CN1963390A - A precision and high efficiency three-dimensional measuring method - Google Patents

Abstract

This invention provides one method to process non-contact three dimension measurement, which comprises the following steps: user uses handle cameral to catch aim to be tested and to put label within visual field to get visual sequence into computer through steps of characteristic point tracing, key frame extracting, radio establishing, self-locking, oral re-establishing to restore the field three dimensional structure composed of measurement aim and label; the structure size can be determined by label geometry information; user selects two points with relative space distance automatically computed out.

Description

A kind of method for three-dimensional measurement of convenient and efficient

Technical field

The present invention relates to target is carried out contactless three-dimensional measurement, be particularly useful for measuring the far away or complicated target of scene of difficult contact, distance, have higher degree of accuracy.

Background technology

Measuring technique has been brought into play considerable effect in the development of the national economy, it is widely used in work such as engineering design, auxiliary construction, structural appraisal, deformation monitoring by the geological information of acquisition targets such as measuring distance, angle.Measuring method commonly used has: 1. directly measure by measurer, as steel ruler, mark do, survey pin and plumb bob etc., be example with the taped measurement, need obtain distance value at last through alignment piling, span, measurement height of pile top, slipstick segment length.2. utilize the optical photoconductor exact instrument: as electro-optical distance instrument, optical theodolite, precision level etc., be example with the electro-optical distance instrument, desire is measured the distance of A, B point-to-point transmission, and in the A point, catoptron is placed in the B point with equipment placement.The modulation of instrument emission turns back to instrument after reflection by A to B, measures this modulated light wave and come and go the phase shift that propagation is produced on survey line, to measure distance.3.GPS satellite positioning surveys, collaborative work by space 21 work satellite and ground monitoring station, the user uses measurement type GPS receiver simultaneous observation more than two or two, observed reading is handled, can obtain WGS-84 baseline vector accurate between two measurement points, pass through work such as adjustment, coordinate transmission, coordinate conversion again, finally obtain the coordinate of measuring point.

Above-mentioned measuring method respectively has quality on operability and accuracy, but common limitation is arranged.At first, they all can only measure some characteristic quantities at every turn, can not measure many characteristic quantities of a certain target simultaneously, such as the distance between the length that will measure some houses and the different parts, layout repeatedly just, repeatedly measure, this certainly will increase workload.Secondly, they must be measured touching target because instrument to be placed in measured target certain a bit, if target far, the layout that is difficult to not accessible or with a varied topography, will increase the difficulty and the error of measurement.Therefore, a kind of non-contact 3-D measuring method that can obtain the many-sided metrical information of target simultaneously of necessary proposition is to improve efficient of measuring and simplification and the accuracy of measuring under given conditions.

Summary of the invention

The invention provides a kind of method, be used to carry out contactless three-dimensional measurement, can overcome the limitation of above-mentioned various measuring methods, may further comprise the steps:

(a) in the scene of measured target place, place mark;

(b) with hand-held camera scene is taken;

(c) video sequence is imported computing machine and carried out three-dimensional reconstruction, obtain the accurate three-dimensional structure of scene;

(d) determine the yardstick of reconstruction of three-dimensional structure by the geological information of mark.

(e) by the reconstructed results of mark degree target is measured.

In above-mentioned steps (a), the mark of use is the object with particular geometric information, and it is known and constant that some monumented points, the distance between these monumented points are promptly arranged on the mark.Guarantee that the mark position do not block certain part of measurement target all the time in shooting, and can not be from the distance of video camera too closely to avoid producing distortion.

In above-mentioned steps (b), use common hand-held camera to take continuously, shooting process must move, but speed wants even, avoids violent shake, guarantees that measurement target and mark are arranged in the visual field.Need not demarcate video camera in advance, promptly the inner parameter of video camera is unknown.

The first step in above-mentioned steps (c), utilization KLT track algorithm, unique point in the video sequence is extracted and followed the tracks of, and utilize life cycle and utmost point geometrical constraint is filtered tracking results and optimizes, obtain the high good unique point track of fiduciary level, these tracks have been represented the coordinate of unique point on each two field picture; Utilization GRIC discrimination standard is chosen certain parallax and is had the wide picture of enough baselines as key frame from successive frame.

In second step in above-mentioned steps (c), utilize key frame and unique point track to carry out projection and rebuild.Adopt the sequence type method, at first use RANSAC to remove the wild point (i.e. Cuo Wu match point) of two width of cloth key frame, calculate the initial value of the fundamental matrix F of this two frame then with 8 methods of normalization, with the MLE maximal possibility estimation F is optimized again, obtain accurate fundamental matrix F.Next utilize the projection matrix and the initial spatial point of trigonometric ratio of fundamental matrix F structure two width of cloth key frame, utilize existing spatial point to calculate the projection matrix and the new spatial point of trigonometric ratio of new key frame again, carry out the space structure expansion with this, simultaneously the spatial point of having rebuild is optimized, to reduce the re-projection error of spatial point on had been friends in the past key frame.Because video camera is not demarcated, obtaining at last is three-dimensional structure under the projective coordinate system.

Deposit the step of the 3rd in the above-mentioned steps (c), carry out Euler's reconstruction by self-calibrating method.At first the video camera projection matrix is carried out normalization, calculate antithesis absolute quadric Ω then ^*, in calculating, adopt 5 parameters to represent 4 * 4 symmetric matrix Ω ^*, at last by decomposing Ω ^*Obtain the transformation matrix T of projective space, the three-dimensional structure under the projective coordinate system be multiply by T transform to Eulerian coordinates system down, finish Euler's reconstruction to the Euler space.

In the 4th step in above-mentioned steps (c), reconstructed results is carried out global optimization.The method " boundling adjustment " that has powerful search capability in the higher-dimension parameter space has been adopted in global optimization, carries out twice altogether, is respectively the result that projection is rebuild and Euler rebuilds is optimized, and the re-projection error is dropped to minimum value.

In above-mentioned steps (d), by the geological information and the actual geological information thereof of mark in Euler's reconstructed results, determine to rebuild the corresponding actual geometry of structural units's length, be translated into the three-dimensional structure of mark degree.

In above-mentioned steps (e), desire is measured the distance between 2 in the scene, on arbitrary two field picture of video sequence, choose this two points, it is also definite thereupon corresponding to the three dimensions point of rebuilding in the structure, then this actual geometric distance of 2 can be directly reconstructed results by the mark degree calculate.

The present invention is by the reconstruction to object scene, obtains accurate three-D space structure, can assess any distance of object scene simultaneously, has greatly improved efficiency of measurement.And owing to adopted video camera to take, operating personnel needn't only need allow it be arranged in the visual field of video camera near measurement target, therefore not limited by site condition, greatly reduce the difficulty of measurement.

Description of drawings

By example, embodiments of the invention are described in conjunction with the accompanying drawings, wherein:

Fig. 1 is the schematic block diagram that has embodied non-contact type measuring method of the present invention;

Fig. 2 is the synoptic diagram that has embodied mark example of the present invention;

Fig. 3 has embodied the schematic block diagram that the present invention utilizes reconstructed results to measure.

Embodiment

Among the embodiment that is described in detail below, method provided by the invention has been finished the non-contact measurement to a specific objective, and this target is a house to be assessed.

At first select a mark post that has monumented point thing as a token of, be positioned over the centre position in video camera and house, the distance of monumented point is known on the mark post.With the normal domestic use hand-held camera house is taken then, moving line is preferably camber line, goes to the right front from the left front in house, and guarantees that house and mark are arranged in the visual field all the time.Taking a period of time is 20 seconds, the video sequence of about 500 frames.

Video sequence is imported computing machine to be handled.At first use KLT signature tracking device that the video of 500 frames is followed the tracks of, obtain the unique point track of about 9000 different lengths.There is greatly the relatively shorter or centre of length comprise wild point in these tracks, therefore very unreliable, must remove.Selecting life cycle is 20 frames, and the track that every like this length is lower than 20 frames is all filtered, and what these tracks often traced into is the very unconspicuous point of feature.Utilization can be removed wild point to utmost point geometrical constraint.Mostly be in the geometric theory that in computer vision the relation of two width of cloth figure can be described with fundamental matrix F.The satisfied x ' that concerns of the corresponding point x of two width of cloth figure → x ' ^TFx=0.Further, for the some x on first width of cloth figure, the polar curve l '=Fx of its correspondence on second width of cloth figure has l=F for x ' equally ^TX '.To utmost point geometrical constraint is exactly that geometric distance δ to match point differentiates, and δ is defined as

δ＝d(x _i′，Fx _i) ²+d(x _i，F ^Tx _i′) ² (1)

If δ is greater than certain thresholding (being 0.5 pixel here), just think that this match point does not satisfy the relation to utmost point geometry, promptly differentiate and be " wild point ", if wild point has appearred in certain bar track, just be considered as untrustworthy track, it the later part of wild point occurred clip, keep the correct part in front.And then, obtain the very high unique point track of about 5000 length long enoughs and good degree like this with life cycle filtration one time.

Next all successive frames are handled, because the parallax of two continuous frames is very little, baseline is very short, can introduce bigger error when the trigonometric ratio spatial point, promptly so-called " narrow baseline " problem.So must select the key frame of enough baseline widths rebuilds.Adopt the GRIC discrimination standard to select key frame, two models of fundamental matrix F and homography matrix H can be used for describing the relation of two width of cloth figure, and describing with H that baseline is narrow is more suitable for.The GRIC standard is calculated the mark of two kinds of models respectively based on the match point of two width of cloth figure, and the model that mark is low more meets data

$\mathrm{GRIC}=\mathrm{\Σ\ρ}\left(e_i^2\right)+(\mathrm{nd}\ln r+k\ln \mathrm{rn})---\left(2\right)$

R is the dimension of data, and n is total number of unique point.When the baseline long enough, promptly F more meets the relation of describing two width of cloth figure, and then the mark of GRIC (F) is lower than GRIC (H).So since the i=0 frame, as GRIC (F _Ij)＜GRIC (H _Ij) time select the j frame as next key frame, continue down to calculate from the i=j frame, final, from 500 width of cloth successive frames, pick out 98 width of cloth key frames.

Projection is rebuild and to be based on that above-mentioned 4000 unique point tracks that obtain and 98 width of cloth key frames carry out.It at first is the fundamental matrix of wanting accurate Calculation two width of cloth key frame.Because still there is wild point inevitably,, obtain big as far as possible interior point (correct match point) set so need remove wild point with the RANSAC algorithm.This is the process of an iteration, and 7 pairs of match points of first picked at random calculate fundamental matrix

Calculate the re-projection error d of every pair of match point then _⊥

$d_{\⊥}=\underset{i}{\mathrm{\Σ}}d(x_i,{\hat{x}}_i{)}^2+d{(x_i^{\′},{\hat{x}}_i^{\′},)}^2---\left(3\right)$

Calculate re-projection error d _⊥The number of the interior point of＜0.5 pixel.Repeat the interior some set of sampling maximum 500 times.8 pairs of match points choosing some set in this then calculate fundamental matrix initial value F linearly.Utilize MLE maximum likelihood estimation approach at last, by minimizing the quadratic sum of all match point re-projection errors:

$\underset{i}{\overset{n}{\mathrm{\Σ}}}d{(x_i,{\hat{x}}_i)}^2+d{(x_i^{\′},{\hat{x}}_i^{\′})}^2---\left(4\right)$

Come F is optimized, obtain the best estimate of fundamental matrix.

The projection matrix of two key frames is built by acquired fundamental matrix.First key frame is P=[I|0], second key frame is P '=[[e '] _xF|e '], e ' is second limit on the key frame, the relation of it and F is F ^TE '=0.For the initialization of spatial point X, by the match point x → x ' on two key frames and two projection matrix P, P ' trigonometric ratio obtains, and promptly separates system of linear equations:

x×(PX)＝0

x′×(P′X)＝0 (5)

To each new key frame k, calculate its projection matrix by the spatial point of having rebuild earlier, the newer spatial point of trigonometric ratio.In the spatial point of having rebuild some still as seen, they form the 3D-2D coupling with the corresponding point on the new key frame in new key frame.Calculate projection matrix P with all 3D-2D match points by equation (6) earlier _kInitial value

x＝PX (6)

Utilize the RANSAC method to seek some set in the maximum 3D-2D once more, in this, optimize P by minimizing formula (7) on the basis of some set then _k

∑d(P _kX，x) ² (7)

Utilize new key frame P then _k, last key frame P _K-1And new match point x in this two frame _K-1→ x _kBy the new spatial point of (5) formula trigonometric ratio.To be optimized existing space structure simultaneously, to reduce the re-projection error of spatial point on current all key frames.Find the solution the weighted linear system of equations

$\frac{1}{P_3\tilde{X}}\left[\begin{array}{c}{P}_{3}a-P_1\\ P_{3}b-P_2\end{array}\right]X=0---\left(8\right)$

Can obtain the optimum value of X.Here

Be the value of X before optimizing, P _iThe i that is P is capable, and (a b) is the coordinate of point on image.After all key frames were handled, the scene structure under the projective coordinate system was rebuild and is finished.

Projective coordinate system is not the statement to real world, and the projection reconstructed results must be switched to the meaning that Eulerian coordinates system just has measurement down, and this is by realizing from demarcating.Antithesis absolute quadric Ω ^*Under projective coordinate system, can be expressed as 4 * 4 orders and be 3 symmetric matrix, and deposit the Ω down of Eulerian coordinates system ^*Be constantly equal to diagonal matrix diag (1,1,1,0).From the target of demarcating is to calculate Ω under the projective coordinate system ^*Value, and find the transition matrix T that it can be transformed to diag (1,1,1,0).According to P Ω ^*P ^T∝ KK ^T, and PO=[I|0], Ω ^*Can be set at

${\mathrm{\Ω}}^{*}=\left[\begin{array}{cc}K{K}^T& a\\ a^T& b\end{array}\right]---\left(9\right)$

A=[a wherein ₁, a ₂, a ₃] ^T, K is the intrinsic parameter of video camera.K is applied strong constraint, and promptly only focal distance f is unknown, then Ω ^*Can use f, a ₁, a ₂, a ₃, b5 parameter represented:

${\mathrm{\Ω}}^{*}=\left[\begin{array}{cccc}{f}^2+\frac{1}{4}{w}^2& \frac{1\mathrm{wh}}{4}& \frac{1}{2}w& a_1\\ \frac{1}{4}\mathrm{wh}& f^2+\frac{1}{4}{h}^2& \frac{1}{2}h& a_2\\ \frac{1}{2}w& \frac{1}{2}h& 1& a_3\\ a_1& a_2& a_3& b\end{array}\right]---\left(10\right)$

Wherein w and h are respectively the wide and high of image.Solving equation group (11) iteratively can be in the hope of Ω ^*

$\frac{1}{9v}(P_1{\mathrm{\Ω}}^{*}{P}_1^T-P_3{\mathrm{\Ω}}^{*}{P}_3^T)=0$

$\frac{1}{9v}(P_2{\mathrm{\Ω}}^{*}{P_2}^T-P_3{\mathrm{\Ω}}^{*}{P_3}^T)=0$

$\frac{1}{9v}(P_1{\mathrm{\Ω}}^{*}{P_1}^T-P_2{\mathrm{\Ω}}^{*}{P_2}^T)=0$

$\frac{1}{0.01v}\left(P_1{\mathrm{\Ω}}^{*}{P_2}^T\right)=0---\left(11\right)$

$\frac{1}{0.1v}\left(P_1{\mathrm{\Ω}}^{*}{P_3}^T\right)=0$

$\frac{1}{0.1v}\left(P_2{\mathrm{\Ω}}^{*}{P_3}^T\right)=0$

Here P _iThe i that is projection matrix P is capable, and v begins to be set as 1, equals P then ₃Ω ^*P ₃ ^TThrough after 5 iteration, solve Ω ^*5 parameters.Decompose Ω ^*Obtained to meet the transformation matrix T of (12) formula:

diag(1，1，1，0)＝TΩ ^*T ^T (12)

Spatial point coordinate X under the last Eulerian coordinates system _MCan obtain by (13) formula

X _M＝TX (13)

To carry out the optimization of the overall situation to the reconstructed results that obtains at last, the re-projection error is reduced to minimum, to improve the precision of measuring.Key frame has 98 width of cloth in this example, and the projection matrix of each width of cloth has 11 degree of freedom, and spatial point is had an appointment 5000, and each point has 3 degree of freedom, then need search at the higher-dimension parameter space of 98 * 11+5000 * 3=16078 parameter when optimizing.Here use based on the boundling method of adjustment of Sparse Levenberg-Marquardt algorithm and carry out global optimization based on equation (14).

$(J_g^T\left(x\right)J_g\left(x\right)+{\mathrm{\λ}}^{\′}I)\mathrm{\δx}=-J_g^T\left(x\right)g\left(x\right)---\left(14\right)$

G (x) is the expense equation (being the re-projection error) that provides, J _g(x) be the Jacobian of g (x).Here need the result that projection is rebuild and Euler rebuilds is carried out a global optimization respectively, obtain quite accurate reconstructed results at last, the re-projection error is 0.25 pixel.

Comprised target house and mark in the scene that Euler rebuilds, but the space structure that rebuild this moment there is not yardstick, just meets the situation of real world on ratio.Here utilize the known geological information of mark, on picture, choose wherein two monumented points, by two monumented points corresponding to the spatial point coordinate in the reconstructed results can calculate they in Eulerian coordinates systems apart from l ', then with the ratio l/l ' of known monumented point actual range l and l ' yardstick as the scene rebuilding structure.

Reconstruct the structure in the house of mark degree, just equaled to have obtained all geological informations in house.For example, desire is measured the height in house, only needs to select on picture the unique point of top, house and bottom, and their corresponding spatial point in reconstructed results determine thereupon that also the distance of point-to-point transmission promptly can be calculated automatically.Same method can be measured interval, story height of house width, window or the like.

From above example as can be seen, method provided by the present invention is just once rebuild the various geological informations of measurement target simultaneously to target, has avoided repeatedly layout, the duplication of labour.Survey crew does not need near measurement target, has improved the security of measuring under the situation of some circumstance complication or danger.And method provided by the present invention can be separated measurement on time and space, promptly survey crew can be taken different places and measure in the different time after take video sequence, has greatly improved the dirigibility of measuring.

The present invention is not limited to above-mentioned embodiment.For example, mark can be selected the flat board or the support of band monumented point for use.Can also replace video camera to take with high-quality camera, this just need import video sequence in the computer in real time with data line.Find the solution the antithesis absolute quadric can also selecting population parameter method (promptly using 10 parameters) for use from timing signal.

Claims (10)

1. one kind is used for the method that non-contact 3-D is measured, and comprises the following steps:

(a) in the scene of measured target place, place mark;

(b) with hand-held camera scene is taken;

(c) video sequence is imported computing machine and carried out three-dimensional reconstruction, obtain the accurate three-dimensional structure of scene;

(d) determine the yardstick of reconstruction of three-dimensional structure by mark.

(e) by the reconstructed results of mark degree target is measured.

2. the method for claim 1, wherein step (a) comprising: mark is the object with particular geometric feature, and some monumented points are promptly arranged on the mark, the distance of these monumented points is known.The position of mark in scene will guarantee can not block measurement target in the visual field of video camera, can not be too near in order to avoid shape produces distortion from video camera.

3. the method for claim 1, wherein step (b) comprising: with hand-held camera object scene is taken continuously, shooting must be moved, and motion requires speed evenly and avoid violent shake, guarantees that object scene and mark are all the time within sweep of the eye.Wherein do not need hand-held camera is demarcated, promptly under the complete condition of unknown of the intrinsic parameter of video camera, video sequence is handled.

4. the method for claim 1, wherein step (c) comprising: at first video sequence is carried out feature point extraction and tracking and chooses key frame; Next utilizes key frame and unique point track to carry out projection and rebuilds; Carry out Euler's reconstruction by self-calibrating method then; At last the result is carried out global optimization.

5. the method for claim 1, wherein step (d) comprising: geological information and actual geological information thereof by mark in the reconstructed results, calculate the corresponding actual geometry of unit length of rebuilding structure, determine the yardstick of scene rebuilding structure.

6. the method for claim 1, wherein step (e) comprising: choose 2 points that desire is measured on the picture, its position corresponding to three dimensions point in the reconstructed results is also determined thereupon, can calculate the actual geometric distance of this point-to-point transmission by the yardstick of rebuilding structure.

7. method as claimed in claim 4, wherein the KLT track algorithm has been used in the extraction of unique point and tracking, and utilizes life cycle and utmost point geometrical constraint is filtered tracking results and optimizes, and obtains the high good unique point track of fiduciary level.The GRIC discrimination standard has been used in the extraction of key frame, chooses certain parallax and have the wide picture of enough baselines as key frame from successive frame.

8. method as claimed in claim 4, wherein the projection process of reconstruction has been used the sequence type method, at first use the RANSAC method to remove the wild point of two width of cloth key frame (i.e. Cuo Wu match point), calculate the initial value of F then with 8 methods of normalization, with the MLE maximal possibility estimation F is optimized again, obtains an accurate fundamental matrix F of two width of cloth key frames.Next utilize F to carry out that coordinate system is built and structure initialization, utilize new key frame to carry out structural extended again.Each also to the structure of having rebuild be optimized when expanding structure, to reduce the re-projection error of spatial point on had been friends in the past key frame.Obtain the three-dimensional structure under the complete projective coordinate system at last.

9. method as claimed in claim 4 is wherein from calibration process elder generation normalization video camera projection matrix, then by calculating antithesis absolute quadric Ω ^*, obtain the transformation matrix of projective space to the Euler space, the three-dimensional structure under the projective coordinate system is transformed to Eulerian coordinates system, finish Euler's reconstruction.Calculate antithesis absolute quadric Ω ^*The time use 5 parameterized methods, symmetric matrix Ω 4 * 4 ^*Represent with 5 parameters, improved the accuracy of calculating.

10 methods as claimed in claim 4, wherein the method " boundling adjustment " that has powerful search capability in the higher-dimension parameter space has been adopted in global optimization, carrying out twice altogether, is respectively to rebuild back and Euler in projection to rebuild and back the result is optimized, and the re-projection error is dropped to minimum value.

Images