CN104820991A - Multi-soft-constraint stereo matching method based on cost matrix - Google Patents

Abstract

The invention discloses a multi-soft-constraint stereo matching method based on a cost matrix. The method comprises the following steps: calculating a three-dimensional matching cost matrix at first, then carrying out multiple-dimension downsampling, and forming a cost matrix pyramid; carrying out the corresponding multiple-dimension downsampling of an image, forming an image pyramid, and carrying out segmentation of all image layers; carrying out cost accumulation of self-adaption weight layer by layer and cost accumulation under the voting-type segmentation constraint secondly, and achieving the cost accumulation under the multiple-dimension constraint through transmitting the cost accumulation results of the upper layer to the lower cost matrix; and achieving stereo matching through combining a reliable point cost diffusion method in a bottom layer cost matrix. The method improves the stability and reliability of matching results, solves the matching problems of weak texture and repeated texture regions and parallax error discontinuous places, and can be used for improving modeling, related to the engineering application of stereo matching, based on images.

Description

A kind of multiple soft-constraint solid matching method based on cost matrix

Technical field

The invention belongs to computation vision and photogrammetric technology field, relate to a kind of solid matching method based on cost matrix, especially relate to a kind of solid matching method carrying out multiple soft-constraint based on cost matrix.

Background technology

Stereo matching is a basis and the key issue in computation vision and photogrammetric field.The concept of Stereo matching proposes as far back as photogrammetric field, for solving the problem of digital aerial surveying robotization mapping.Stereo matching is also the key issue of computer vision field, has influence on the vision system modeling to people, robot navigation and operation, mixes outdoor scene action etc. in the image that 3D model is set up and generated in a computer.

Stereo matching is an ill-conditioning problem, is especially difficult to comprehensive solution for shortage texture with repetition texture region and parallax discontinuous error hiding problem always.For obtaining better matching result, need rationally to include the constraint of more priori conditions in.Matching algorithm generally adopts local based on the coupling of window, utilizes image local information to improve Window match in matching process, utilizes Pyramid technology matching strategy etc. to improve stability and accuracy simultaneously.But these methods often need larger match window, cause the details being difficult to retain image.Also there are some algorithms to include Image Segmentation constraint in, but have very strong dependence for image segmentation result.In a word, traditional solid matching method is often direct reduces disparity search scope by various constraint condition, therefore very strong for priori conditions dependence, easily produces error hiding.

Occur that some pass through to improve the algorithm that cost agglomeration approach realizes coupling in recent years, self-adaptation cost agglomeration approach is such as adopted to improve local window matching algorithm, adopt multiple dimensioned cost agglomeration approach to solve the matching problem repeating texture, adopt half overall cost agglomeration approach to solve the discontinuous matching problem of parallax.The common ground of these algorithms is all realize respective constraint in cost matrix, but due to the constraint included in abundant not, be therefore difficult to solve error hiding problem more all sidedly.

Summary of the invention

The present invention mainly solves the existing problem excessively strong for priori conditions constraint dependence of prior art; Provide a kind of multiple soft-constraint cost agglomeration approach based on cost matrix simultaneously including multiple prior-constrained condition in and complete Stereo matching, effectively can solve the error hiding problem lacking texture and repeat texture region and edge.

The technical solution adopted in the present invention is: a kind of multiple soft-constraint solid matching method based on cost matrix, is characterized in that, comprise the following steps:

Step 1: concentrate at raw video and select a wherein width core line image as reference images, and to utilize AD-Census-Sobel as similarity measure to wherein each pixel, calculate the Matching power flow of each candidate disparity values, generate a three-dimensional Matching power flow matrix;

Step 2: carry out multiple dimensioned down-sampled to the cost matrix obtained in step 1, forms cost matrix pyramid;

Step 3: also carry out corresponding multiple dimensioned down-sampled to reference images, forms image pyramid;

Step 4; Respectively Iamge Segmentation is carried out to each layer image of the image pyramid obtained in step 3;

Step 5: to the cost matrix pyramid obtained in step 2, gathers the cost that each layer cost matrix carries out adaptive weighting from top layer cost matrix;

Step 6: according to the image segmentation result described in step 4, gathers the cost that the cost matrix after process in step 5 carries out under " ballot formula " segmentation constraint further;

Step 7: this layer of cost is gathered result and passes to lower floor's cost matrix;

Step 8: repeat the operation of step 5 to step 7 to pyramid cost matrix from top to bottom, until original cost matrix, finally carries out step 5 to original cost matrix and step 6 operates;

Step 9: to the cost matrix obtained after process in step 8, Least-cost is got to each pixel and mates degree of confidence and be greater than the parallax of predetermined threshold as final parallax, then in cost matrix, cost diffusion being carried out to non-match point as controlling with the point obtaining parallax, finally adopting " winner overwhelm the market " method to generate disparity map;

Step 10: concentrate at raw video and select another image as reference images, repeat above-mentioned steps 1 to step 9, generate another width disparity map, detect Mismatching point by the disparity consistency contrasting two width disparity map respective pixel, obtain removing Mismatching point disparity map; Parallax interpolation is carried out to removal Mismatching point disparity map, fills up the black hole that error hiding causes, generate complete disparity map; Utilize complete disparity map and image elements of exterior orientation, according to forward intersection principle compute depth figure and digital surface model.

As preferably, in step 1,

Described AD-Census-Sobel similarity measure is defined as follows:

Cost (p, d)=exp (α * AD (p, d)+β * Census (p, d)+γ * Sobel (p, d)) (formula one);

Wherein, p represents the pixel coordinate p (x in reference images, y), d is parallax value, α, β and γ is weight coefficient, General Requirements is positive number and alpha+beta+γ=1, AD (p, d) refer to the gray scale difference absolute value of reference images pixel p and the conjugation pixel p+d on coupling image thereof, Census (p, d) refers to the Census match measure value of reference images pixel p and the conjugation pixel p+d on coupling image thereof, Sobel (p, d) refers to the Sobel match measure value of reference images pixel p and the conjugation pixel p+d on coupling image thereof;

Calculate Matching power flow value corresponding to different parallax d according to formula one pair of reference images by pixel, form a three-dimensional Matching power flow matrix represented by image horizontal ordinate W, ordinate H and parallax value D.

As preferably, the pyramidal generation method of cost matrix described in step 2 is, keep parallax value D direction constant to the three-dimensional Matching power flow matrix obtained in step 1 and adopt gaussian pyramid to carry out down-sampled at horizontal ordinate W and ordinate H direction, multiple dimensioned cost matrix can be generated through repeatedly down-sampled in this way.

As preferably, the multiple dimensioned down-sampled method described in step 3 is, adopts gaussian pyramid to carry out down-sampled, can generate image pyramid in this way through repeatedly down-sampled to raw video at horizontal ordinate W and ordinate H direction.

As preferably, first image partition method described in step 4 for adopt existing image partition method, then after Iamge Segmentation completes, block cluster is carried out, and the classification that its element number is greater than predetermined threshold is split, again until the element number of each classification is not more than predetermined threshold.

As preferably, the cost agglomeration approach of the adaptive weighting described in step 5 is, according to Euclidean distance dist (p, q) and the grey value difference of the pixel q in window certain centered by each pixel p and this point | I _p-I _q| determine the contribution margin of pixel q for pixel p cost, all pixels of accumulation window like this for pixel p cost contribution and obtain the final cost value cost (p, d) of p, concrete computing method are as follows:

$\cos t(p,d)=\underset{q\∈W_p}{\mathrm{\Σ}}\cos t(q,d)*\mathrm{weight}(q,d)$ (formula two);

weight(q,d)＝exp(α*dis(p,q)+β*|I _p-I _q|)

Wherein, weight (q, d) represents the weighted value carrying out neighborhood territory pixel when cost is gathered, the parameter related to when α, β are and calculate this weighted value, respectively corresponding Euclidean distance dist (p, q) and grey value difference | I _p-I _q| weight coefficient, General Requirements is positive number and alpha+beta=1.

As preferably, the specific implementation of step 6 comprises following sub-step:

Step 6.1: to each Iamge Segmentation block, total the cost that in this Iamge Segmentation block that adds up respectively, each parallax value of all pixels is corresponding, obtains in whole Iamge Segmentation block, the parallax that Matching power flow changes with parallax value-cost histogram;

Step 6.2: according to histogram calculation Matching power flow average, the difference of each parallax value and average and the average of these differences, and be negative to its difference and the parallax value that absolute value is greater than difference average is composed with punishment, namely compose with a negative cost value, according to said method obtain one and correspond to the histogrammic parallax of parallax-cost described in step 6.1-cost punishment histogram;

To the parallax value of each pixel, step 6.3: apply the punishment of above-mentioned cost by pixel to Iamge Segmentation block, namely according to the parallax described in step 6.2-cost punishment histogram, adds that corresponding cost is punished;

Wherein the computing method of cost punishment are as follows:

$p\left(d\right)=\left\{\begin{array}{cc}0,& c_d>\stackrel{\&OverBar;}{c_d}\\ \mathrm{\&alpha;}*(c_d-\stackrel{\&OverBar;}{c_d})/\stackrel{\&OverBar;}{|c_d-\stackrel{\&OverBar;}{c_d}|},& c_d<\stackrel{\&OverBar;}{c_d}\end{array}\right.$ (formula three);

Wherein, c _dfor block corresponds to the cost of parallax d, α is weight coefficient.

As preferably, this layer of cost gathered result pass to lower floor's cost matrix described in step 7, its specific implementation comprises following sub-step:

Step 7.1: respectively one dimension least square convolution is carried out to the cost value of each pixel in disparity range; Note, the object of convolution makes upper strata Matching power flow more level and smooth with parallax change here, therefore also can adopt minimum and convolution or other convolution methods;

Step 7.2: according to pyramid pixel corresponding relation, the cost value of this layer of each pixel is added to the cost value of the corresponding parallax of lower floor's respective pixel.

As preferably, the ratio of the coupling degree of confidence described in step 9 and Matching power flow sub-minimum and minimum value, could obtain parallax first for the pixel only having coupling degree of confidence to be greater than predetermined threshold; As follows as controlling to carry out the method for cost diffusion to non-match point in cost matrix according to the point obtaining parallax: to mark match point and non-match point, then certain penalty value is deducted, shown in formula specific as follows to the parallax cost that the non-matched pixel in each match point certain window neighborhood corresponds to this match point:

$\cos t(q,d)=\left\{\begin{array}{c}\cos t(q,d)-P,D\left(p\right)=d\\ \cos t(q,d),D\left(p\right)\&NotEqual;d\end{array}\right.,q\&Element;N_p,D\left(q\right)=\mathrm{NULL}$ (formula wantonly);

Wherein, p represents match point, and q represents non-match point, and P represents penalty value, and D (p) represents the value of initial parallax figure D at p point place.

Tool of the present invention has the following advantages: in Stereo matching process, consider local grain structure and half-tone information, enhances same place ga s safety degree; Incorporate multiple dimensioned constraint, enhance the matching capacity of repetition texture region; Incorporate segmentation constraint, enhance the matching capacity of weak texture region; Carry out cost diffusion according to elementary reliable matching result, enhance coupling stability; In cost matrix, complete above institute Constrained, often kind of constraint is all the soft-constraint that directly can not determine final matching results, further enhancing stability and the reliability of matching result.

Accompanying drawing explanation

Fig. 1: be the overview flow chart of the embodiment of the present invention;

Fig. 2: multiple dimensioned down-sampled method schematic diagram is carried out to cost matrix for the embodiment of the present invention;

Fig. 3: be " ballot formula " based on cost matrix segmentation constraint cost agglomeration approach schematic diagram of the embodiment of the present invention.

Embodiment

Understand for the ease of those of ordinary skill in the art and implement the present invention, below in conjunction with drawings and Examples, the present invention is described in further detail, should be appreciated that exemplifying embodiment described herein is only for instruction and explanation of the present invention, is not intended to limit the present invention.

Ask for an interview Fig. 1, a kind of multiple soft-constraint solid matching method based on cost matrix provided by the invention, comprises the following steps:

Step 1: concentrate at raw video and select a wherein width core line image as reference images, and to utilize AD-Census-Sobel as similarity measure to wherein each pixel, calculate the Matching power flow of each candidate disparity values, generate a three-dimensional Matching power flow matrix;

AD-Census-Sobel similarity measure is defined as follows:

Cost (p, d)=exp (α * AD (p, d)+β * Census (p, d)+γ * Sobel (p, d)) (formula one);

Wherein, p represents the pixel coordinate p (x in reference images, y), d is parallax value, α, β and γ is weight coefficient, General Requirements is positive number and alpha+beta+γ=1, AD (p, d) refer to the gray scale difference absolute value of reference images pixel p and the conjugation pixel p+d on coupling image thereof, Census (p, d) refers to the Census match measure value of reference images pixel p and the conjugation pixel p+d on coupling image thereof, Sobel (p, d) refers to the Sobel match measure value of reference images pixel p and the conjugation pixel p+d on coupling image thereof;

Calculate Matching power flow value corresponding to different parallax d according to formula one pair of reference images by pixel, form a three-dimensional Matching power flow matrix represented by image horizontal ordinate W, ordinate H and parallax value D, size is W*H*D.

Step 2: carry out multiple dimensioned down-sampled to the cost matrix obtained in step 1, forms cost matrix pyramid;

Ask for an interview Fig. 2, the pyramidal generation method of cost matrix is, keep parallax value D direction constant to the three-dimensional Matching power flow matrix obtained in step 1 and adopt gaussian pyramid to carry out down-sampled at horizontal ordinate W and ordinate H direction, multiple dimensioned cost matrix can be generated through repeatedly down-sampled in this way.

Step 3: also carry out corresponding multiple dimensioned down-sampled to reference images, forms image pyramid; Wherein multiple dimensioned down-sampled method is, adopts gaussian pyramid to carry out down-sampled, can generate image pyramid in this way through repeatedly down-sampled to raw video at horizontal ordinate W and ordinate H direction.

Step 4; Respectively Iamge Segmentation is carried out to each layer image of the image pyramid obtained in step 3; First image partition method for adopt existing image partition method, then after Iamge Segmentation completes, block cluster is carried out, and the classification that its element number is greater than predetermined threshold is split, again until the element number of each classification is not more than predetermined threshold.

Step 5: to the cost matrix pyramid obtained in step 2, gathers the cost that each layer cost matrix carries out adaptive weighting from top layer cost matrix; The cost agglomeration approach of adaptive weighting is, according to Euclidean distance dist (p, q) and the grey value difference of the pixel q in window certain centered by each pixel p and this point | I _p-I _q| determine the contribution margin of pixel q for pixel p cost, all pixels of accumulation window like this for pixel p cost contribution and obtain the final cost value cost (p, d) of p, concrete computing method are as follows:

$\cos t(p,d)=\underset{q\&Element;W_p}{\mathrm{\&Sigma;}}\cos t(q,d)*\mathrm{weight}(q,d)$ (formula two);

weight(q,d)＝exp(α*dis(p,q)+β*|I _p-I _q|)

Wherein, weight (q, d) represents the weighted value carrying out neighborhood territory pixel when cost is gathered, the parameter related to when α, β are and calculate this weighted value, respectively corresponding Euclidean distance dist (p, q) and grey value difference | I _p-I _q| weight coefficient, General Requirements is positive number and alpha+beta=1.

Step 6: according to the image segmentation result described in step 4, gathers the cost that the cost matrix after process in step 5 carries out under " ballot formula " segmentation constraint further;

Ask for an interview Fig. 3, specific implementation comprises following sub-step:

Step 6.1: to each Iamge Segmentation block, total the cost that in this Iamge Segmentation block that adds up respectively, each parallax value of all pixels is corresponding, obtains in whole Iamge Segmentation block, the parallax that Matching power flow changes with parallax value-cost histogram;

Step 6.2: according to histogram calculation Matching power flow average, the difference of each parallax value and average and the average of these differences, and be negative to its difference and the parallax value that absolute value is greater than difference average is composed with punishment, namely compose with a negative cost value, according to said method obtain one and correspond to the histogrammic parallax of parallax-cost described in step 6.1-cost punishment histogram;

To the parallax value of each pixel, step 6.3: apply the punishment of above-mentioned cost by pixel to Iamge Segmentation block, namely according to the parallax described in step 6.2-cost punishment histogram, adds that corresponding cost is punished;

Wherein the computing method of cost punishment are as follows:

$p\left(d\right)=\left\{\begin{array}{cc}0,& c_d>\stackrel{\&OverBar;}{c_d}\\ \mathrm{\&alpha;}*(c_d-\stackrel{\&OverBar;}{c_d})/\stackrel{\&OverBar;}{|c_d-\stackrel{\&OverBar;}{c_d}|},& c_d<\stackrel{\&OverBar;}{c_d}\end{array}\right.$ (formula three);

Wherein, c _dfor block corresponds to the cost of parallax d, α is weight coefficient.

Step 7: this layer of cost is gathered result and passes to lower floor's cost matrix; Its specific implementation comprises following sub-step:

Step 7.1: respectively according to formula, one dimension least square convolution is wantonly carried out to the cost value of each pixel in disparity range;

${\cos t}^{\&prime;}\left(d\right)=\underset{d}{\min }(\cos t\left(d\right)+{(d-d^{\&prime;})}^2)$ (formula wantonly);

Wherein, cost'(d) represent cost value corresponding to parallax d after convolution, d' is other parallax value within the scope of disparity search;

Step 7.2: according to pyramid pixel corresponding relation, the cost value of this layer of each pixel is added to the cost value of the corresponding parallax of lower floor's respective pixel.

Step 8: repeat the operation of step 5 to step 7 to pyramid cost matrix from top to bottom, until original cost matrix, finally carries out step 5 to original cost matrix and step 6 operates;

Step 9: to the cost matrix that obtains after process in step 8, gets Least-cost by formula 5 to each pixel and coupling degree of confidence is greater than the parallax of predetermined threshold as final parallax, coupling degree of confidence T _confthe i.e. ratio of Matching power flow sub-minimum and minimum value, could obtain parallax first for the pixel only having coupling degree of confidence to be greater than predetermined threshold.

(formula 5);

Then match point and non-match point is marked, right back-pushed-type land deducts certain penalty value P to the parallax cost that the non-matched pixel in each match point 8 neighborhood corresponds to this match point, thus matching result is diffused into the non-match point of neighborhood with the form of the punishment of cost.

$\cos t(q,d)=\left\{\begin{array}{c}\cos t(q,d)-P,D\left(p\right)=d\\ \cos t(q,d),D\left(p\right)\&NotEqual;d\end{array}\right.,q\&Element;N_p,D\left(q\right)=\mathrm{NULL}$ (formula land);

Wherein, p represents match point, and q represents non-match point, and P represents penalty value, and D (p) represents the value of initial parallax figure D at p point place.

" winner overwhelm the market " method shown in formula seven is finally adopted to generate disparity map.

$D\left(p\right)=\arg \underset{d}{\min }\left(\cos t(p,d)\right)$ (formula seven);

Step 10: concentrate at raw video and select another image as reference images, repeat above-mentioned steps 1 to step 9, generate another width disparity map, detect Mismatching point by the disparity consistency contrasting two width disparity map respective pixel, obtain removing Mismatching point disparity map; Parallax interpolation is carried out to removal Mismatching point disparity map, fills up the black hole that error hiding causes, generate complete disparity map; Utilize complete disparity map and image elements of exterior orientation, according to forward intersection principle compute depth figure and digital surface model.

The mathematical measure of horizontal parallax consistency detection is:

LRC (p)=| D _l-R(p)-D _r-L(p+D _l-R(p)) | (formula eight);

Wherein, D _l-Rexpression is the matching result with reference to image with left figure, D _r-Lfor taking right figure as the matching result with reference to image, the match point that LRC (p) is greater than certain threshold value is considered to illegal point.

Carry out occlusion detection and parallax interpolation to above-mentioned disparity map, the existing correct match point parallax of utilization fills up the black hole that its neighborhood error hiding causes, and generates complete disparity map;

Wherein, occlusion detection method is: illegally put p in reference images, and coupling image is searched for along horizontal epipolar line direction, if there is a parallax d to make D (p+d)-d<T _d, then this point is error hiding, otherwise for blocking, shown in nine:

$\mathrm{flag}\left(p\right)=\left\{\begin{array}{c}\mathrm{occluded},D(p+d)-d<T_d\\ \mathrm{mismatch},D(p+d)-d<T_d\end{array}\right.$ (formula nine);

Parallax interpolating method is:

$D\left(p\right)=\left\{\begin{array}{c}{D}_{\mathrm{low}}(q\&Element;N_p),\mathrm{flag}\left(p\right)=\mathrm{occluded}\\ D_{\mathrm{median}}(q\&Element;N_p),\mathrm{flag}\left(p\right)=\mathrm{mismatched}\\ D\left(p\right),\mathrm{otherwise}\end{array}\right.$ (formula is picked up);

Wherein, D _low(q ∈ N _p) represent and from neighborhood, select less parallax, D _median(q ∈ N _p) represent from neighborhood, select parallax intermediate value.

Utilize above-mentioned disparity map and image elements of exterior orientation can by forward intersection compute depth figure or digital surface model.

Should be understood that, the part that this instructions does not elaborate all belongs to prior art.

Should be understood that; the above-mentioned description for preferred embodiment is comparatively detailed; therefore the restriction to scope of patent protection of the present invention can not be thought; those of ordinary skill in the art is under enlightenment of the present invention; do not departing under the ambit that the claims in the present invention protect; can also make and replacing or distortion, all fall within protection scope of the present invention, request protection domain of the present invention should be as the criterion with claims.

Claims (9)

1., based on a multiple soft-constraint solid matching method for cost matrix, it is characterized in that, comprise the following steps:

Step 1: concentrate at raw video and select a wherein width core line image as reference images, and to utilize AD-Census-Sobel as similarity measure to wherein each pixel, calculate the Matching power flow of each candidate disparity values, generate a three-dimensional Matching power flow matrix;

Step 2: carry out multiple dimensioned down-sampled to the cost matrix obtained in step 1, forms cost matrix pyramid;

Step 3: also carry out corresponding multiple dimensioned down-sampled to reference images, forms image pyramid;

Step 4; Respectively Iamge Segmentation is carried out to each layer image of the image pyramid obtained in step 3;

Step 5: to the cost matrix pyramid obtained in step 2, gathers the cost that each layer cost matrix carries out adaptive weighting from top layer cost matrix;

Step 6: according to the image segmentation result described in step 4, gathers the cost that the cost matrix after process in step 5 carries out under " ballot formula " segmentation constraint further;

Step 7: this layer of cost is gathered result and passes to lower floor's cost matrix;

Step 8: repeat the operation of step 5 to step 7 to pyramid cost matrix from top to bottom, until original cost matrix, finally carries out step 5 to original cost matrix and step 6 operates;

Step 9: to the cost matrix obtained after process in step 8, Least-cost is got to each pixel and mates degree of confidence and be greater than the parallax of predetermined threshold as final parallax, then in cost matrix, cost diffusion being carried out to non-match point as controlling with the point obtaining parallax, finally adopting " winner overwhelm the market " method to generate disparity map;

Step 10: concentrate at raw video and select another image as reference images, repeat above-mentioned steps 1 to step 9, generate another width disparity map, detect Mismatching point by the disparity consistency contrasting two width disparity map respective pixel, obtain removing Mismatching point disparity map; Parallax interpolation is carried out to removal Mismatching point disparity map, fills up the black hole that error hiding causes, generate complete disparity map; Utilize complete disparity map and image elements of exterior orientation, according to forward intersection principle compute depth figure and digital surface model.

2. the multiple soft-constraint solid matching method based on cost matrix according to claim 1, is characterized in that: in step 1,

Described AD-Census-Sobel similarity measure is defined as follows:

Cost (p, d)=exp (α * AD (p, d)+β * Census (p, d)+γ * Sobel (p, d)) (formula one);

Wherein, p represents the pixel coordinate p (x in reference images, y), d is parallax value, α, β and γ is weight coefficient, General Requirements is positive number and alpha+beta+γ=1, AD (p, d) refer to the gray scale difference absolute value of reference images pixel p and the conjugation pixel p+d on coupling image thereof, Census (p, d) refers to the Census match measure value of reference images pixel p and the conjugation pixel p+d on coupling image thereof, Sobel (p, d) refers to the Sobel match measure value of reference images pixel p and the conjugation pixel p+d on coupling image thereof;

Calculate Matching power flow value corresponding to different parallax d according to formula one pair of reference images by pixel, form a three-dimensional Matching power flow matrix represented by image horizontal ordinate W, ordinate H and parallax value D.

3. the multiple soft-constraint solid matching method based on cost matrix according to claim 2, it is characterized in that: the pyramidal generation method of the cost matrix described in step 2 is, keep parallax value D direction constant to the three-dimensional Matching power flow matrix obtained in step 1 and adopt gaussian pyramid to carry out down-sampled at horizontal ordinate W and ordinate H direction, multiple dimensioned cost matrix can be generated through repeatedly down-sampled in this way.

4. the multiple soft-constraint solid matching method based on cost matrix according to claim 2, it is characterized in that: the multiple dimensioned down-sampled method described in step 3 is, adopt gaussian pyramid to carry out down-sampled to raw video at horizontal ordinate W and ordinate H direction, can image pyramid be generated through repeatedly down-sampled in this way.

5. the multiple soft-constraint solid matching method based on cost matrix according to claim 1, it is characterized in that: first the image partition method described in step 4 for adopt existing image partition method, then after Iamge Segmentation completes, block cluster is carried out, and the classification that its element number is greater than predetermined threshold is split, again until the element number of each classification is not more than predetermined threshold.

6. the multiple soft-constraint solid matching method based on cost matrix according to claim 1, it is characterized in that: the cost agglomeration approach of the adaptive weighting described in step 5 is, according to Euclidean distance dist (p, q) and the grey value difference of the pixel q in window certain centered by each pixel p and this point | I _p-I _q| determine the contribution margin of pixel q for pixel p cost, all pixels of accumulation window like this for pixel p cost contribution and obtain the final cost value cost (p, d) of p, concrete computing method are as follows:

$\cos t(p,d)=\underset{q\&Element;W_p}{\mathrm{\&Sigma;}}\cos t(q,d)*\mathrm{weight}(q,d)$ (formula two);

weight(q,d)＝exp(α*dis(p,q)+β*|I _p-I _q|)

Wherein, weight (q, d) represents the weighted value carrying out neighborhood territory pixel when cost is gathered, the parameter related to when α, β are and calculate this weighted value, respectively corresponding Euclidean distance dist (p, q) and grey value difference | I _p-I _q| weight coefficient, General Requirements is positive number and alpha+beta=1.

7. the multiple soft-constraint solid matching method based on cost matrix according to claim 1, is characterized in that: the specific implementation of step 6 comprises following sub-step:

Step 6.1: to each Iamge Segmentation block, total the cost that in this Iamge Segmentation block that adds up respectively, each parallax value of all pixels is corresponding, obtains in whole Iamge Segmentation block, the parallax that Matching power flow changes with parallax value-cost histogram;

Step 6.2: according to histogram calculation Matching power flow average, the difference of each parallax value and average and the average of these differences, and be negative to its difference and the parallax value that absolute value is greater than difference average is composed with punishment, namely compose with a negative cost value, according to said method obtain one and correspond to the histogrammic parallax of parallax-cost described in step 6.1-cost punishment histogram;

To the parallax value of each pixel, step 6.3: apply the punishment of above-mentioned cost by pixel to Iamge Segmentation block, namely according to the parallax described in step 6.2-cost punishment histogram, adds that corresponding cost is punished;

Wherein the computing method of cost punishment are as follows:

$P\left(d\right)=\left\{\begin{array}{cc}0,& c_d>\stackrel{\&OverBar;}{c_d}\\ \mathrm{\&alpha;}*(c_d-\stackrel{\&OverBar;}{c_d})/\stackrel{\&OverBar;}{|c_d-\stackrel{\&OverBar;}{c_d}|},& c_d<c_d\end{array}\right.$ (formula three);

Wherein, c _dfor block corresponds to the cost of parallax d, α is weight coefficient.

8. the multiple soft-constraint solid matching method based on cost matrix according to claim 1, is characterized in that: this layer of cost is gathered result pass to lower floor's cost matrix described in step 7, and its specific implementation comprises following sub-step:

Step 7.1: respectively one dimension least square convolution is carried out to the cost value of each pixel in disparity range;

Step 7.2: according to pyramid pixel corresponding relation, the cost value of this layer of each pixel is added to the cost value of the corresponding parallax of lower floor's respective pixel.

9. the multiple soft-constraint solid matching method based on cost matrix according to claim 1, it is characterized in that: the ratio of the coupling degree of confidence described in step 9 and Matching power flow sub-minimum and minimum value, parallax could be obtained first for the pixel only having coupling degree of confidence to be greater than predetermined threshold; As follows as controlling to carry out the method for cost diffusion to non-match point in cost matrix according to the point obtaining parallax: to mark match point and non-match point, then certain penalty value is deducted, shown in formula specific as follows to the parallax cost that the non-matched pixel in each match point certain window neighborhood corresponds to this match point:

$\cos t(q,d)=\left\{\begin{array}{c}\cos t(q,d)-P,D\left(p\right)=d\\ \cos t(q,d),D\left(p\right)\&NotEqual;d\end{array}\right.,q\&Element;N_d,D\left(q\right)=\mathrm{NULL}$ (formula wantonly);

Wherein, p represents match point, and q represents non-match point, and P represents penalty value, and D (p) represents the value of initial parallax figure D at p point place.