CN103325120A - Rapid self-adaption binocular vision stereo matching method capable of supporting weight - Google Patents

Abstract

The invention discloses a fast self-adaptive weight binocular vision stereo matching method, comprising the following steps: reading an existing binocular image pair to be matched; calculating matching cost; weighted aggregation matching cost; calculating initial parallax; Correction to obtain the final disparity matrix; generate a disparity map and output the result. The invention can be used in the technical field of stereo display to improve the effect of stereo matching.

Description

A Fast Adaptive Weighted Binocular Stereo Matching Method

technical field

The invention relates to the technical field of image display, in particular to a fast adaptive support weight binocular vision stereo matching method.

Background technique

Vision is an important means for human beings to understand and perceive the world. About 75% of the information in human cognition of the outside world is obtained through the visual system. From traditional black-and-white photos and black-and-white TVs to current high-resolution color digital photos and high-definition digital TVs, human beings have higher and higher requirements for visual experience. Although traditional two-dimensional video can already provide high-definition two-dimensional plane information, human beings live in a three-dimensional world, and two-dimensional plane video can never give people an "immersive" visual experience. Binocular stereo vision breaks the limitation of traditional two-dimensional video "seeing the world with one eye", uses computer to simulate the human visual system, and obtains three-dimensional information of the scene through two or more two-dimensional views in the scene. Can feel the real three-dimensional world. Binocular stereo vision is an important research field of computer vision, including four steps: image acquisition, camera calibration, stereo matching and 3D reconstruction, among which stereo matching is the key technology, and the accuracy of stereo matching directly affects the effect of 3D reconstruction.

Although there are a large number of stereo matching methods, there are still many problems in practical applications. According to different optimization methods, stereo matching can be divided into two categories: global stereo matching method and local stereo matching method. The global stereo matching method has high matching accuracy, but the calculation structure is complex, and it is not easy to implement in hardware. The local stereo matching method has a simple structure and is easy to implement in hardware, but the matching accuracy is relatively low. Since Yoon proposed the adaptive support weight method, the stereo matching performance of the local method has been greatly improved, even surpassing some global methods, but there is still an important problem in the Yoon adaptive support weight method: the calculation speed is slow, and the calculation time is slower than other methods. The local algorithm is long. Therefore, it is of great significance to invent a fast adaptive support weight method with fast calculation speed and high matching performance, which is conducive to the application of stereo matching technology to practical problems.

Contents of the invention

The technical problem to be solved by the present invention is to overcome the deficiencies of the Yoon adaptive support weight method, and propose a fast adaptive support weight method based on extended rank transformation, with the left view of the binocular stereo image pair as the image to be matched, and The right view is the matching image, and the corresponding matching point is found in the right view for each pixel to be matched in the left view, that is, the disparity map of the left view is obtained.

The present invention is for solving its technical problem, and the technology adopted comprises the following steps:

S1. Read the existing pair of binocular images I _l and I _r to be matched, and obtain the size and color channel information of the image, wherein, I _l represents the left view, which is the image to be matched, and I _r represents the right view, which is the matched image;

S2. Calculate the matching cost between pixels in I _l and I _r , including:

S21. Determine the square support window N for the image I _l to be matched, and calculate the gray difference diff(p,q) between the center pixel and the support pixel in the support window, where diff(p,q)=I(p)−I( q), I(p) and I(q) are the gray values of the center pixel p and the support pixel q, respectively;

S22. According to the gray difference diff(p,q) obtained in S21, each pixel in the square support window N is defined into five levels, and the five levels are as follows:

其中，s和t是根据经验设置的阈值，满足尽可能减小图像噪声的影响这一原则；

Among them, s and t are thresholds set according to experience, which satisfy the principle of reducing the influence of image noise as much as possible;

S23. Statistically obtain the initial similarity measure value S _d :

Let fuz represent the numerical matrix obtained after rank transformation, and the size of the matrix is the same as that of the square support window N, $\mathrm{fuz}=\left\{\begin{array}{cc}-2& \mathrm{diff}(p,q)<-\mathrm{the\; s}\\ -1& -\mathrm{the\; s}\&le;\mathrm{diff}(p,q)<-t\\ 0& -t\&le;\mathrm{diff}(p,q)\&le;t\\ 1& t<\mathrm{diff}(p,q)\&le;\mathrm{the\; s}\\ 2& \mathrm{the\; s}<\mathrm{diff}(p,q)\end{array}\right.,$ Perform rank transformation on the pixels to be matched in I _l and the candidate matching pixels in I _r respectively, and two rank transformation matrices fuz _l and fuz _r can be obtained. The statistical rank transformation matrix fuz _l and the rank transformation matrix fuz _r are in the square support window The number of corresponding positions with the same level in the corresponding position, and the initial similarity measure S _d is obtained, where, $S_d={\mathrm{\&Sigma;}}_{q\&Element;N}m,$ $m=\left\{\begin{array}{cc}1& \mathrm{if}{\mathrm{fuz}}_1={\mathrm{fuz}}_r\\ 0& \mathrm{otherwise}\end{array}\right.,$ m indicates whether the rank transformation matrices fuz _l and fuz _r have the same level at the corresponding positions, if they are the same, then m=1, otherwise m=0;

其中，d表示待匹配像素和候选匹配像素在水平方向的视差，D＝{d_min,...d_max}， $C_{\mathrm{ERT}}(q,{\stackrel{\&OverBar;}{q}}_d)=\underset{n\&times;n}{\mathrm{\&Sigma;}}{S}_d\left(q\right);$ S24. Count the initial similarity measurement value corresponding to each disparity value d∈D in the n×n square statistical window M centered on the pixel to be matched in the image to be matched, and obtain the pixel to be matched and the matching value according to the ERT similarity measurement function The matching cost between candidate matching pixels in the image

Among them, d represents the disparity between the pixel to be matched and the pixel to be matched in the horizontal direction, D={d _min ,...d _max }, $C_{\mathrm{ERT}}(q,{\stackrel{\&OverBar;}{q}}_d)=\underset{\mathrm{no}\&times;\mathrm{no}}{\mathrm{\&Sigma;}}{S}_d\left(q\right);$

S3, weighted aggregation matching cost, including:

其中，f_s(Δc_pq)表示由颜色相似性确定的聚类强度，f_p(Δg_pq)表示由几何接近性确定的聚类强度，Δc_pq表示两个像素颜色c_p和c_q在RGB颜色空间的欧几里德距离，c_p＝[R_p,G_p,B_p]，c_q＝[R_q,G_q,B_q]， ${\mathrm{\&Delta;c}}_{\mathrm{pq}}=\sqrt{{(R_p-R_q)}^2+{(G_p-G_q)}^2+{(B_p-B_q)}^2},$ Δg_pq表示中心像素和支持像素的在空间位置上的欧几里德距离，设像素p在图像域的坐标为p(x,y)，像素q在图像域的坐标为q(x',y')，则γ_c，γ_p是用户指定参数，分别用来调节颜色相似性和几何接近性对支持权值大小的影响；S31. Calculating the support weight w(p,q): Using color similarity and geometric proximity, calculate the support weight w(p,q),w(p,q) of the support pixel q in the matching support window to be matched with the pixel p )=f _s (Δc _pq )·f _p (Δg _pq ),

where f _s (Δc _pq ) represents the clustering strength determined by color similarity, f _p (Δg _pq ) represents the clustering strength determined by geometric proximity, and Δc _pq represents the two pixel colors c _p and c _q in RGB Euclidean distance of color space, c _p = [R _p , G _p , B _p ], c _q = [R _q , G _q , B _q ], ${\mathrm{\&Delta;\; c}}_{\mathrm{pq}}=\sqrt{{(R_p-R_q)}^2+{(G_p-G_q)}^2+{(B_p-B_q)}^2},$ Δg _pq represents the Euclidean distance between the central pixel and the supporting pixel in the spatial position. Let the coordinates of pixel p in the image domain be p(x, y), and the coordinates of pixel q in the image domain be q(x', y '),but γ _c , γ _p are user-specified parameters, which are used to adjust the influence of color similarity and geometric proximity on the support weights;

和S31所得支持权值w(p,q)，加权聚合匹配代价，得到： $E(p,{\stackrel{\&OverBar;}{p}}_d)=\frac{{\mathrm{\&Sigma;}}_{q\&Element;N_P,\stackrel{\&OverBar;}{q}\&Element;N_{{\stackrel{\&OverBar;}{p}}_d}}w(p,q)c_{\mathrm{ERT}}(q,{\stackrel{\&OverBar;}{q}}_d)}{{\mathrm{\&Sigma;}}_{q\&Element;N_P,\stackrel{\&OverBar;}{q}\&Element;N_{{\stackrel{\&OverBar;}{p}}_d}}w(p,q)},$ 其中，

分别表示当视差为d时待匹配图像像素p、q在匹配图像中对应的匹配像素点，N_p表示参考图像中的支持窗口大小，

表示目标图像中的对应支持窗口大小，且 S32. The matching cost obtained according to S24

Combined with the support weight w(p,q) obtained in S31, and the weighted aggregation matching cost, we get: $\mathrm{E.}(p,{\stackrel{\&OverBar;}{p}}_d)=\frac{{\mathrm{\&Sigma;}}_{q\&Element;N_P,\stackrel{\&OverBar;}{q}\&Element;N_{{\stackrel{\&OverBar;}{p}}_d}}w(p,q)c_{\mathrm{ERT}}(q,{\stackrel{\&OverBar;}{q}}_d)}{{\mathrm{\&Sigma;}}_{q\&Element;N_P,\stackrel{\&OverBar;}{q}\&Element;N_{{\stackrel{\&OverBar;}{p}}_d}}w(p,q)},$ in,

Respectively represent the matching pixel points corresponding to the image pixels p and q to be matched in the matching image when the disparity is d, N _p represents the support window size in the reference image,

denote the corresponding support window size in the target image, and

S4. Calculating the initial disparity: The local optimization method WTA (Winner-Take-All, the winner is king) is used for the weighted aggregation matching cost obtained in S3 to obtain the maximum weighted aggregation result, and the disparity corresponding to the maximum weighted aggregation result is The initial disparity d _p of the pixel, the initial disparity result of each pixel is saved to the initial disparity matrix, and the initial disparity matrix is: $d_p=\arg \underset{d\&Element;\mathrm{D.}}{\max }\mathrm{E.}(p,{\stackrel{\&OverBar;}{p}}_d);$

S5. Correcting the initial parallax obtained in S4 to obtain a final parallax matrix, including:

S51. Determine the correction window N _c centered on the pixel p to be corrected, and adaptively assign an appropriate support weight w _c to each pixel in the correction window according to color similarity and geometric proximity,

S52. Observe the initial parallax distribution of all pixels in the correction window, count the number of times the parallax d∈D appears in the correction window, and aggregate the number of times each parallax value d appears in the correction window and the corresponding weight. The maximum aggregation result corresponds to The disparity is the final disparity d _{p_final} of the pixel to be corrected, and the result is saved to the final disparity matrix, where $d_{p\_\mathrm{final}}=\arg \underset{d\&Element;\mathrm{D.}}{\max }\{{\mathrm{\&Sigma;}}_{q\&Element;N_c}{w}_c(p,q)\&times;k\},$ $k=\left\{\begin{array}{cc}1& \mathrm{if}{d}_p\left(q\right)=d\\ 0& \mathrm{otherwise}\end{array}\right.,$ k indicates whether the initial parallax of the pixels in the correction window is equal to the parallax d to be counted, if they are equal, then k=1, otherwise k=0;

S6. Generate a disparity map, and output the result: map the final disparity value d _{p_final} obtained in S5 to the corresponding grayscale space [0, 255] with a mapping ratio of t, and obtain a grayscale image representing disparity information.

Further, in S22, s<t.

Further, the pixels in the correction window in S51 should come from the same depth as much as possible, and the adaptive support weight square correction window satisfies this condition.

The present invention adopts a fast self-adaptive support weight method based on extended rank transform, takes the left view of the binocular stereo image pair as the image to be matched, takes the right view as the matching image, and uses the left view as the matching image for each pixel point to be matched in the left view Find the corresponding matching point in the right view, that is, obtain the disparity map of the left view. Compared with other calculation methods, the calculation speed is fast and the matching performance is high, which is more conducive to the application of stereo matching technology to practical problems.

Description of drawings

Fig. 1 is a schematic diagram of the method steps of the present invention.

Detailed ways

The present invention will be further described in detail below in conjunction with specific embodiments.

A fast adaptive support weight binocular vision stereo matching method, the purpose of which is to quickly obtain a high-precision dense disparity map of an image pair to be matched. This embodiment takes the Teddy standard test image pair provided by the Middlebury test platform as the experimental object, wherein the left view is the image to be matched, the right view is the matching image, and each pixel point to be matched in the left view is found in the right view corresponding matching points. In order to solve this problem, according to the flow chart 1, the following steps need to be taken:

S1: Read the binocular image pair to be matched. Input the Teddy standard test image pair provided by the Middlebury test platform, in which the left view I _l and the right view I _r . Read the image pair to be matched, including information such as image size and color channel.

S2: Use the extended rank transformation function to calculate the matching cost between I _l and I _r pixels.

First determine the pixel p to be matched, determine the size of the square support window N=25, and calculate the gray difference diff(p,q)=I(p)-I(q) between the center pixel p and the support pixel q in the square support window, where I(p), I(q) represent the gray values of pixels p and q, respectively.

Then, according to the value of diff(p,q), each pixel in the support window is defined to 5 levels.

统计左右视图中方形支持窗口内秩变换矩阵fuz_l(p)和

在对应位置具有相同等级的个数，得到待匹配像素点p和候选匹配像素点

间的初始相似性测度值S_d。Perform rank transformation on the left view and right view to obtain two rank transformation matrices fuz _l (p),

Statistical rank transformation matrices fuz _l (p) and

The number of the same level in the corresponding position, get the pixel point p to be matched and the candidate matching pixel point

The initial similarity measure S _d between them.

${\mathrm{<span\; class="notranslate">\; S</span>}}_{\mathrm{<span\; class="notranslate">\; d</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; q</span>}<span\; class="notranslate">\; ,</span>{\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; q</span>}}}_{\mathrm{<span\; class="notranslate">\; d</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; q</span>}<span\; class="notranslate">\; \&Element;</span>\mathrm{<span\; class="notranslate">\; N</span>}}\mathrm{<span\; class="notranslate">\; m</span>}$

$\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; =</span>\left\{\begin{array}{cc}\mathrm{<span\; class="notranslate">\; 1</span>}& \mathrm{<span\; class="notranslate">\; if</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; fuz</span>}}_{\mathrm{<span\; class="notranslate">\; l</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; fuz</span>}}_{\mathrm{<span\; class="notranslate">\; r</span>}}<span\; class="notranslate">\; (</span>{\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; p</span>}}}_{\mathrm{<span\; class="notranslate">\; d</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; )</span>\\ \mathrm{<span\; class="notranslate">\; 0</span>}& \mathrm{<span\; class="notranslate">\; otherwise</span>}\end{array}\right.$

其中，初始化视差集D={0,1,2,3,…56,57,58,59}。Finally, count the initial similarity measure value corresponding to each disparity value d∈D in the 3×3 square statistical window with the pixel p to be matched as the center of the window, and obtain the matching cost between the pixel to be matched and the candidate matching pixel

Among them, the initial disparity set D={0,1,2,3,...56,57,58,59}.

S3: Weight aggregation matching cost:

First, using the color similarity and geometric proximity, calculate the support weight w(p,q) of the support pixel q in the matching support window to be matched with the pixel p:

$\mathrm{<span\; class="notranslate">\; w</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; q</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; f</span>}}_{\mathrm{<span\; class="notranslate">\; the\; s</span>}}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; \&Delta;\; c</span>}}_{\mathrm{<span\; class="notranslate">\; pq</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; \&Center\; Dot;</span>{\mathrm{<span\; class="notranslate">\; f</span>}}_{\mathrm{<span\; class="notranslate">\; p</span>}}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; \&Delta;\; g</span>}}_{\mathrm{<span\; class="notranslate">\; pq</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; exp</span>}<span\; class="notranslate">\; (</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\frac{{\mathrm{<span\; class="notranslate">\; \&Delta;\; c</span>}}_{\mathrm{<span\; class="notranslate">\; pq</span>}}}{\mathrm{<span\; class="notranslate">\; 19</span>}}<span\; class="notranslate">\; +</span>\frac{{\mathrm{<span\; class="notranslate">\; \&Delta;g</span>}}_{\mathrm{<span\; class="notranslate">\; pq</span>}}}{\mathrm{<span\; class="notranslate">\; 12.5</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; )</span>$

${\mathrm{<span\; class="notranslate">\; \&Delta;\; c</span>}}_{\mathrm{<span\; class="notranslate">\; pq</span>}}<span\; class="notranslate">\; =</span>\sqrt{{<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; R</span>}}_{\mathrm{<span\; class="notranslate">\; p</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; R</span>}}_{\mathrm{<span\; class="notranslate">\; q</span>}}<span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; +</span>{<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; G</span>}}_{\mathrm{<span\; class="notranslate">\; p</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; G</span>}}_{\mathrm{<span\; class="notranslate">\; q</span>}}<span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; +</span>{<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; B</span>}}_{\mathrm{<span\; class="notranslate">\; p</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; B</span>}}_{\mathrm{<span\; class="notranslate">\; q</span>}}<span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}$

${\mathrm{<span\; class="notranslate">\; \&Delta;\; g</span>}}_{\mathrm{<span\; class="notranslate">\; pq</span>}}<span\; class="notranslate">\; =</span>\sqrt{{<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; q</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; x</span>}}^{<span\; class="notranslate">\; \&prime;</span>}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; the\; y</span>}}^{<span\; class="notranslate">\; \&prime;</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}$

Then, using a square support weight window, the matching cost and the corresponding support weight are aggregated, and the size of the support window is $N=N_p=N_{{\stackrel{\&OverBar;}{p}}_d}=25.$

$\mathrm{<span\; class="notranslate">\; E.</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; ,</span>{\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; p</span>}}}_{\mathrm{<span\; class="notranslate">\; d</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\frac{{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; q</span>}<span\; class="notranslate">\; \&Element;</span>{\mathrm{<span\; class="notranslate">\; N</span>}}_{\mathrm{<span\; class="notranslate">\; P</span>}}<span\; class="notranslate">\; ,</span>\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; q</span>}}<span\; class="notranslate">\; \&Element;</span>{\mathrm{<span\; class="notranslate">\; N</span>}}_{{\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; p</span>}}}_{\mathrm{<span\; class="notranslate">\; d</span>}}}}\mathrm{<span\; class="notranslate">\; w</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; q</span>}<span\; class="notranslate">\; )</span>{\mathrm{<span\; class="notranslate">\; c</span>}}_{\mathrm{<span\; class="notranslate">\; ERT</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; q</span>}<span\; class="notranslate">\; ,</span>{\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; q</span>}}}_{\mathrm{<span\; class="notranslate">\; d</span>}}<span\; class="notranslate">\; )</span>}{{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; q</span>}<span\; class="notranslate">\; \&Element;</span>{\mathrm{<span\; class="notranslate">\; N</span>}}_{\mathrm{<span\; class="notranslate">\; P</span>}}<span\; class="notranslate">\; ,</span>\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; q</span>}}<span\; class="notranslate">\; \&Element;</span>{\mathrm{<span\; class="notranslate">\; N</span>}}_{{\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; p</span>}}}_{\mathrm{<span\; class="notranslate">\; d</span>}}}}\mathrm{<span\; class="notranslate">\; w</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; q</span>}<span\; class="notranslate">\; )</span>}$

S4: Calculate the initial disparity from the aggregation results. Using the local optimization method WTA, the disparity value corresponding to the maximum weighted result is the initial disparity d _p of the pixel, and the result is saved to the initial disparity matrix.

${\mathrm{<span\; class="notranslate">\; d</span>}}_{\mathrm{<span\; class="notranslate">\; p</span>}}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; arg</span>}\underset{\mathrm{<span\; class="notranslate">\; d</span>}<span\; class="notranslate">\; \&Element;</span>\mathrm{<span\; class="notranslate">\; D.</span>}}{\mathrm{<span\; class="notranslate">\; max</span>}}\mathrm{<span\; class="notranslate">\; E.</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; ,</span>{\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; p</span>}}}_{\mathrm{<span\; class="notranslate">\; d</span>}}<span\; class="notranslate">\; )</span>$

S5: Correct the initial parallax

First, determine the correction window N _c =21 centered on the pixel p to be corrected, and calculate the support weight _wc (p,q) of the pixel p to be corrected within the correction window based on the color similarity and geometric proximity :

${\mathrm{<span\; class="notranslate">\; w</span>}}_{\mathrm{<span\; class="notranslate">\; c</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; q</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; exp</span>}<span\; class="notranslate">\; (</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\frac{{\mathrm{<span\; class="notranslate">\; \&Delta;\; c</span>}}_{\mathrm{<span\; class="notranslate">\; pq</span>}}}{\mathrm{<span\; class="notranslate">\; 19</span>}}<span\; class="notranslate">\; +</span>\frac{{\mathrm{<span\; class="notranslate">\; \&Delta;g</span>}}_{\mathrm{<span\; class="notranslate">\; pq</span>}}}{\mathrm{<span\; class="notranslate">\; 10.5</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; )</span>$

Then observe the initial disparity distribution of all pixels in the correction window, count the number of times that the disparity d∈D appears in the correction window, and aggregate the number of times each disparity value d appears in the calibration window and the corresponding weight. The maximum aggregation result corresponds to The disparity is the final disparity d _{p_final} of the pixel to be corrected, and the final disparity result is saved to the final disparity matrix.

${\mathrm{<span\; class="notranslate">\; d</span>}}_{\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; \_</span>\mathrm{<span\; class="notranslate">\; final</span>}}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; arg</span>}\underset{\mathrm{<span\; class="notranslate">\; d</span>}<span\; class="notranslate">\; \&Element;</span>\mathrm{<span\; class="notranslate">\; D.</span>}}{\mathrm{<span\; class="notranslate">\; max</span>}}<span\; class="notranslate">\; \{</span>{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; q</span>}<span\; class="notranslate">\; \&Element;</span>{\mathrm{<span\; class="notranslate">\; N</span>}}_{\mathrm{<span\; class="notranslate">\; c</span>}}}{\mathrm{<span\; class="notranslate">\; w</span>}}_{\mathrm{<span\; class="notranslate">\; c</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; q</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; \&times;</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; \}</span>$

$\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; =</span>\left\{\begin{array}{cc}\mathrm{<span\; class="notranslate">\; 1</span>}& \mathrm{<span\; class="notranslate">\; if</span>}{\mathrm{<span\; class="notranslate">\; d</span>}}_{\mathrm{<span\; class="notranslate">\; p</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; q</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; d</span>}\\ \mathrm{<span\; class="notranslate">\; 0</span>}& \mathrm{<span\; class="notranslate">\; otherwise</span>}\end{array}\right.$

S6: Generate a disparity map and output the result. Map the disparity value in the final disparity matrix to the corresponding grayscale space [0, 255], the mapping ratio t=4, and the mapping result of the disparity value in the disparity set D is as follows:

0×4＝0 1×4＝4 ... 13×4＝52 14×4＝56

15×4=60 16×4=64 ... 28×4=112 29×4=116

30×4=120 31×4=124 ... 43×4=172 44×4=176

45×4=180 46×4=184 ... 58×4=232 59×4=236

Among them, the larger the disparity value, the closer to 255 after mapping, the brighter it is in the disparity map, the smaller the disparity value, the closer to 0 after mapping, and the darker it is in the disparity map.

The present invention uses I _l as the image to be matched and I _r as the matching image, finds a corresponding matching point in the right view for each pixel in the left view, and obtains the disparity map of the left view. Table 1 is a quantitative comparison between the results of the embodiment of the present invention and the results of the Yoon adaptive support weight method. As can be seen from Table 1, the present invention is in the non-occlusion region, the depth discontinuity region, and the false matching rate of all regions is all lower than the false matching rate of the Yoon adaptive support weight method, and the stereo matching time of the present invention is about Yoon adaptive supports 1/20 of the matching time of the weight method, and the matching speed is faster.

Table 1

It must be pointed out here that the above embodiments are only used to further illustrate the present invention, so that those skilled in the art can better understand the present invention. The present invention has disclosed its preferred embodiment by text, but can comprehend optimization and modifiability wherein by reading these technical text description, and can improve without departing from the scope and spirit of the present invention, but such improvement should Still belong to the protection scope of the claims of the present invention.

Claims (3)

1. a quick self-adapted support-weight binocular vision solid matching method is characterized in that, may further comprise the steps:

S1, read existing binocular image to be matched to I _lAnd I _r, size and the Color Channel information of acquisition image, wherein, I _lThe expression left view is image to be matched, I _rThe expression right view is matching image;

S2, calculating I _l, I _rCoupling cost between middle pixel comprises:

S21, treat matching image I _l, determine square support window N, calculate and support center pixel and the gray scale difference diff (p that supports pixel in the window, q), wherein, diff (p, q)=and I (p)-I (q), I (p) and I (q) are respectively central pixel point p and support the gray-scale value of pixel q;

S22, according to S21 gained gray scale difference diff (p, q), with each pixel definition is in 5 grades in the square support window N, five grades are as follows:

Wherein, s and t are the threshold values that rule of thumb arranges, and satisfy this principle that affects that reduces as far as possible picture noise;

S23, statistics draw initial similarity measure value S _d:

Make fuz represent the numerical matrix that obtains after the order conversion, matrix size is identical with square support window N size, $\mathrm{fuz}=\left\{\begin{array}{cc}-2& \mathrm{diff}(p,q)<-s\\ -1& -s\&le;\mathrm{diff}(p,q)<-t\\ 0& -t\&le;\mathrm{diff}(p,q)\&le;t\\ 1& t<\mathrm{diff}(p,q)\&le;s\\ 2& s<\mathrm{diff}(p,q)\end{array}\right.,$ Respectively to I _lIn pixel to be matched and I _rMiddle candidate matches pixel carries out the order conversion, can obtain two order transformation matrix fuz _lAnd fuz _r, statistics order transformation matrix fuz _lWith order transformation matrix fuz _rCorrespondence position has the number of same levels in square support window, obtains initial similarity measure value S _d, wherein, $S_d={\mathrm{\&Sigma;}}_{q\&Element;N}m,$ $m=\left\{\begin{array}{cc}1& \mathrm{if}{\mathrm{fuz}}_l={\mathrm{fuz}}_r\\ 0& \mathrm{otherwise}\end{array}\right.,$ M represents order transformation matrix fuz _lAnd fuz _rWhether have same levels at correspondence position, if identical, m=1 then, otherwise m=0;

Initial similarity measure value corresponding to each parallax value d ∈ D in the square statistic window M of n * n in S24, the statistics image to be matched centered by the pixel to be matched obtains the coupling cost between the candidate matches pixel in pixel to be matched and the matching image according to ERT similarity measurement function

Wherein, d represents pixel to be matched and candidate matches pixel parallax in the horizontal direction, D={d _Min... d _Max, $C_{\mathrm{ERT}}(q,{\stackrel{\&OverBar;}{q}}_d)=\underset{n\&times;n}{\mathrm{\&Sigma;}}{S}_d\left(q\right);$

S3, weighting polymerization coupling cost comprise:

S31, calculating support-weight w (p, q): utilize color similarity and how much proximities, calculate the support-weight w (p, q) that the interior support of Matching supporting window pixel q treats matched pixel p, w (p, q)=f _s(Δ c _Pq) f _p(Δ g _Pq),

Wherein, f _s(Δ c _Pq) represent by the definite cluster intensity of color similarity, f _p(Δ g _Pq) represent by how much definite cluster intensity of proximity, Δ c _PqRepresent two pixel color c _pAnd c _qAt the Euclidean distance of RGB color space, c _p=[R _p, G _p, B _p], c _q=[R _q, G _q, B _q], ${\mathrm{\&Delta;c}}_{\mathrm{pq}}=\sqrt{{(R_p-R_q)}^2+{(G_p-G_q)}^2+{(B_p-B_q)}^2},$ Δ g _PqThe Euclidean distance on the locus of expression center pixel and support pixel, establishing pixel p is p (x, y) at the coordinate of image area, pixel q is q (x', y') at the coordinate of image area, then

γ _c, γ _pBe user-specified parameters, be used for respectively the impact on the support-weight size of adjustable colors similarity and how much proximities;

S32, according to S24 gained coupling cost

With S31 gained support-weight w (p, q), weighting polymerization coupling cost obtains: $E(p,{\stackrel{\text{-}}{p}}_d)=\frac{{\mathrm{\&Sigma;}}_{q\&Element;N_P,\stackrel{\&OverBar;}{q}\&Element;N_{{\stackrel{\&OverBar;}{p}}_d}}w(p,q)c_{\mathrm{ERT}}(q,{\stackrel{\&OverBar;}{q}}_d)}{{\mathrm{\&Sigma;}}_{q\&Element;N_P,\stackrel{\&OverBar;}{q}\&Element;N_{{\stackrel{\&OverBar;}{p}}_d}}w(p,q)},$ Wherein,

The matched pixel point that represents respectively image pixel p to be matched, q correspondence in matching image when parallax is d, N _pSupport window size in the expression reference picture,

Corresponding support window size in the expression target image, and

S4, calculating initial parallax: gained weighting polymerization coupling cost among the S3 is adopted suboptimization method WTA(Winner-Take-All, the victor is a king), draw the maximum weighted polymerization result, the parallax value that the maximum weighted polymerization result is corresponding is the initial parallax d of this pixel _p, the initial parallax result of each pixel is saved in the initial parallax matrix, and the initial parallax matrix is: $d_p=\arg \underset{d\&Element;D}{\max }E(p,{\stackrel{\&OverBar;}{p}}_d);$

S5, S4 gained initial parallax is proofreaied and correct, is obtained final parallax matrix, comprising:

S51, definite correcting window N centered by pixel p to be corrected _c, be that each pixel is distributed a suitable support-weight w adaptively in the correction window according to color similarity and how much proximities _c,

The initial parallax of all pixels distributes in S52, the observation correction window, the number of times that statistical parallax d ∈ D occurs in correction window, and the number of times that in correction window, occurs of each parallax value of polymerization d and corresponding weights, the parallax that maximum polymerization result is corresponding then is the final parallax d of pixel to be corrected _{P_final}, and the result is saved in final parallax matrix, wherein $d_{p\_\mathrm{final}}=\arg \underset{d\&Element;D}{\max }\{{\mathrm{\&Sigma;}}_{q\&Element;N_c}{w}_c(p,q)\&times;k\},$ $k=\left\{\begin{array}{cc}1& \mathrm{if}{d}_p\left(q\right)=d\\ 0& \mathrm{otherwise}\end{array}\right.,$ K represents whether the initial parallax of pixel equals to treat statistical parallax d in the correction window, if equate, and k=1 then, otherwise k=0;

S6, generation disparity map, Output rusults: with the final parallax value d of S5 gained _{P_final}Be mapped to corresponding gray space [0,255], the mapping ratio is t, obtains representing the gray level image of parallax information.

2. a kind of quick self-adapted support-weight binocular vision solid matching method according to claim 1 is characterized in that: s＜t described in the S22.

3. a kind of quick self-adapted support-weight binocular vision solid matching method according to claim 1 is characterized in that: pixel should be as far as possible from same depth in the described correction window of S51, and the square correction window of self-adaptation support-weight satisfies this condition.