CN103996201A - Stereo matching method based on improved gradient and adaptive window - Google Patents

Abstract

Aiming at the problems that the existing local stereo matching algorithm has low precision and is easily affected by amplitude distortion, the present invention proposes a stereo matching method based on improved gradient and self-adaptive window. Firstly, on the basis that the traditional gradient vector only contains amplitude information, the phase information is introduced, and the original matching cost is transformed to further eliminate outliers; then a cross-based adaptive window generation method is adopted, which can be based on the adjacent pixels An adaptive window is constructed based on the relationship between color and spatial position. In low-texture areas, larger windows are provided to improve matching accuracy; while in high-texture areas, smaller windows are generated to protect details such as object edges. For the aggregated cost, the "Winner-Takes-All (WTA)" strategy is used to select the disparity value corresponding to the minimum total cost as the initial matching result; finally, a disparity value of the local disparity histogram is proposed. Refining the method, a high-precision disparity map is obtained. Experimental results show that the proposed method has high matching accuracy and is robust to amplitude distortion conditions.

Description

A kind of solid matching method based on improving gradient and self-adapting window

Technical field:

The present invention relates to a kind of binocular image Stereo matching and obtain depth map method, be particularly related to a kind of solid matching method based on improving gradient and self-adapting window, the depth map result that the present invention obtains can further be applied to the fields such as vision measurement, three-dimensionalreconstruction, drawing virtual view image.

Background technology:

Stereo matching is a classical problem in computer vision, is the focus of research always.For many years, researchers have proposed a large amount of algorithms and have attempted to address this problem, but due to the pathosis of problem itself, at present perfect solution without comparison also.Scharstein is (referring to Daniel Scharstein, Richard Szeliski.A taxonomy and evaluation of dense two-frame stereo correspondence algorithms[J] .International Journal of Computer Vision, 2002,47 (1): 7-42.) etc. furtherd investigate some typical Stereo Matching Algorithm, various main method have been carried out to more comprehensive summary.They Stereo matching process be summarised as that coupling cost is calculated, coupling cost polymerization, initial parallax calculate and parallax four steps of refining, and according to cost polymerization methods, Stereo Matching Algorithm be divided into partial approach and global approach.Global approach generally has higher matching precision, but efficiency is lower; Partial approach travelling speed is fast, be easy to realize, but the coupling cost computing method of the support window that How to choose is suitable and pixel are difficult problem (referring to YANG Qing-xiong.A non-local cost aggregation method for stereo matching[C] .IEEE Conference on Computer Vision and Pattern Recognition, 2012:1402-1409.).

The similarity measure of most of Stereo Matching Algorithm is all the gray-scale value based on pixel, the same unique point in two width images should have identical gray-scale value (referring to Wang Junzheng under desirable illumination condition, Zhu Huajian, Li Jing. a kind of variable weight Stereo Matching Algorithm [J] based on Census conversion. Beijing Institute of Technology's journal, 2013,33 (7): 704-710.).For example gray scale difference absolute value and (AD), gray scale difference quadratic sum (SD), Adapt Weight is (referring to Yoon K, Kweon S.Locally adaptive support weight approch for visual correspondence search[J] .IEEE Transactions on Pattern Analysis and Machine Intelligence, 2006, 28 (4): 924-931.), Segment Support is (referring to Tombari F, Mattoccia S.Segmentation based adaptive support for accurate stereo correspondence[C] .IEEE Pacific Rim Symposium on Video and Technology, 2007:427-438.) etc. can obtain the matching result of degree of precision for ideal image, but these methods are for by illumination variation, difference in exposure, the picture amplitude distortion that the factors such as the dark angle of camera cause is very responsive, therefore be difficult to use in the coupling of real scene image.The insensitive coupling cost of amplitude distortion is mainly contained to normalized crosscorrelation (NCC), gradient (Gradient) is (referring to Daniel Scharstein.View synthesis using stereo vision.Phd thesis, 1997, 23 (5): 98-109.), Rank and Census convert (referring to Ramin Zabih, John Woodfill.Non-parametric local transforms for computing visual correspondence[C] .Proceedings of European Conference on Computer Vision, 1994:151-158.) etc.

There is the selection problem of mapping window size in traditional stationary window: if window is too little, mates cost discrimination too low, easily occur mistake coupling at low texture region; If window is excessive, can occur mating at degree of depth discontinuity zone again (referring to Zhou Long, Xu Guili, Li Kaiyu etc. based on the Stereo Matching Algorithm [J] of Census conversion and improvement self-adapting window. aviation journal; by mistake 2012,33 (5): 886-892.).Fusiello and Roberto are (referring to Fusiello A, Roberto V.Efficient stereo with multiple windowing[C] .IEEE Computer Society Conference on Computer vision and Pattern Recognition, 1997:858-863.) propose to select optimum window as support window in multiple windows given in advance; Veksler (referring to Veksler O.Fast variable window for stereo correspondence using integral image[C] .IEEE Computer Society Conference on Computer vision and Pattern Recognition, 2003:556-561.) proposes pointwise self-adaptation and chooses shape and the size of support window; Zhang (referring to Zhang K.Cross-based local stereo matching using orthogonal integral images[J] .IEEE Transactions on Circuits and systems for Video Technology, 2009,19 (7): 1073-1079.), according to the adaptively selected arbitrary shape of the color relations of neighbor and big or small support window, obtain good parallax result.

Summary of the invention:

For above-mentioned coupling cost and window selection problem, the present invention proposes a kind of matching process based on improving gradient coupling cost and self-adapting window.First only comprise on the basis of amplitude information in traditional gradient coupling cost, introduce Gradient Phase information, and original match cost is converted, further eliminate exceptional value; Then utilize picture structure and color information to build that self-adapting window carries out cost polymerization and " the victor is a king " (Winner-Takes-All (WTA)) strategy carries out parallax selection; Finally, propose the histogrammic parallax refined method of a kind of local parallax, obtained high-precision disparity map.

The technical problem to be solved in the present invention is:

1. traditional local matching method is lower to picture amplitude distortion-robust, the problem that matching precision reduces rapidly under illumination distortion condition.

2. the cost polymerization of stationary window is difficult to obtain at the low texture of image and the degree of depth discontinuity zone problem of higher matching precision simultaneously.

The technical solution adopted for the present invention to solve the technical problems: based on the matching process that improves gradient coupling cost and self-adapting window, it is characterized in that comprising the following steps:

Step 1: coupling cost is calculated: coupling cost is the tolerance of corresponding point similarity between the image of left and right, utilizes two components of gradient vector in image x, y direction, mould m and the phase angle of definition gradient vector then adopt the linearity of mould and phase angle to be combined as coupling cost function, to utilize to greatest extent gradient information.Finally utilize Geman-McClure function original match cost function to be converted to eliminate the impact of abnormal cost value.

Step 2: adaptive windows outlet structure: treat the polymerization window of a self-adaptation size of each pixel structure of matching image, it is how many that the large young pathbreaker of window directly determines to participate in the neighborhood territory pixel of polymerization.The present invention adopts a kind of improved right-angled intersection self-adapting window generation method, can build self-adapting window according to the color of neighbor and spatial relation.At low texture region, provide larger window to improve matching precision; Produce less window at high texture region, to protect the detailed information such as object edge.

Step 3: cost polymerization: after determining the self-adapting window of each pixel, need carry out polymerization to the original match cost of each single pixel in window and obtain total cost, finally select to make parallax value that total Least-cost is corresponding as initial matching result.

Step 4: parallax is refined: the initial parallax obtaining by above-mentioned steps also exists some Mismatching points and insincere value with true parallax, need to carry out the parallax processing of refining.The present invention proposes one and based on the histogrammic parallax refined method of local parallax, initial parallax figure is further processed.Then, adopt left and right consistency check to detect the Mismatching point still existing, utilize the value that in adjacent available point, parallax is less to carry out assignment to Mismatching point.

Brief description of the drawings:

Fig. 1 is the process flow diagram of the inventive method.

Fig. 2 is self-adapting window construction process schematic diagram.

Fig. 3 (a) is the right experimental result of Tsukuba image.

Fig. 3 (b) is the right experimental result of Venus image.

Fig. 3 (c) is the right experimental result of Teddy image.

Fig. 3 (d) is the right experimental result of Cones image.

Fig. 4 (a) is the experimental result under same light photograph and conditions of exposure.

Fig. 4 (b) is the experimental result under different illumination conditions.

Fig. 4 (c) is the experimental result under different exposure.

Fig. 4 (d) is the experimental result under different light and conditions of exposure.

Embodiment:

Be described in further detail the present invention below in conjunction with the drawings and the specific embodiments.

The present invention proposes a kind of matching process based on improving gradient coupling cost and self-adapting window, mainly contain four steps.

Step 1: coupling cost is calculated.Concrete methods of realizing is as follows:

Image gradient is defined as the single order partial derivative of image along x and y direction:

$G=\left(\begin{array}{c}{G}_x\\ G_y\end{array}\right)=\left(\begin{array}{c}\frac{\∂I}{\∂x}\\ \frac{\∂I}{\∂y}\end{array}\right)---\left(1\right)$

Wherein I is gradation of image, in practical application, and can be by the formwork calculation gradient vector of horizontal direction and vertical direction.Like this, we just can obtain the gradient map G of left and right image _l=(G _lx, G _ly) ^t, G _r=(G _rx, G _ry) ^t; Consider the image after proofreading and correct, establishing p (x, y) is a bit on left image, and on right image, the match point of corresponding parallax d is pd (x-d, y).

Utilize two components of gradient vector in x, y direction, mould and the phase angle of definition gradient vector:

$m=\sqrt{G_x^2+G_y^2}---\left(2\right)$

The mould m of gradient characterizes rate of gray level, phase angle direction while characterizing rate of gray level maximum, they provide the different information of neighborhood of pixels, and illumination distortion is had to different unchangeability.Input picture can affect gradient-norm to gain distortions, and phase angle can not change, but they can not be subject to bias distortion impact.Thereby, the mould of gradient and phase angle are separately considered to be more conducive to the susceptibility of control method to noise.The present invention adopts the linearity of mould and phase angle to be combined as coupling cost function, to utilize to greatest extent gradient information.Be expressed as follows:

M in formula _c, represent respectively corresponding to coloured image R, G, gradient mode and the phase angle of tri-passages of B, α is weighting coefficient.Because phase angle is taking π as the cycle, need to be normalized in the monocycle, therefore definition f:

$f\left(x\right)=\left\{\begin{array}{cc}x& 0\&le;x\&le;\mathrm{\&pi;}\\ 2\mathrm{\&pi;}-x& \mathrm{\&pi;}<x<2\mathrm{\&pi;}\end{array}\right.---\left(5\right)$

Owing to having introduced weighting coefficient α, we can be by adjusting the value change method of the parameter alpha robustness to illumination distortion and noise.α is less, and the impact of phase place is larger, and α is larger, and the impact of mould value is larger.Because different images has illumination distortion in various degree, in reality, need to come by experiment to determine the reasonable value scope of α.

Due to e (p, d) represent be the original match cost of single pixel, in actual conditions, still can there are some excessive exceptional values, need to get rid of to improve matching precision.Conventional method is to adopt a truncated function, compares with a constant by e (p, d), gets its minimum value as coupling cost.The method is very little to the improvement of result, and the present invention adopts Geman-McClure function to process exceptional value:

$\mathrm{\&rho;}\left(x\right)=\frac{x^2}{x^2+{\mathrm{\&sigma;}}^2}---\left(6\right)$

When input, x exceedes after certain value, and it will drop to 0 smoothly on the impact of output valve, and the coupling cost after conversion will converge to 1, and can be controlled by parameter σ.Thereby, no matter input original match cost much, after Geman-McClure functional transformation, its output valve will can not exceed 1.

Step 2: adaptive windows outlet structure.Concrete methods of realizing is as follows:

The core of method therefor of the present invention is to build self-adapting window according to the color of neighbor and spatial relation, and concrete construction process as shown in Figure 2.First, determine a right-angled intersection region of current pixel p to be matched according to picture structure and color information, this right-angled intersection district inclusion horizontal and vertical direction, uses respectively H (p) and V (p) to represent, the size in region is by the brachium of 4 directions determine, and can be according to the structure of image and color information adaptively modifying.With for example, the criterion of brachium is as follows:

1.D _c(p _i, p) < τ ₁and D _c(p _i, p _i+ (1,0)) < τ ₁;

2.D _s(p _i,p)<L ₁；

3.D _c(p _i,p)<τ ₂,L ₂<D _s(p _i,p)<L ₁。

Wherein, D _s(p _i, p) be pixel p _ipoor with the space length of p; D _c(p _i, p) be color difference, be defined as τ ₁> τ ₂, L ₁>L ₂, be default color threshold value and distance threshold.Criterion 1 not only defines p _ithe color difference opposite sex with p requires p simultaneously _iwith its right side neighbor p _ithe color difference opposite sex of+(1,0) is less than τ ₁, avoided brachium to stride across borderline region; Criterion 2 and 3 has been relaxed brachium scope, uses larger distance threshold L at low texture region ₁can obtain larger window; And in the time that brachium exceedes preset value L ₂time, will adopt stricter threshold tau ₂ensure that brachium is only in the very close low texture region expansion of color, make high texture region and the degree of depth discontinuity zone window can be not excessive.

Utilize said method can determine respectively 4 brachiums and then obtain orthogonal right-angled intersection region H (p) and V (p):

$\left\{\begin{array}{c}H\left(p\right)=\left\{(x,y)\right|x\&Element;[x_p-h_p^{-},x_p+h_p^{+}],y=y_p\}\\ V\left(p\right)=\left\{(x,y)\right|x=x_p,y\&Element;[y_p-v_p^{-},y_p+v_p^{+}]\}\end{array}\right.---\left(7\right)$

Finally, along vertical direction, each pixel q in V (p) is repeated to said process, the adaptive region of trying to achieve any pixel p in image is:

$U\left(p\right)=\underset{q\&Element;V\left(p\right)}{\&cup;}H\left(q\right)---\left(8\right)$

Step 3: cost polymerization.Concrete methods of realizing is as follows:

The present invention will consider left and right image local support area separately symmetrically.For two corresponding match point p (x in the image of left and right, and pd (x-d y), y), utilize said method can generate respectively adaptive region U (p) and U'(pd), their associating public domain is defined as final support area by we:

U _d(p)＝{(x,y)|(x,y)∈U(p),(x-d,y)∈U'(pd)} (9)

Then, in associating support area, original single pixel matching cost is carried out to polymerization, tries to achieve total cost in region:

$E_d\left(p\right)=\frac{1}{N}\underset{q\&Element;U_d\left(p\right)}{\mathrm{\&Sigma;}}e(q,d)---\left(10\right)$

In formula, N is zone of convergency U _d(p) the total number of pixel in.Finally adopt " the victor is a king " (Winner-Takes-All (WTA)) strategy, in parallax interval, select the point of coupling Least-cost to carry out parallax selection as matching double points p point, obtain initial parallax:

$d_p^0=\arg \underset{0\&le;d\&le;d_{\max }}{\min }{E}_d\left(p\right)---\left(11\right)$

Wherein d represents the possible parallax in disparity space, and its value is generally 0 to maximum disparity d _maxbetween integer.

Step 4: parallax is refined.Concrete methods of realizing is as follows:

The present invention proposes one and based on the histogrammic parallax refined method of local parallax, initial parallax figure is further processed.For certain pixel p in disparity map, we construct a local parallax histogram centered by it within the scope of its neighborhood the number of times that in statistics field, each parallax value occurs.In histogram, will there is a peak value, represent the maximum times that parallax occurs.The parallax value that this peak value is corresponding is the optimum parallax value in statistical significance the present invention adopts this optimal value to replace the initial parallax of pixel p

This process has been utilized the self-adapting window that generates in the previous step neighborhood as pixel, thereby does not produce extra computation burden.In addition, the present invention carries out iteration three times to this process, makes optimum parallax value more accurate.

Then, adopt left and right consistency check to detect the Mismatching point still existing, be labeled as available point by the point of consistency check, otherwise be Null Spot.For the Mismatching point detecting, first available point of horizontal scan direction left and right, and utilize the value that in both, parallax is less to carry out assignment to Mismatching point.

For the validity of verification method, on VS2008 platform, to the inventive method realization of programming, and the standard stereotome of being issued by Middlebury website that adopts academic circles at present to generally acknowledge is tested and is evaluated and tested method.This website provides 4 groups of reference color images to Tsukuba, Venus, Teddy and Cones, and corresponding real depth map.Experimental result and true disparity map (Ground truth) relatively can be obtained to the matching error quantizing, thus evaluation method precision objectively.Table 1 is that the inventive method is at limits of error δ _dthe mistake matched pixel number percent data of=1 o'clock.The row at Noocc, All, Disc place be respectively unshielding region mistake matched pixel ratio, always by mistake matched pixel than and degree of depth discontinuity zone mistake matched pixel ratio.And with GC+occ (referring to Kolmogorov V, Rabih R.Computing visual correspondence with occlusions using graph cuts[C] .IEEE Conference on Computer Vision, 2001.), SemiGlob (referring to Hirschm ü ller H.Accurate and efficient stereo processing by semi-global matching and mutual information[J] .IEEE Transactions on Pattern Analysis and Machine Intelligence, 2008,30 (2): 328-341.), AdaptWeight is (referring to Yoon K, Kweon S.Locally adaptive support weight approch for visual correspondence search[J] .IEEE Transactions on Pattern Analysis and Machine Intelligence, 2006,28 (4): 924-931.) and Enhanced BP (referring to Larsen S, Mordohai P, Pollefeys M et al.Temporally consistent reconstruction from multiple video streams using enhanced belief propagation[C] .IEEE International Conference on Computer Vision, 2007:1-8.) method contrasts.

Table 1 matching result objective evaluation

Fig. 3 has reacted the precision of the inventive method operation result intuitively.(a), (b), (c), (d) are followed successively by Tsukuba, Venus, Teddy and the right experimental result of Cones image; In figure, the 1st classifies original image to be matched as; The 2nd classifies corresponding true disparity map as; The 3rd classifies the disparity map that the inventive method obtains as; The 4th classifies the mistake matched pixel figure of the inventive method as, and in figure, white bulk zone is the point that coupling is correct, and gray area and black region represent respectively the Mismatching point in occlusion area and unobstructed region.

Fig. 4 is the operation result of the present invention under illumination distortion condition, and is that the method for mating cost compares with using traditional gray scale absolute difference (AD), to verify the robustness of the inventive method to illumination distortion.As shown in Figure 4, in figure, (a), (b), (c), (d) are respectively the experimental result under same light photograph and exposure, different light, different exposure and different light and conditions of exposure to experimental result.The 1st of every group of result classified original left image as, and the 2nd classifies original right image as, and the 3rd classifies the Variable Cross method matching result that adopts SAD coupling cost as, and the 4th classifies the inventive method matching result as.Experimental result shows, the present invention, owing to having adopted the gradient information that amplitude distortion is had to repellence as coupling cost, in the situation that having illumination distortion, still can obtain the matching result that precision is higher, has good robustness.

Claims (4)

1. A stereo matching method based on improved gradient and adaptive window, characterized in that: the method introduces gradient phase information on the basis that the traditional gradient matching cost only includes magnitude information, and transforms the original matching cost, further Eliminate outliers; then use the image structure and color information to construct an adaptive window for cost aggregation and "Winner-Takes-All (WTA)" strategy for disparity selection; finally, a local disparity histogram is proposed The disparity refinement method obtains a high-precision disparity map. The specific steps of the method are as follows: Step 1: Matching cost calculation: The matching cost is a measure of the similarity of corresponding points between the left and right images. Using the two components of the gradient vector in the x and y directions of the image, define the modulus m and phase angle of the gradient vector A linear combination of magnitude and phase angle is then adopted as the matching cost function to maximize the use of gradient information. Finally, the Geman-McClure function is used to transform the original matching cost function to eliminate the influence of abnormal cost value. Step 2: Adaptive window construction: Construct an adaptive size aggregation window for each pixel of the image to be matched, and the size of the window will directly determine the number of neighboring pixels participating in the aggregation. The invention adopts an improved cross adaptive window generation method, which can construct an adaptive window according to the color and spatial position relationship of adjacent pixels. In low-texture areas, larger windows are provided to improve matching accuracy; while in high-texture areas, smaller windows are generated to protect details such as object edges. Step 3: Cost aggregation: After determining the adaptive window for each pixel, it is necessary to aggregate the original matching cost of each single pixel in the window to obtain the total cost, and finally select the disparity value corresponding to the minimum total cost as the initial matching result. Step 4: Disparity refinement: There are still some mismatching points and unreliable values between the initial disparity obtained through the above steps and the real disparity, and disparity refinement processing is required. The present invention proposes a parallax refinement method based on a local parallax histogram to further process the initial parallax map. Then, the left-right consistency test is used to detect the still existing mis-matching points, and the value of the mis-matching points is assigned with the value of the smaller parallax among the adjacent valid points. 2. according to a kind of stereo matching method based on improved gradient and adaptive window described in claim 1, it is characterized in that: the gradient cost of the matching cost application improvement of described step 1, concrete calculation process is as follows: The image gradient is defined as the first-order partial derivative of the image along the x and y directions: $\mathrm{<span\; class="notranslate">\; G</span>}<span\; class="notranslate">\; =</span>\left(\begin{array}{c}{\mathrm{<span\; class="notranslate">\; G</span>}}_{\mathrm{<span\; class="notranslate">\; x</span>}}\\ {\mathrm{<span\; class="notranslate">\; G</span>}}_{\mathrm{<span\; class="notranslate">\; the\; y</span>}}\end{array}\right)<span\; class="notranslate">\; =</span>\left(\begin{array}{c}\frac{<span\; class="notranslate">\; \&PartialD;</span>\mathrm{<span\; class="notranslate">\; I</span>}}{<span\; class="notranslate">\; \&PartialD;</span>\mathrm{<span\; class="notranslate">\; x</span>}}\\ \frac{<span\; class="notranslate">\; \&PartialD;</span>\mathrm{<span\; class="notranslate">\; I</span>}}{<span\; class="notranslate">\; \&PartialD;</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}}\end{array}\right)$ Where I is the grayscale of the image. In practical applications, the gradient vector can be calculated through templates in the horizontal and vertical directions. In this way, we can get the gradient maps of the left and right images G _L =(G _Lx ,G _Ly ) ^T , G _R =(G _Rx ,G _Ry ) ^T ; considering the corrected image, set p(x,y) is a point on the left image, then the matching point corresponding to the parallax d on the right image is pd(xd,y). Using the two components of the gradient vector in the x and y directions, define the magnitude and phase angle of the gradient vector: $\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; =</span>\sqrt{{\mathrm{<span\; class="notranslate">\; G</span>}}_{\mathrm{<span\; class="notranslate">\; x</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; G</span>}}_{\mathrm{<span\; class="notranslate">\; the\; y</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}$ The modulus m of the gradient represents the rate of change of the gray level, and the phase angle Characterizing the direction when the rate of gray change is the largest, they provide different information about the pixel neighborhood and have different invariance to illumination distortion. The input image's gain distortion affects the gradient mode, while the phase angle does not change, but neither of them is affected by the bias distortion. Therefore, it is more beneficial to control the sensitivity of the method to noise if the mode and phase angle of the gradient are considered separately. The present invention adopts the linear combination of modulus and phase angle as the matching cost function, so as to maximize the utilization of gradient information. The expression is as follows: where m _c , Respectively represent the gradient vector modulus and phase angle corresponding to the three channels of the color image R, G, and B, and α is the weighting coefficient. Since the phase angle is π as a period, it needs to be normalized to a single period, so define f(x): $\mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\left\{\begin{array}{cc}\mathrm{<span\; class="notranslate">\; x</span>}& \mathrm{<span\; class="notranslate">\; 0</span>}<span\; class="notranslate">\; \&le;</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; \&le;</span>\mathrm{<span\; class="notranslate">\; \&pi;</span>}\\ \mathrm{<span\; class="notranslate">\; 2</span>}\mathrm{<span\; class="notranslate">\; \&pi;</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; x</span>}& \mathrm{<span\; class="notranslate">\; \&pi;</span>}<span\; class="notranslate">\; <</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; <</span>\mathrm{<span\; class="notranslate">\; 2</span>}\mathrm{<span\; class="notranslate">\; \&pi;</span>}\end{array}\right.$ Due to the introduction of the weighting coefficient α, we can change the robustness of the method to illumination distortion and noise by adjusting the value of the parameter α. The smaller the α, the greater the influence of the phase, and the greater the α, the greater the influence of the modulus. Since different images will have different degrees of illumination distortion, it is necessary to determine the reasonable value range of α through experiments in practice. Since e(p,d) represents the original matching cost of a single pixel, there will still be some overly large outliers in the actual situation, which need to be excluded to improve the matching accuracy. A commonly used method is to use a truncated function, that is, to compare e(p,d) with a constant, and take the minimum value as the matching cost. This method improves the result very little, and the present invention adopts the Geman-McClure function to handle outliers: $\mathrm{<span\; class="notranslate">\; \&rho;</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\frac{{\mathrm{<span\; class="notranslate">\; x</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}{{\mathrm{<span\; class="notranslate">\; x</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; \&sigma;</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}$ When the input x exceeds a certain value, its influence on the output value will smoothly drop to 0, and the transformed matching cost will converge to 1, which can be controlled by the parameter σ. Therefore, no matter how much the original matching cost of the input is, after the Geman-McClure function transformation, its output value will not exceed 1. 3. according to a kind of stereo matching method based on improved gradient and adaptive window described in claim 1, it is characterized in that: described step 2 and step 3 application improved cross adaptive window generation method, concrete process is as follows: First, according to the image structure and color information, determine a cross area of the current pixel p to be matched. The cross area includes horizontal and vertical directions, denoted by H(p) and V(p) respectively, and the size of the area consists of four directions arm length Determined, and can be adaptively changed according to the structure and color information of the image. by For example, the criteria for judging the arm length are as follows: 1. D _c (p _i ,p)<τ ₁ and D _c (p _i ,p _i +(1,0))<τ ₁ ; 2.D _s (p _i ,p)<L ₁ ; 3. D _c (p _i ,p)<τ ₂ , L ₂ <D _s (p _i ,p)<L ₁ . Among them, D _s (p _i ,p) is the spatial distance difference between pixels p _i and p; D _c (p _i ,p) is the color difference, defined as τ ₁ >τ ₂ , L ₁ >L ₂ , are preset color thresholds and distance thresholds. Criterion 1 not only limits the color difference between p _i and p, but also requires the color difference between p _i and its right adjacent pixel p _i + (1,0) to be less than τ ₁ , so as to avoid the arm length crossing the boundary area; Criteria 2 and 3 relax the range of arm length, and a larger window can be obtained by using a larger distance threshold L ₁ in low-textured areas; and when the arm length exceeds the preset value L ₂ , a stricter threshold τ ₂ will be used To ensure that the arm length is only extended in low-textured areas with very similar colors, so that the windows of high-textured areas and depth discontinuous areas will not be too large. Using the above method, the four arm lengths can be determined respectively In turn, the orthogonal crossing areas H(p) and V(p) are obtained: $\left\{\begin{array}{c}\mathrm{<span\; class="notranslate">\; h</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span><span\; class="notranslate">\; \{</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">\; x</span>}<span\; class="notranslate">\; \&Element;</span><span\; class="notranslate">\; [</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; p</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; h</span>}}_{\mathrm{<span\; class="notranslate">\; p</span>}}^{<span\; class="notranslate">\; -</span>}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; p</span>}}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; h</span>}}_{\mathrm{<span\; class="notranslate">\; p</span>}}^{<span\; class="notranslate">\; +</span>}<span\; class="notranslate">\; ]</span><span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; the\; y</span>}}_{\mathrm{<span\; class="notranslate">\; p</span>}}<span\; class="notranslate">\; \}</span>\\ \mathrm{<span\; class="notranslate">\; V</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span><span\; class="notranslate">\; \{</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">\; x</span>}<span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; p</span>}}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; \&Element;</span><span\; class="notranslate">\; [</span>{\mathrm{<span\; class="notranslate">\; the\; y</span>}}_{\mathrm{<span\; class="notranslate">\; p</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; v</span>}}_{\mathrm{<span\; class="notranslate">\; p</span>}}^{<span\; class="notranslate">\; -</span>}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; the\; y</span>}}_{\mathrm{<span\; class="notranslate">\; p</span>}}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; v</span>}}_{\mathrm{<span\; class="notranslate">\; p</span>}}^{<span\; class="notranslate">\; +</span>}<span\; class="notranslate">\; ]</span><span\; class="notranslate">\; \}</span>\end{array}\right.$ Finally, repeat the above process for each pixel q in V(p) along the vertical direction, and obtain the adaptive area of any pixel p in the image as: $\mathrm{<span\; class="notranslate">\; u</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\underset{\mathrm{<span\; class="notranslate">\; q</span>}<span\; class="notranslate">\; \&Element;</span>\mathrm{<span\; class="notranslate">\; V</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; )</span>}{<span\; class="notranslate">\; \&cup;</span>}\mathrm{<span\; class="notranslate">\; h</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; q</span>}<span\; class="notranslate">\; )</span>$ The present invention will symmetrically consider the respective local regions of support of the left and right images. For the two corresponding matching points p(x,y) and pd(x-d,y) in the left and right images, the above method can be used to generate adaptive regions U(p) and U'(pd) respectively, and their joint common region Areas of support identified as final: U _d (p)={(x,y)|(x,y)∈U(p),(xd,y)∈U'(pd)} Then, the original single-pixel matching cost is aggregated in the joint support region to obtain the total cost in the region: ${\mathrm{<span\; class="notranslate">\; E.</span>}}_{\mathrm{<span\; class="notranslate">\; d</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; N</span>}}\underset{\mathrm{<span\; class="notranslate">\; q</span>}<span\; class="notranslate">\; \&Element;</span>{\mathrm{<span\; class="notranslate">\; u</span>}}_{\mathrm{<span\; class="notranslate">\; d</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; )</span>}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}\mathrm{<span\; class="notranslate">\; e</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; q</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; d</span>}<span\; class="notranslate">\; )</span>$ In the formula, N is the total number of pixels in the aggregation area U _d (p). Finally, the "Winner-Takes-All (WTA)" strategy is adopted, and the point with the smallest matching cost is selected as the matching point in the disparity interval to select the disparity of point p to obtain the initial disparity: ${\mathrm{<span\; class="notranslate">\; d</span>}}_{\mathrm{<span\; class="notranslate">\; p</span>}}^{\mathrm{<span\; class="notranslate">\; 0</span>}}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; arg</span>}\underset{\mathrm{<span\; class="notranslate">\; 0</span>}<span\; class="notranslate">\; \&le;</span>\mathrm{<span\; class="notranslate">\; d</span>}<span\; class="notranslate">\; \&le;</span>{\mathrm{<span\; class="notranslate">\; d</span>}}_{\mathrm{<span\; class="notranslate">\; max</span>}}}{\mathrm{<span\; class="notranslate">\; min</span>}}{\mathrm{<span\; class="notranslate">\; E.</span>}}_{\mathrm{<span\; class="notranslate">\; d</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; )</span>$ Where d represents the possible disparity in the disparity space, and its value is generally an integer between 0 and the maximum disparity d _max . 4. according to a kind of stereo matching method based on improved gradient and adaptive window described in claim 1, it is characterized in that: described step 4 proposes the parallax refinement method based on local parallax histogram, and concrete calculation process is as follows: For a certain pixel p in the disparity map, construct a local disparity histogram in its neighborhood with it as the center Counts the number of occurrences of each disparity value within the domain. A peak will appear in the histogram, representing the maximum number of occurrences of parallax. The disparity value corresponding to this peak is the optimal disparity value in the statistical sense The present invention uses this optimal value to replace the initial disparity of pixel p This process utilizes the adaptive window generated in the previous step as a neighborhood of pixels, thus incurring no additional computational burden. In addition, the present invention performs three iterations on this process to make the optimal parallax value more accurate. Then, the left-right consistency test is used to detect the still existing mismatch points, and the points that pass the consistency test are marked as valid points, otherwise they are invalid points. For the detected mismatching points, scan the first effective point on the left and right in the horizontal direction, and use the value of the smaller parallax between the two to assign a value to the mismatching point.