CN101271591A - Interactive multi-vision point three-dimensional model reconstruction method - Google Patents

Abstract

The invention relates to a three-dimensional reconstruction algorithm introducing an interactive operation for a user, belonging to the technical field of a computer multimedia. The method is as followings: the user utilizes a plurality of cameras to obtain a two-dimensional image Ii of a plurality of viewpoints of an object and extracts a contour map of the object; the computer obtains a visual shell model of a real scene according to a geometric parameter of every camera and the contour map, and turns the visual shell model into a point cloud form; a two-dimensional characteristic in the image is extracted to optimize a part of a real surface model corresponding to the characteristics; the user divides the image Ii into a region of a smoothness and the region of a subsidence; a nonuniform weighted graph is set up according to the areas selected above to find out a minimal cut of the image to obtain the optimized scene surface model. Combined with an auxiliary operation of the user, the three-dimensional reconstruction algorithm is capable of reconstructing the high-quality three-dimensional model of the object relatively fast according to the multi-view image which is shot of the actual object and the camera parameter corresponding to each image.

Description

A kind of interactive multi-vision point three-dimensional model reconstruction method

Technical field

The invention belongs to technical field of computer multimedia, particularly many viewpoints three-dimensional rebuilding method.

Background technology

Three-dimensional reconstruction promptly adopts the suitable data structure, the process that the three-dimensional geometry form on the surface of scene in the real world is carried out modeling and statement.The three-dimensional reconstruction technology, in amusement, CAD, many-sides such as historical relic's protection all will have very big application.Interim in the time of very long one section, the realization of three-dimensional reconstruction mainly can directly obtain the instrument of depth information by means of spatial digitizer etc., is referred to as scanning type three-dimensional and rebuilds.And copy human visual function, people also can from the two dimension view at a plurality of visual angles of object, obtain three-dimensional geometric information, thereby reach the purpose of three-dimensional reconstruction equally.More than the three-dimensional reconstruction realized of a plurality of two dimension view of this foundation, be referred to as many viewpoints three-dimensional reconstruction.Because many viewpoints three-dimensional reconstruction mainly depends on two dimension view, so it only needing to realize common and relatively inexpensive two dimensional image collecting device, as digital camera etc., therefore for practical application, many viewpoints three-dimensional reconstruction technology has advantage low-cost and that easily promote.Though early stage many viewpoints three-dimensional reconstruction technology is inferior to the scanning type three-dimensional reconstruction technique greatly on reconstruction quality, along with the development of computer vision technique, both gaps on reconstruction quality are more and more littler.But because computer data is handled and the restriction of storage capacity, and the restriction of Network Transmission bandwidth, current many viewpoints three-dimensional reconstruction technology still lags significantly behind scanning type three-dimensional and rebuilds on reconstruction speed.So under current development trend, practical application to an urgent demand of many viewpoints three-dimensional reconstruction technology is: it should further improve speed under the prerequisite that guarantees certain reconstruction quality.

A kind of important early stage many viewpoints reconstruction technique is visual skinning technique (Visual Hull), it can be by taking from the profile diagram of each viewpoint, obtain reconstructed results (claiming this result to be visual shell) soon, but its result's poor quality, fully truly the depression position of reflection surface.But the thought of visual skinning technique is very simple, and is easy to realize, so it still is widely used in a lot of occasions.And, using current computer hardware, visual skinning technique can reach very high reconstruction speed.

Some that propose are paid close attention to more new many viewpoints three-dimensional reconstruction technology recently, then can be considered as is a process that visual shell is revised, as the Jonathan Starck of Surrey (Surrey) university, many viewpoints three-dimensional reconstruction technology that the George Vogiatzis of Britain Camb (Cambridge) university etc. is invented separately.In these new many viewpoints three-dimensional reconstruction technology, mainly comprise obtaining of (1) image, (2) profile diagram obtains, (3) generation of visual shell, the some cloudization of (4) visual shell, (5) Feature Extraction, (6) screening of feature, (7) reconstruction of feature, (8) Graph Cut optimizes reconstruction: these new technology have had very big leap on reconstruction quality, even can be very near the level of those scanning type three-dimensional reconstruction techniques.But with regard to reconstruction speed, the performance of these technology is all more not fully up to expectations.

Summary of the invention

The objective of the invention is for overcoming the weak point of prior art, a kind of interactive multi-vision point three-dimensional model reconstruction method is proposed, this method can be from gathering from the image of a plurality of viewpoints, in conjunction with the geometric parameter of the gamma camera of each viewpoint, reconstructs the surface model of scene.And can operate speed and the quality that improves reconstruction by introducing the user.

A kind of interactive multi-vision point three-dimensional model reconstruction method that the present invention proposes is characterized in that this method may further comprise the steps:

1) obtaining of image: the user utilizes a plurality of video cameras, obtains the two-dimensional image I to a plurality of viewpoints of object _i, footnote i wherein represents the numbering of each corresponding viewpoint; And utilize Zhang Zhengyou camera marking method, obtain the geometric parameter P of each video camera _i

2) obtaining of profile diagram: the user utilize Lazy Snapping algorithm on computer screen to each viewpoint T _iThe two-dimensional image I at place _iCarry out the division of prospect and background, thereby extract the profile of prospect, generate corresponding profile diagram S _i

3) generation of visual shell: computing machine is according to each profile diagram S _iGeometric parameter P with each video camera _i, utilize the EPVH algorithm, reconstruct the visual shell V of body surface, and express with grid configuration;

4) the some cloudization of visual shell: computing machine forms some cloud V ' (set of point) with visual shell V discretize; Its concrete grammar is: uses and x, and y, many straight lines that the z three axes is parallel with the space uniform sampling, obtain a series of evenly distributed intersection points, are called the site; Get the site of the wire side intersection of these straight lines and V then, as the point among the cloud V '; Any 1 v at a cloud V ' _kIn the pairing data structure, comprise coordinate Coor (v _k), level and smooth degree Smt (v _k), sinking degree Conc (v _k), and the numbering N (v of the corresponding wire side in visual shell V _k) four data;

5) Feature Extraction: computing machine adopts DOG or Harris operator, extracts described two-dimensional image I _iTwo dimensional character set F _i{ p _I1, p _I2P _IN;

6) screening of feature: computing machine is according to the video camera geometric parameter P at three viewpoint places _i, P _I+1With P _I-1, seek and viewpoint T _iTwo the most contiguous viewpoint T _I+1With T _I-1Image I _I+1With I _I-1In, with image I _iUnique point p _IjCharacteristic of correspondence point p _{(i+1) j}With p _{(i-1) j}

7) reconstruction of feature: in two width of cloth image I _I+1With I _I-1In find image I _iLast unique point p _IjCorresponding point p _{(i+1) j}, p _{(i-1) j}, p _IjWith T _iPlace's video camera center O _iDetermined straight line p _IjO _iMeet at 2 q with visual shell V ₁ ^(ij)With q ₂ ^(ij), promptly project to p on the actual object surface _IjSome q ^(ij)At q ₁ ^(ij)With q ₂ ^(ij)Between line segment on, computing machine is to line segment q ₁ ^(ij)q ₂ ^(ij)Carry out segmentation by equal length, as node, adjacent node is connected to form weighted graph (s-t), and adopt the max-flow algorithm to try to achieve the minimal cut (minimum cut) of this weighted graph, and from this minimal cut, obtain unique point p with each Along ent _IjPairing local surfaces is represented with a this form of cloud Vp;

8) area dividing of visual shell: the user is in image I _iIn mark off the smoothness region R _Smt ^{I, j}, and depression degree region R _Conc ^{I, j}, and according to video camera geometric parameter P _i, to the each point v among the cloud V ' corresponding to described zone _kLevel and smooth degree value Smt (v _k), sinking degree value Conc (v _k) set, computing machine is revised automatically to the zone boundary again;

9) rebuild determining of scope: computing machine is gone up each point v by some cloud V ' _kLevel and smooth degree Smt (v _k) carry out up-sampling, obtain new some cloud V of denseization "; Then according to each point v _kSinking degree Conc (v _k), determine the needed point set S of next step Optimizing operation _Between, wherein the some cloud Vp of step 7) gained is this point set S _BetweenA subclass;

10) Graph Cut optimizes reconstruction: computing machine is according to the some cloud Vp that reconstructs in the step 7), the geometric parameter P of each video camera _iWith each image I _i, utilize resulting point set S in the step 9) _BetweenIn point set up cum rights (s-t) figure as node, for the pairing node of point among the Vp because it is determined the real surface that belongs to object, its the weights on limit also can be revised accordingly.Finally adopt the push-relabel algorithm, draw the minimal cut (minimum cut) of this figure, therefrom obtain out the accurate point cloud model of body surface, as the three-dimensional model of rebuilding.

Characteristics of the present invention:

The present invention is directed to the operation automatically that places one's entire reliance upon of existing new technology, and ignored the weak point of user's subjective initiative.Add the access customer operation, utilize people's priori, can only from a width of cloth two dimensional image, judge the different smooth degrees that the prospect different piece is had, and can the possible depression in finding surface.These recognition capabilities all are better than the development level of current computer vision far away on order of accuarcy and speed.

The present invention is according to the requirement of current practical application to the raising speed of many viewpoints three-dimensional reconstruction technology proposition, principle according to Graph Cut Optimization Model, in conjunction with consideration to user's actual operational capacity, this method is according to multi-view image and each image pairing camera parameters captured to actual object of input, in conjunction with user's non-productive operation, can reconstruct the three-dimensional model of object in high quality with fast speeds.

The present invention can reach following beneficial effect:

1. significantly improved accuracy rate by the appropriate users operation, and reduced the processing time;

2. the intermediate result that respectively goes on foot gained is all exportable, and can be applicable to other aspects;

3. clear in structure;

4. reconstruction quality height.

Description of drawings

Fig. 1 is a method overall procedure block diagram of the present invention;

Fig. 2 is by the two dimensional image of the gamma camera gained of 10 viewpoints in the embodiments of the invention;

Fig. 3 is the profile diagram of the two dimensional image of 3 viewpoints in the present embodiment;

The visual shell that Fig. 4 is obtained by the EPVH algorithm for present embodiment;

Fig. 5 is the visual shell of present embodiment by a cloud formal representation;

The model that Fig. 6 finally reconstructs for present embodiment.

Embodiment

A kind of interactive multi-vision point three-dimensional model reconstruction method that the present invention proposes reaches embodiment in conjunction with the accompanying drawings and is described in detail as follows:

Method flow of the present invention as shown in Figure 1, may further comprise the steps:

(1) obtaining of image: the user utilizes a plurality of video cameras, obtains the two-dimensional image I to a plurality of viewpoints of object _i, footnote i wherein represents the numbering of each corresponding viewpoint; In the embodiments of the invention by the two dimensional image of the gamma camera gained of 10 viewpoints as shown in Figure 2.

And utilize Zhang Zhengyou camera marking method, obtain the geometric parameter P of each video camera _i, show as the matrix of a 3*3 and the matrix of a 3*4;

(2) obtaining of profile diagram: the user utilize Lazy Snapping algorithm on computer screen to each viewpoint T _iThe two-dimensional image I at place _iCarry out the division of prospect and background, thereby extract the profile of prospect, generate corresponding profile diagram S _iFor the profile diagram of the two dimensional image of 3 viewpoints in the present embodiment as shown in Figure 3;

(3) generation of visual shell: computing machine is according to each profile diagram S _iGeometric parameter P with each video camera _i, utilize the EPVH algorithm, reconstruct the visual shell V of body surface, and express with grid configuration; The visual shell that present embodiment is obtained by the EPVH algorithm as shown in Figure 4;

(4) the some cloudization of visual shell: computing machine forms a some cloud V ' (set of point) with visual shell V discretize, and present embodiment by the visual shell of a cloud formal representation as shown in Figure 5; Its concrete grammar is: uses and x, and y, many straight lines that the z three axes is parallel with the space uniform sampling, obtain a series of evenly distributed intersection points, are called the site; Get the site of the wire side intersection of these straight lines and V then, as the point among the cloud V '; Any 1 v in a cloud V ' _kIn the pairing data structure, comprise coordinate Coor (v _k), level and smooth degree Smt (v _k), sinking degree Conc (v _k), and the numbering N (v of the corresponding wire side in visual shell V _k) four data;

(5) Feature Extraction: adopt DOG or Harris operator, extract described two-dimensional image I _iTwo dimensional character set F _i{ p _I1, p _I2P _IN;

(6) screening of feature: computing machine is according to the video camera geometric parameter P at three viewpoint places _i, P _I+1With P _I-1, seek and viewpoint T _iTwo the most contiguous viewpoint T _I+1With T _I-1Image I _I+1With I _I-1In, with image I _iUnique point p _IjCharacteristic of correspondence point p _{(i+1) j}With p _{(i-1) j}

(7) reconstruction of feature: in two width of cloth image I _I+1With I _I-1In find image I _iLast unique point p _IjCorresponding point p _{(i+1) j}, p _{(i-1) j}, p _IjWith T _iPlace's video camera center O _iDetermined straight line p _IjO _iMeet at 2 q with visual shell V ₁ ^(ij)With q ₂ ^(ij), promptly project to p on the actual object surface _IjSome q ^(ij)At q ₁ ^(ij)With q ₂ ^(ij)Between line segment on, computing machine is to line segment q ₁ ^(ij)q ₂ ^(ij)Carry out segmentation by equal length, as node, adjacent node is connected to form weighted graph (s-t), and adopt the max-flow algorithm to try to achieve the minimal cut (minimum cut) of this weighted graph, and from this minimal cut, obtain unique point p with each Along ent _IjPairing local surfaces is represented with a this form of cloud Vp; Arbitrfary point vp among the Vp _kBe confirmed as belonging to body surface.

(8) area dividing of visual shell: the user is in image I _iIn mark off the smoothness region R _Smt ^{I, j}, and depression degree region R _Conc ^{I, j}, and according to video camera geometric parameter P _i, to the each point v among the cloud V ' corresponding to described zone _kLevel and smooth degree value Smt (v _k), sinking degree value Conc (v _k) set, computing machine is revised automatically to the zone boundary again; Specifically comprise:

User's range estimation is selected surface smoothness Ra and is reached the above roughness region R of 1cm-5cm _Smt ^{I, j}(as the fold place of clothing among the figure), and specify the smoothness value in selected zone for computing machine, the order that the span of this smoothness value rises successively by level and smooth degree is 0,1,2 ..., 8,9,10 expressions, the zone that all are not selected, its smoothness all is set at 5;

User's range estimation is selected the degree of depth that is absorbed in the surface and is reached the above sunk area R of 5cm _Conc ^{I, j}, and specify the depression degree value in selected zone for computing machine, the order that the span of this depression degree value rises successively by sinking degree is 5,4,3...0;

Computing machine is revised automatically to the zone boundary then:

At first will be according to the two dimensional image characteristic, to the region R that the user divided _Smt ^{I, j}And R _Conc ^{I, j}The border be optimized; Concrete optimized Algorithm is selected SNAKE or based on the method for Graph Cut framework;

A kind of specific implementation of fairly simple employing SNAKE algorithm is:

Suppose that zone boundary that the user provides is by set of pixels { Px ₁..., Px _LComposition, wherein Px _i=(x _i, y _i), i={1 ..., L}.Be positioned at Px for each _iSome Px in the neighborhood _i', the energy term that is calculated as follows:

E _i(Px _i′)＝E _int(Px _i′)+μ ₁×E _ext(Px _i′)

Then with Px _iMove to the Px that makes the energy term minimum _i' the position on.

E _Int(Px _i') the expression internal energy, its computing formula is:

$E_{\mathrm{int}}\left({{\mathrm{Px}}_i}^{\′}\right)=\frac{1}{I}{\left|\right|{{\mathrm{Px}}_i}^{\′}-\frac{1}{2\cos (2\mathrm{\π}/L)}({\mathrm{Px}}_{i+1}-{\mathrm{Px}}_i)\left|\right|}^2+{\mathrm{\μ}}_2\×{n_i}^T({\mathrm{Px}}_i-{{\mathrm{Px}}_i}^{\′})$ Wherein, n _iExpression Px _iOutside normal vector in initial boundary, its tangent line vector t by will be herein _iRevolve and turn 90 degrees and get t _iComputing formula be:

$t_i=\frac{{\mathrm{Px}}_i-{\mathrm{Px}}_{i-1}}{\left|\right|{\mathrm{Px}}_i-{\mathrm{Px}}_{i-1}\left|\right|}+\frac{{\mathrm{Px}}_{i+1}-{\mathrm{Px}}_i}{\left|\right|{\mathrm{Px}}_{i+1}-{\mathrm{Px}}_i\left|\right|}$

E _Ext(Px _i') the expression external energy, its computing formula is:

$E_{\mathrm{int}}\left({{\mathrm{Px}}_i}^{\′}\right)=\mathrm{Gs}\left({{\mathrm{Px}}_i}^{\′}\right)-{n_i}^T\▿\mathrm{Gs}\left({{\mathrm{Px}}_i}^{\′}\right)$

Wherein, Gs (Px _i') expression gets the gray-scale value at this pixel place,

The gradient vector of representing the gray scale at this pixel place.

μ in the formula ₁μ ₂μ ₃The expression adjustable parameter generally can all be taken as 0.5.

After the zone boundary is optimized end and is determined, will be at first according to viewpoint T _iThe camera geometric parameter P at place _i, among the visual shell V ' of judging point cloud form, can project to the region R that marks off above _IjIn point { v ₁..., v _N; v ₁..., v _NIn from T _iThe nearest some v of camera at place _mBe determined and region R _Smt ^{I, j}Or R _Conc ^{I, j}Corresponding; Then to a v _mSmoothness value Smt (v _m), and depression degree value Conc (v _m) revise by the value of its corresponding region.

(9) rebuild determining of scope: computing machine is gone up each point v by some cloud V ' _kLevel and smooth degree Smt (v _k) carry out up-sampling, obtain new some cloud V of denseization "; Then according to each point v _kSinking degree Conc (v _k), determine the needed point set S of next step Optimizing operation _Between, wherein the some cloud Vp of step 7) gained is this point set S _BetweenA subclass;

This step specifically comprises:

For any 1 v among the V ' _k, be taken at its pairing wire side N (v _k) interior 8 neighborhoods point { v _K1..., v _K8The average Smt of smoothness value _Avg, determine and will in this 8 neighborhood institute localized area, use bilinear interpolation, increase that quantity be f (Smt _Avg) the point; The depression degree value Conc (v of newly-increased point _k) calculate by interpolation, obtain the V as a result behind the up-sampling ";

Then with V " the middle pairing surface of each point, as outside surface V _Ex(being the upper bound of reconstructed results), the S set that is designated as a little in the space that it comprised _ExOuter surface S then _ExIn any 1 v _k, be the centre of sphere with it, be the ball of r as radius, the size of r is its depression degree value Conc (v _k) function g (Conc (v _k)), if in this ball, have V _ExOutside point, v then _kBe included into S set _Between

In concrete enforcement, function f can be chosen as this simple form of f (x)=α * x, and wherein parameter alpha can be selected by the user.Function g can be chosen as this simple form of g (x)=β * x, and wherein parameter beta can be selected by the user;

(10) Graph Cut optimizes reconstruction: computing machine is according to the some cloud Vp that reconstructs in the step 7), the geometric parameter P of each video camera _iWith each image I _i, utilize resulting point set S in the step 9) _BetweenIn point as ordinary node n _kSet up cum rights (s-t) figure, except these ordinary nodes n _kAlso will comprise two special joint source and sink among this figure outward.Each ordinary node all connects to source, and the weights on this limit are taken as 0.239; S _BetweenIn the ordinary node of close visual case surface, also connect to sink simultaneously, the weights on this limit are taken as 9999.

For any two adjacent ordinary node n _kWith n _l, connect limit e each other _KlWeight w _KlBe set to node n _kWith n _lThe colour consistency C (n ') (Photo Consistency) of mid point n '.

The concrete computing method of C (n ') are: the image collection { Iv that is not blocked at n ' ₁..., Iv _MIn (M≤10), appoint and get two width of cloth image I v _jWith Iv _k, in this two width of cloth image, the two-dimensional coordinate of subpoint is respectively prj with n ' spot projection _jWith prj _kThen respectively with prj _jWith prj _kBe the center, obtain all interior pixel { Px of scope of 11*11 _J1..., Px _J121And { Px _K1..., Px _K121, calculate the NCC value (Normalized Cross Correlation) of these two groups of pixel set, its expression formula is as follows:

$\mathrm{NCC}(j,k)=\frac{\underset{i=1}{\overset{121}{\mathrm{\Σ}}}P{x}_{\mathrm{ji}}^2\underset{i=1}{\overset{121}{\mathrm{\Σ}}}{\mathrm{Px}}_{\mathrm{ki}}^2}{{(\underset{i=1}{\overset{121}{\mathrm{\Σ}}}{\mathrm{Px}}_{\mathrm{ji}}\×{\mathrm{Px}}_{\mathrm{ki}})}^2}$

Then,, its summation is averaged, obtains C (n ') for by each a plurality of NCC value in the set to the image gained:

$C\left(n^{\′}\right)=\frac{\underset{j=1...M-1,k=j...M}{\mathrm{\Σ}}\mathrm{NCC}(j,k)}{C_M^2}$

If above-mentioned n _k, n _lInstitute's corresponding point all belong to Vp simultaneously, then connect limit e each other _KlWeight w _KlNeed be modified to w _Kl=9999 * C (n ').And if n _k, n _lIn have one to belong to Vp, and another person does not belong to Vp, then w _KlNeed be modified to w _Kl=0.001 * C (n ').

Finally adopt the push-relabel algorithm, draw the minimal cut (minimum cut) of this figure, therefrom obtain out the accurate point cloud model of body surface, as the three-dimensional model of rebuilding.The model that present embodiment finally reconstructs as shown in Figure 6.

(11) according to customer requirements described point cloud model is converted into grid configuration, and carries out texture mapping and handle.

Claims (4)

1, a kind of interactive multi-vision point three-dimensional model reconstruction method is characterized in that, this method may further comprise the steps:

1) obtaining of image: the user utilizes a plurality of video cameras, obtains the two-dimensional image I to a plurality of viewpoints of object _i, footnote i wherein represents the numbering of each corresponding viewpoint; And utilize Zhang Zhengyou camera marking method, obtain the geometric parameter P of each video camera _i

2) obtaining of profile diagram: the user utilize Lazy Snapping algorithm on computer screen to each viewpoint T _iThe two-dimensional image I at place _iCarry out the division of prospect and background, thereby extract the profile of prospect, generate corresponding profile diagram S _i

3) generation of visual shell: computing machine is according to each profile diagram S _iGeometric parameter P with each video camera _i, utilize the EPVH algorithm, reconstruct the visual shell V of body surface, and express with grid configuration;

4) the some cloudization of visual shell: computing machine forms some cloud V ' (set of point) with visual shell V discretize; Its concrete grammar is: uses and x, and y, many straight lines that the z three axes is parallel with the space uniform sampling, obtain a series of evenly distributed intersection points, are called the site; Get the site of the wire side intersection of these straight lines and V then, as the point among the cloud V '; Any 1 v in a cloud V ' _kIn the pairing data structure, comprise coordinate Coor (v _k), level and smooth degree Smt (v _k), sinking degree Conc (v _k), and the numbering N (v of the corresponding wire side in visual shell V _k) four data;

5) Feature Extraction: computing machine adopts DOG or Harris operator, extracts described two-dimensional image I _iTwo dimensional character set F _i{ p _I1, p _I2P _IN;

6) screening of feature: computing machine is according to the video camera geometric parameter P at three viewpoint places _i, P _I+1With P _I-1, seek and viewpoint T _iTwo the most contiguous viewpoint T _I+1With T _I-1Image I _I+1With I _I-1In, with image I _iUnique point p _IjCharacteristic of correspondence point p _{(i+1) j}With p _{(i-1) j}

7) reconstruction of feature: in two width of cloth image I _I+1With I _I-1In find image I _iLast unique point p _IjCorresponding point p _{(i+1) j}, p _{(i-1) j}, p _IjWith T _iPlace's video camera center O _iDetermined straight line p _IjO _iMeet at 2 q with visual shell V ₁ ^(ij)With q ₂ ^(ij), promptly project to p on the actual object surface _IjSome q ^(ij)At q ₁ ^(ij)With q ₂ ^(ij)Between line segment on, computing machine is to line segment q ₁ ^(ij)q ₂ ^(ij)Carry out segmentation by equal length, as node, adjacent node is connected to form weighted graph (s-t), and adopt the max-flow algorithm to try to achieve the minimal cut (minimum cut) of this weighted graph, and from this minimal cut, obtain unique point p with each Along ent _IjPairing local surfaces is represented with a this form of cloud Vp;

8) area dividing of visual shell: the user is in image I _iIn mark off the smoothness region R _Smt ^{I, j}, and depression degree region R _Conc ^{I, j}, and according to video camera geometric parameter P _i, to the each point v among the cloud V ' corresponding to described zone _kLevel and smooth degree value Smt (v _k), sinking degree value Conc (v _k) set, computing machine is revised automatically to the zone boundary again;

9) rebuild determining of scope: computing machine is gone up each point v by some cloud V ' _kLevel and smooth degree Smt (v _k) carry out up-sampling, obtain new some cloud V of denseization "; Then according to each point v _kSinking degree Conc (v _k), determine the needed point set S of next step Optimizing operation _Between, wherein the some cloud Vp of step 7) gained is this point set S _BetweenA subclass;

10) Graph Cut optimizes reconstruction: computing machine is according to the some cloud Vp that reconstructs in the step 7), the geometric parameter P of each video camera _iWith each image I _i, utilize resulting point set S in the step 9) _BetweenIn point set up cum rights (s-t) figure as node, for the pairing node of point among the Vp because it is determined the real surface that belongs to object, its the weights on limit also can be revised accordingly.Finally adopt the push-relabel algorithm, draw the minimal cut (minimum cut) of this figure, therefrom obtain out the accurate point cloud model of body surface, as the three-dimensional model of rebuilding.

2, the method for claim 1 is characterized in that, this method also comprises:

11) according to the customer requirements computing machine described point cloud model is converted into grid configuration, and carries out texture mapping and handle, as the three-dimensional model of last reconstruction.

3, the method for claim 1 is characterized in that, level and smooth degree value Smt (v in the described step 8) _k) and sinking degree value Conc (v _k) method set is:

User's range estimation is selected surface smoothness Ra and is reached the above roughness region R of 1cm-5cm _Smt ^{I, j}, and specify the smoothness value in selected zone for computing machine, and the order that the span of this smoothness value rises successively by level and smooth degree is 0,1,2 ..., 8,9,10 expressions, the zone that all are not selected, its smoothness all is set at 5;

User's range estimation is selected the degree of depth that is absorbed in the surface and is reached the above sunk area R of 5cm _Conc ^{I, j}, and specify the depression degree value in selected zone for computing machine, the order that the span of this depression degree value rises successively by sinking degree is 5,4,3...0;

Described computing machine is modified to automatically to the zone boundary:

(1) will be according to the characteristics of image (as texture, color etc.) of two dimension, to the region R that the user divided _Smt ^{I, j}And R _Conc ^{I, j}The border be optimized; Concrete optimized Algorithm is selected SNAKE or based on the method for Graph Cut framework;

(2) after the zone boundary is optimized end and is determined, will be at first according to viewpoint T _iThe video camera geometric parameter P at place _i, among the visual shell V ' of judging point cloud form, can project to described region R _Smt ^{I, j}And R _Conc ^{I, j}Point { v ₁..., v _N; v ₁..., v _NIn from T _iThe nearest some v of video camera at place _mBe determined and region R _Smt ^{I, j}Or R _Conc ^{I, j}Corresponding; Then to a v _mSmoothness value Smt (v _m), and depression degree value Conc (v _m) revise by the value of its corresponding region.

4, the method for claim 1 is characterized in that, the up-sampling in the described step 9) is operating as:

For any 1 v among the cloud V ' _k, be taken at its pairing wire side N (v _k) interior 8 neighborhoods point { v _K1..., v _K8The average Smt of smoothness value _Avg, determine and will in this 8 neighborhood institute localized area, use bilinear interpolation, increase that quantity be f (Smt _Avg) the point; The depression degree value Conc (v of newly-increased point _k) calculate by interpolation, obtain new some cloud V behind the up-sampling ";

Described point set S _BetweenBe defined as:

With new some cloud V " the middle pairing surface of each point, as outside surface V _Ex(being the upper bound of reconstructed results), the S set that is designated as a little in the space that it comprised _ExOuter surface S then _ExIn any 1 v _k, be the centre of sphere with it, be the ball of r as radius, the size of r is its depression degree value Conc (v _k) function g (Conc (v _k)), if in this ball, have V _ExOutside point, v then _kBe included into S set _Between

Images