate of rounded edge and then carry out the space circle fitting method, its weak point is that this method needs artificial position and the quantity of determining polar curve, last measuring accuracy can be subjected to artificial factor, and the video camera measuring distance under its precision does not provide simultaneously.

4. Chinese patent application numbers 200710043742.2, applying date 2007.07.11, publication number CN101093160, open day 2007.12.26, applicant Shanghai Communications University, invention and created name is based on the measuring method of the geometric parameters of spatial circle of binocular stereo vision, the binocular stereo vision space circle measuring method that this application case proposes, though it is very suitable for independent navigation, but when industrial products are measured, world coordinate system is based upon circle centre position, and perhaps the method with left camera coordinate system opening relationships is unfavorable for global calibration.

Summary of the invention

Technical matters to be solved by this invention is to have overcome the problem that prior art exists, and a kind of binocular stereo vision measurement method of geometric parameters of spatial circle is provided.Make that space circle (comprising round place plane equation, radius of a circle and central coordinate of circle) measuring process automaticity is strong, both can satisfy the industrial products accuracy requirement, can reach very high precision again, and real-time.

For solving the problems of the technologies described above, the present invention adopts following technical scheme to be achieved.The binocular stereo vision measurement method of geometric parameters of spatial circle adopts calibration phase, Flame Image Process stage and space circle match stage.

The concrete steps of described calibration phase are as follows:

1) sets the 3 D stereo target, be covered with chequered with black and white gridiron pattern on the target face, tessellated size is 20mm * 20mm, with the public angle point of black box and white square as feature point for calibration, the quantity of feature point for calibration is 250～600, is that initial point is set up world coordinate system with three target hand-deliver points of three-dimensional target;

2) the fixed installation left and right cameras is set up the binocular tri-dimensional vision system, places three-dimensional target in the public view field scope of two video cameras, and photographic images for obtaining the image coordinate of feature point for calibration, improves Harris angle point extraction algorithm;

3) with left and right sides Method of Homonymy Points on Image to coordinate (u _Il, v _Il, u _Ir, v _Ir) as input, (z) as output, whole coordinate points is to being divided into three groups respectively as training sample, confirmatory sample and the test sample book of neural network for x, y for the feature point for calibration three-dimensional coordinate;

4) sample data is carried out normalized, make data area between-1～1;

5) set up network, the hidden neuron number is made as variable, mean square deviation is carried out network training as objective function, chooses the final neuron number of variable conduct that makes the objective function minimum, and suitably adopts the premature termination strategy, preserves network after training is finished;

6) utilization confirmatory sample and test sample book are confirmed the neural network of having built up and are tested, the generalization ability and the precision of checking network.

The concrete steps in described Flame Image Process stage are as follows:

1) space circle to be measured is placed in the public view field of binocular tri-dimensional vision system and takes, obtaining two width of cloth, to comprise the image of oval image of space circle right;

2) the logical OR computing between the utilization image, circle to be measured zone is separated from background image;

3) utilize 3 * 3 mean filter that image is handled earlier, remove the noise of some type in the image; Utilize median filter further to suppress noise again;

4) with the publish picture whole pixel edge of oval image of space circle on the picture plane of canny operator extraction, it is right to obtain the point that one group of coordinate by oval image edge pixel constitutes on the image of the left and right sides respectively;

5) with the shade of gray of image as matching characteristic, the utilization image is realized the coupling one by one of 2 image rounded edge points to shade of gray distance, neighborhood support and three constraint conditions of uniqueness, promptly respectively to the sad value setting threshold of shade of gray distance between the image border point of the left and right sides and marginal point neighborhood inside gradient and require satisfied uniqueness of mating to be limited;

6) utilization method of interpolation is rapidly and efficiently carried out floating-point operation to match point to coordinate, is accurate to the sub-pixel precision.

The concrete steps in described space circle match stage are as follows:

1), obtains the three dimensional space coordinate of space circle marginal point according to the neural network of having set up;

2) match on plane, space circle place;

3) rotation of space plane;

4) match of flat circle;

Obtain the plane equation on geometric center, radius and the plane, circle place of space circle.

Harris angle point extraction algorithm is improved described in the technical scheme is meant: be provided with one and merge threshold value on the basis of former algorithm, if two angle points just average interpolation arithmetic in merging threshold range, otherwise then keep former angular coordinate.The average interpolation operational formula is as follows:

$\left\{\begin{array}{c}{u}_m=\frac{u_i+u_j}{2}\\ v_n=\frac{v_i+v_j}{2}\end{array}\right.$

u _i, u _jThe angle point horizontal ordinate that extracts for original Harris algorithm;

v _i, v _jThe angle point ordinate that extracts for original Harris algorithm;

u _m, v _nFor improving angle point horizontal stroke, the ordinate that back Harris algorithm extracts.

The match on the plane, space circle place described in the technical scheme is meant: the equation of space plane can be expressed as z=Ax+By+C, utilize the method for regretional analysis according to Gauss-markov linear model, distribution situation by the space circle coordinate points draws the regression plane equation, can obtain the normal vector of space plane.

The rotation of the space plane described in the technical scheme is meant: the normal vector of establishing space plane is for (C), α, β, γ are respectively plane normal and x for A, B _w, y _w, z _wThe angle of axle then has relational expression:

$\left\{\begin{array}{c}\cos \mathrm{\α}=\frac{A}{\sqrt{A^2+B^2+C^2}}\\ \cos \mathrm{\β}=\frac{B}{\sqrt{A^2+B^2+C^2}}\\ \cos \mathrm{\γ}=\frac{C}{\sqrt{A^2+B^2+C^2}}\end{array}\right.---\left(1\right)$

Knowledge by space analysis geometry can pivot to space plane the position parallel with coordinate plane.

The match of the flat circle described in the technical scheme is meant: through the match on plane, space circle place and rotation two steps of space plane, space circle has been rotated on the plane parallel, used least square method that flat circle is carried out match again with coordinate plane.If the data point of match circular curve has n, circular curve can be expressed as x _i ²+ y _i ²+ Dx _i+ Ey _i+ F=0.

Error function e then _i=x _iD+y _iE+F+l _i, l wherein _i=x _i ²+ y _i ²Can list n error equation by n error function.Because of having only 3 unknown numbers, so can form 3 rank normal equations:

$\left[\begin{array}{ccc}\left[\mathrm{xx}\right]& \left[\mathrm{xy}\right]& \left[x\right]\\ \left[\mathrm{xy}\right]& \left[\mathrm{yy}\right]& \left[y\right]\\ \left[x\right]& \left[y\right]& n\end{array}\right]\left[\begin{array}{c}D\\ E\\ F\end{array}\right]=\left[\begin{array}{c}-\left[x(x^2+y^2)\right]\\ -\left[y(x^2+y^2)\right]\\ -[x^2+y^2]\end{array}\right]---\left(2\right)$

Then $\left[\begin{array}{c}D\\ E\\ F\end{array}\right]=-{\left[\begin{array}{ccc}\left[\mathrm{xx}\right]& \left[\mathrm{xy}\right]& \left[x\right]\\ \left[\mathrm{xy}\right]& \left[\mathrm{yy}\right]& \left[y\right]\\ \left[x\right]& \left[y\right]& n\end{array}\right]}^{-1}\left[\begin{array}{c}\left[x(x^2+y^2)\right]\\ \left[y(x^2+y^2)\right]\\ [x^2+y^2]\end{array}\right]---\left(3\right)$

Use the adjoint matrix method of inverting and solve (3) formula, at last:

Δ＝nOT+2QUP-Q ²T-U ²O-nP ² (4)

$D=\frac{1}{\mathrm{\Δ}}(U^{2}R-\mathrm{nTR}+\mathrm{nPV}-\mathrm{QUV}+\mathrm{QTW}-\mathrm{PUW})---\left(5\right)$

$E=\frac{1}{\mathrm{\Δ}}(\mathrm{nPR}-\mathrm{QUR}+Q^{2}V-\mathrm{nOV}+\mathrm{OUW}-\mathrm{PQW})---\left(6\right)$

$F=\frac{1}{\mathrm{\Δ}}(\mathrm{QTR}-\mathrm{PUR}+\mathrm{OUV}-\mathrm{QPV}+P^{2}W-\mathrm{OTW})---\left(7\right)$

Wherein: O=[xx], P=[xy], Q=[x] and, R=[x (x ²+ y ²)], T=[yy], U=[y], V=[y (x ²+ y ²)],

Try to achieve D, E, after F three parameters, then central coordinate of circle and radius of circle are:

$\left\{\begin{array}{c}{x}_0=-\frac{D}{2}\\ y_0=-\frac{E}{2}\\ R=\frac{1}{2}\sqrt{D^2+E^2-4F}\end{array}\right.$

Passing through contrary twiddle operation more just can be in the hope of space circle heart coordinate.

Compared with prior art the invention has the beneficial effects as follows:

1. the present invention can disposablely accurately measure all geometric parameters of space circle, and the utilization nerual network technique has been avoided the computing of loaded down with trivial details inside and outside direction parameter, and all distortion factors are included in the middle of the network.

2. the present invention can separate zone to be measured from background image, significantly reduced calculated amount, the entire image processing process is very fast, satisfies the needs that automatic online is measured in real time, in contactless industrial products on-line measurement very high using value is arranged.

3. the matching algorithm and the approximating method of the present invention's employing are easily understood, and have reduced the error of manual operation, therefore, make the measuring accuracy of the binocular stereo vision measurement method of geometric parameters of spatial circle satisfy the industrial products accuracy requirement, and can reach more high precision.

Description of drawings

The present invention is further illustrated below in conjunction with accompanying drawing:

Fig. 1 is the binocular tri-dimensional sense sensor mathematical model of the binocular stereo vision measurement method of geometric parameters of spatial circle;

Fig. 2 is the three-dimensional target synoptic diagram that is adopted in the binocular stereo vision measurement method of geometric parameters of spatial circle;

Fig. 3 is the image that adopts original Harris angle point extraction algorithm to extract;

Fig. 4 is the image that adopts the Harris angle point extraction algorithm after improving to extract;

Fig. 5 is neural metwork training figure;

Fig. 6 is the left measurement image that adopts the binocular tri-dimensional vision system in the binocular stereo vision measurement method of geometric parameters of spatial circle to take;

Fig. 7 is the right measurement image that adopts the binocular tri-dimensional vision system in the binocular stereo vision measurement method of geometric parameters of spatial circle to take;

Fig. 8 is the logical OR computing between the image that adopts in the binocular stereo vision measurement method of geometric parameters of spatial circle splits circle to be measured zone from background image the left measurement image through amplifying;

Fig. 9 is the logical OR computing between the image that adopts in the binocular stereo vision measurement method of geometric parameters of spatial circle splits circle to be measured zone from background image the right measurement image through amplifying;

Figure 10 adopts the canny operator extraction to publish picture to measure the whole pixel edge image of oval image as the left side of space circle on the plane;

Figure 11 adopts the canny operator extraction to publish picture to measure the whole pixel edge image of oval image as the right side of space circle on the plane;

Figure 12 is the image coordinate of left image bore edges match point to be measured;

Figure 13 is the image coordinate of right image bore edges match point to be measured;

Figure 14 is the three dimensional space coordinate at space circle edge.

Embodiment

Below in conjunction with accompanying drawing the present invention is explained in detail:

Consult Fig. 1, shown in the figure is the binocular tri-dimensional vision system of optional position.O _wx _wy _wz _wBe world coordinate system, O _lx _ly _lz _lAnd O _rx _ry _rz _rBe respectively the left and right cameras coordinate system, o _lu _lv _lAnd o _ru _rv _rBeing respectively with the pixel is the left and right sides image coordinate system of unit.The binocular stereo vision measurement method of geometric parameters of spatial circle can be divided into calibration phase, Flame Image Process stage and space circle match stage, and its concrete implementation step is as follows:

1. calibration phase

1) sets the 3 D stereo target, be covered with chequered with black and white gridiron pattern on the target face, tessellated size is 20mm * 20mm, with the public angle point of black box and white square as feature point for calibration, the quantity of feature point for calibration is 250～600, with three target hand-deliver points of three-dimensional target is that initial point is set up world coordinate system, so the volume coordinate of feature point for calibration under world coordinate system just known.

2) fix left and right cameras and set up the binocular tri-dimensional vision system, in the public view field scope of two video cameras, place three-dimensional target, photographic images.For obtaining the image coordinate of feature point for calibration, to Harris angle point extraction algorithm (C.G.Harris and M.J.Stephens.A combined corner and edgedetector[J] .Proceedings Fourth Alvey Vision Conference, Manchester.1988,147-151) improve, thereby eliminate the polysemy phenomenon of former algorithm existence and improve coordinate precision.The improvement algorithm is: be provided with one and merge threshold value on the basis of former algorithm, if two angle points just average interpolation arithmetic in merging threshold range, otherwise then keep former angular coordinate.The average interpolation operational formula is as follows:

$\left\{\begin{array}{c}{u}_m=\frac{u_i+u_j}{2}\\ v_n=\frac{v_i+v_j}{2}\end{array}\right.$

u _i, u _jThe angle point horizontal ordinate that extracts for original Harris algorithm;

v _i, v _jThe angle point ordinate that extracts for original Harris algorithm;

u _m, v _nFor improving angle point horizontal stroke, the ordinate that back Harris algorithm extracts;

3) with left and right sides Method of Homonymy Points on Image to coordinate (u _Il, v _Il, u _Ir, v _Ir) as input, (z) as output, whole coordinate points is to being divided into three groups respectively as training sample, confirmatory sample and the test sample book of neural network for x, y for the feature point for calibration three-dimensional coordinate.

4) sample data is carried out normalized (A.S.Weigand, D.E.Rumelhart, B.A.Hubeman.Generalization by weight elimination with application toforcasting.In Advances in Neural Information Processing System3, R.P.Lippman, J.E.Moody and D.J.Touretzky, eds, San Mateo, CA:MorganKaufmann, 1991,575-582), make data area between-1～1.

5) set up network, the hidden neuron number is made as variable, mean square deviation is carried out network training as objective function, choose the final neuron number of variable conduct that makes the objective function minimum, and suitably adopt premature termination (W.S.Sarle.Stopped training and other remendies for overfitting, toappear in Proceedings of the 27 ^ThSymposium on the Interface, 1995) strategy, after finishing, training preserves network.

6) utilization confirmatory sample and test sample book are confirmed the neural network of having built up and are tested, the generalization ability and the precision of checking network.

2. Flame Image Process stage

1) space circle to be measured is placed in the public view field of stereo visual system and takes, obtaining two width of cloth, to comprise the image of oval image of space circle right.

2) the logical OR computing between the utilization image, circle to be measured zone is separated from background image.

3) utilize 3 * 3 mean filter that image is handled earlier, remove the noise of some type in the image; Utilize median filter further to suppress noise again, can protect marginal information, for edge extracting is laid a good foundation.

4) with the publish picture whole pixel edge of oval image of space circle on the picture plane of canny operator extraction, it is right to obtain the point that one group of coordinate by oval image boundary pixel constitutes on the image of the left and right sides respectively.

5) with the shade of gray of image as matching characteristic, the utilization image is realized the coupling one by one of 2 image rounded edge points to three constraint conditions such as a shade of gray distance, neighborhood support and uniqueness, promptly respectively SAD (Sum of AbsoluteDifferences, absolute difference sum) the value setting threshold of shade of gray distance between the image border point of the left and right sides and marginal point neighborhood inside gradient and the uniqueness that requires to satisfy coupling are limited.

6) utilization method of interpolation (Wu Xiaobo etc. rapidly and efficiently, use the resolution that polynomial interpolating function improves the area array CCD dimensional measurement, be published in Chinese journal of scientific instrument, 1996,7 (2): 154) match point is carried out floating-point operation to coordinate, be accurate to the sub-pixel precision.

3. space circle match stage

1), obtains the three dimensional space coordinate of space circle marginal point according to the neural network of having set up.

2) match on plane, space circle place

The equation of space plane can be expressed as z=Ax+By+C, utilize the method for regretional analysis according to Gauss's one markov linear model, distribution situation by the space circle coordinate points draws regression plane equation (Zhou Fugong, Huang Yuncheng, use linear regression analysis [M], publishing house of the Renmin University of China, 1989.8), can obtain the normal vector of space plane.

3) rotation of space plane

If the normal vector of space plane is that (C), α, β, γ are respectively plane normal and x for A, B _w, y _w, z _wThe angle of axle then has relational expression:

$\left\{\begin{array}{c}\cos \mathrm{\α}=\frac{A}{\sqrt{A^2+B^2+C^2}}\\ \cos \mathrm{\β}=\frac{B}{\sqrt{A^2+B^2+C^2}}\\ \cos \mathrm{\γ}=\frac{C}{\sqrt{A^2+B^2+C^2}}\end{array}\right.---\left(1\right)$

Knowledge by space analysis geometry can pivot to space plane the position parallel with coordinate plane.

4) match of flat circle

Through preceding two steps, space circle has been rotated on the plane parallel with coordinate plane, use least square method that flat circle is carried out match again.If the data point of match circular curve has n, circular curve can be expressed as x _i ²+ y _i ²+ Dx _i+ Ey _i+ F=0.

Error function e then _i=x _iD+y _iE+F+l _i, l wherein _i=x _i ²+ y _i ²Can list n error equation by n error function.Because of having only 3 unknown numbers, so can form 3 rank normal equations:

$\left[\begin{array}{ccc}\left[\mathrm{xx}\right]& \left[\mathrm{xy}\right]& \left[x\right]\\ \left[\mathrm{xy}\right]& \left[\mathrm{yy}\right]& \left[y\right]\\ \left[x\right]& \left[y\right]& n\end{array}\right]\left[\begin{array}{c}D\\ E\\ F\end{array}\right]=\left[\begin{array}{c}-\left[x(x^2+y^2)\right]\\ -\left[y(x^2+y^2)\right]\\ -[x^2+y^2]\end{array}\right]---\left(2\right)$

Then $\left[\begin{array}{c}D\\ E\\ F\end{array}\right]=-{\left[\begin{array}{ccc}\left[\mathrm{xx}\right]& \left[\mathrm{xy}\right]& \left[x\right]\\ \left[\mathrm{xy}\right]& \left[\mathrm{yy}\right]& \left[y\right]\\ \left[x\right]& \left[y\right]& n\end{array}\right]}^{-1}\left[\begin{array}{c}\left[x(x^2+y^2)\right]\\ \left[y(x^2+y^2)\right]\\ [x^2+y^2]\end{array}\right]---\left(3\right)$

Use the adjoint matrix method of inverting and solve (3) formula, at last:

Δ＝nOT+2QUP-Q ²T-U ²O-nP ² (4)

$D=\frac{1}{\mathrm{\Δ}}(U^{2}R-\mathrm{nTR}+\mathrm{nPV}-\mathrm{QUV}+\mathrm{QTW}-\mathrm{PUW})---\left(5\right)$

$E=\frac{1}{\mathrm{\Δ}}(\mathrm{nPR}-\mathrm{QUR}+Q^{2}V-\mathrm{nOV}+\mathrm{OUW}-\mathrm{PQW})---\left(6\right)$

$F=\frac{1}{\mathrm{\Δ}}(\mathrm{QTR}-\mathrm{PUR}+\mathrm{OUV}-\mathrm{QPV}+P^{2}W-\mathrm{OTW})---\left(7\right)$

Wherein:

O＝[xx]，P＝[xy]，Q＝[x]，R＝[x(x ²+y ²)]，T＝[yy]，U＝[y]，V＝[y(x ²+y ²)]，

Try to achieve D, E, after F three parameters, then central coordinate of circle and radius of circle are:

$\left\{\begin{array}{c}{x}_0=-\frac{D}{2}\\ {\mathrm{<figure-callout\; id="0"\; label="y"\; filenames="C20081005063000151.png"\; state="\{\{state\}\}">y</figure-callout>}}_0=-\frac{E}{2}\\ R=\frac{1}{2}\sqrt{D^2+E^2-4F}\end{array}\right.$

Passing through contrary twiddle operation more just can be in the hope of space circle heart coordinate.

Like this, all geometric parameters of space circle comprise that plane, space circle place, space central coordinate of circle and space circle radius size all measured.

Embodiment

According to the step of the method for narrating above, the model of employing left and right cameras is the CMOS of DH-HV1302UM-T, and resolution is 1280 * 1024, and camera lens is the CCTV lens of 12.5mm-75mm F1.8.

Represented among Fig. 2 is three-dimensional target figure.Represented among Fig. 3 is the image that the Harris algorithm angle point before utilization improves extracts.Represented among Fig. 4 is the image that the Harris algorithm angle point after utilization improves extracts.To as importing, the volume coordinate under the world coordinate system is as output with the angular coordinate of the same name that obtains.Have 588 pairs of coordinates, the choosing wherein 294 pairs as training sample set, 147 pairs is the confirmatory sample collection, 147 pairs as the test sample book collection.Represented among Fig. 5 is the training plan of neural network, and drawing best hidden neuron number is 16, preserves network.Represented among Fig. 6 is that the binocular tri-dimensional vision system contains the left measurement image that space circular hole part is taken to one.Represented among Fig. 7 is that the binocular tri-dimensional vision system contains the right measurement image that space circular hole part is taken to one.Represented among Fig. 8 is the logical OR computing between image splits circle to be measured zone from background image the left measurement image through amplifying.Represented among Fig. 9 is the logical OR computing between image splits circle to be measured zone from background image the right measurement image through amplifying.Represented among Figure 10 is to adopt the canny operator extraction to publish picture to measure the whole pixel edge image of oval image as the left side of space circle on the plane.Represented among Figure 11 is to adopt the canny operator extraction to publish picture to measure the whole pixel edge image of oval image as the right side of space circle on the plane.Represented among Figure 12 is the distribution situation of coordinate on the plane of delineation of left image bore edges match point to be measured.Represented among Figure 13 is the distribution situation of coordinate on the plane of delineation of right image bore edges match point to be measured.Represented among Figure 14 is the three dimensional space coordinate at space circle edge.Normal vector through match space circle place plane equation is (0.30606,0.014453 ,-1), then space plane is rotated 1.2738 (Circular measures, down together) around the x axle, and then around z axle rotation 2.8443, space plane just rotates to and x like this _wo _wz _wParallel plane position.The match of planar justifying again just can obtain the space circle parameter after contrary rotation.The plane, space circle place and the plane of delineation be into about 50 ° angle in the example, the about 2.5m of video camera measuring distance, and the space circle parameter is:

Central coordinate of circle: (246.204,197.341,261.810) mm

The normal vector on plane, circle place: (0.30606,0.014453 ,-1)

Radius of circle: R=12.986mm differs 0.076mm with the real radius of justifying.

Claims (5)

1. the binocular stereo vision measurement method of a geometric parameters of spatial circle adopts calibration phase, Flame Image Process stage and space circle match stage, it is characterized in that,

The concrete steps of described calibration phase are as follows:

1) sets the 3 D stereo target, be covered with chequered with black and white gridiron pattern on the target face, tessellated size is 20mm * 20mm, with the public angle point of black box and white square as feature point for calibration, the quantity of feature point for calibration is 250～600, is that initial point is set up world coordinate system with three target hand-deliver points of three-dimensional target;

2) the fixed installation left and right cameras is set up the binocular tri-dimensional vision system, places three-dimensional target in the public view field scope of two video cameras, and photographic images for obtaining the image coordinate of feature point for calibration, improves Harris angle point extraction algorithm;

3) with left and right sides Method of Homonymy Points on Image to coordinate (u _Il, v _Il, u _Ir, v _Ir) as input, (z) as output, whole coordinate points is to being divided into three groups respectively as training sample, confirmatory sample and the test sample book of neural network for x, y for the feature point for calibration three-dimensional coordinate;

4) sample data is carried out normalized, make data area between-1～1;

5) set up network, the hidden neuron number is made as variable, mean square deviation is carried out network training as objective function, chooses the final neuron number of variable conduct that makes the objective function minimum, and suitably adopts the premature termination strategy, preserves network after training is finished;

6) utilization confirmatory sample and test sample book are confirmed the neural network of having built up and are tested, the generalization ability and the precision of checking network;

The concrete steps in described Flame Image Process stage are as follows:

1) space circle to be measured is placed in the public view field of binocular tri-dimensional vision system and takes, obtaining two width of cloth, to comprise the image of oval image of space circle right;

2) the logical OR computing between the utilization image, circle to be measured zone is separated from background image;

3) utilize 3 * 3 mean filter that image is handled earlier, remove the noise of some type in the image; Utilize median filter further to suppress noise again;

4) with the publish picture whole pixel edge of oval image of space circle on the picture plane of canny operator extraction, it is right to obtain the point that one group of coordinate by oval image edge pixel constitutes on the image of the left and right sides respectively;

5) with the shade of gray of image as matching characteristic, the utilization image is realized the coupling one by one of 2 image rounded edge points to shade of gray distance, neighborhood support and three constraint conditions of uniqueness, promptly respectively to the sad value setting threshold of shade of gray distance between the image border point of the left and right sides and marginal point neighborhood inside gradient and require satisfied uniqueness of mating to be limited;

6) utilization method of interpolation is rapidly and efficiently carried out floating-point operation to match point to coordinate, is accurate to the sub-pixel precision;

The concrete steps in described space circle match stage are as follows:

1), obtains the three dimensional space coordinate of space circle marginal point according to the neural network of having set up;

2) match on plane, space circle place;

3) rotation of space plane;

4) match of flat circle;

Obtain the plane equation on geometric center, radius and the plane, circle place of space circle.

2. according to the binocular stereo vision measurement method of the described geometric parameters of spatial circle of claim 1, it is characterized in that, described Harris angle point extraction algorithm is improved is meant: be provided with one and merge threshold value on the basis of former algorithm, if two angle points just average interpolation arithmetic in merging threshold range, otherwise then keep former angular coordinate; The average interpolation operational formula is as follows:

$\left\{\begin{array}{c}{u}_m=\frac{u_i+u_j}{2}\\ v_n=\frac{v_i+v_j}{2}\end{array}\right.$

u _i, u _jThe angle point horizontal ordinate that extracts for original Harris algorithm;

v _i, v _jThe angle point ordinate that extracts for original Harris algorithm;

u _m, v _nFor improving angle point horizontal stroke, the ordinate that back Harris algorithm extracts.

3. according to the binocular stereo vision measurement method of the described geometric parameters of spatial circle of claim 1, it is characterized in that, the match on plane, described space circle place is meant: the equation of space plane can be expressed as z=Ax+By+C, utilize the method for regretional analysis according to Gauss-markov linear model, distribution situation by the space circle coordinate points draws the regression plane equation, can obtain the normal vector of space plane.

4. according to the binocular stereo vision measurement method of the described geometric parameters of spatial circle of claim 1, it is characterized in that the rotation of described space plane is meant: the normal vector of establishing space plane is for (C), α, β, γ are respectively plane normal and x for A, B _w, y _w, z _wThe angle of axle then has relational expression:

$\left\{\begin{array}{c}\cos \mathrm{\&alpha;}=\frac{A}{\sqrt{A^2+B^2+C^2}}\\ \cos \mathrm{\&beta;}=\frac{B}{\sqrt{A^2+B^2+C^2}}\\ \cos \mathrm{\&gamma;}=\frac{C}{\sqrt{A^2+B^2+C^2}}\end{array}\right.---\left(1\right)$

Knowledge by space analysis geometry can pivot to space plane the position parallel with coordinate plane.

5. according to the binocular stereo vision measurement method of the described geometric parameters of spatial circle of claim 1, it is characterized in that, the match of described flat circle is meant: through the match on plane, space circle place and rotation two steps of space plane, space circle has been rotated on the plane parallel with coordinate plane, use least square method that flat circle is carried out match again, if the data point of match circular curve has n, circular curve can be expressed as x _i ²+ y _i ²+ Dx _i+ Ey _i+ F=0,

Error function e then _i=x _iD+y _iE+F+l _i, l wherein _i=x _i ²+ y _i ², can list n error equation by n error function, because of having only 3 unknown numbers, so can form 3 rank normal equations:

$\left[\begin{array}{ccc}\left[\mathrm{xx}\right]& \left[\mathrm{xy}\right]& \left[x\right]\\ \left[\mathrm{xy}\right]& \left[\mathrm{yy}\right]& \left[y\right]\\ \left[x\right]& \left[y\right]& n\end{array}\right]\left[\begin{array}{c}D\\ E\\ F\end{array}\right]=\left[\begin{array}{c}-\left[x(x^2+y^2)\right]\\ -\left[y(x^2+y^2)\right]\\ -[x^2+y^2]\end{array}\right]---\left(2\right)$

Then $\left[\begin{array}{c}D\\ E\\ F\end{array}\right]=-{\left[\begin{array}{ccc}\left[\mathrm{xx}\right]& \left[\mathrm{xy}\right]& \left[x\right]\\ \left[\mathrm{xy}\right]& \left[\mathrm{yy}\right]& \left[y\right]\\ \left[x\right]& \left[y\right]& n\end{array}\right]}^{-1}\left[\begin{array}{c}\left[x(x^2+y^2)\right]\\ \left[y(x^2+y^2)\right]\\ [x^2+y^2]\end{array}\right]---\left(3\right)$

Use the adjoint matrix method of inverting and solve (3) formula, at last:

Δ＝nOT+2QUP-Q ²T-U ²O-nP ² (4)

$D=\frac{1}{\mathrm{\&Delta;}}(U^{2}R-\mathrm{nTR}+\mathrm{nPV}-\mathrm{QUV}+\mathrm{QTW}-\mathrm{PUW})---\left(5\right)$

$E=\frac{1}{\mathrm{\&Delta;}}(\mathrm{nPR}-\mathrm{QUR}+Q^{2}V-\mathrm{nOV}+\mathrm{OUW}-\mathrm{PQW})---\left(6\right)$

$F=\frac{1}{\mathrm{\&Delta;}}(\mathrm{QTR}-\mathrm{PUR}+\mathrm{OUV}-\mathrm{QPV}+P^{2}W-\mathrm{OTW})---\left(7\right)$

Wherein: O=[xx], P=[xy], Q=[x] and, R=[x (x ²+ y ²)], T=[yy], U=[y], V=[y (x ²+ y ²)],

Try to achieve D, E, after F three parameters, then central coordinate of circle and radius of circle are:

$\left\{\begin{array}{c}{x}_0=-\frac{D}{2}\\ y_0=-\frac{E}{2}\\ R=\frac{1}{2}\sqrt{D^2+E^2-4F}\end{array}\right.$

Passing through contrary twiddle operation more just can be in the hope of space circle heart coordinate.

Images