CN103491361B - A kind of method improving sparse corresponding points images match precision and stereo image correction - Google Patents

Abstract

The invention discloses a kind of method improving sparse corresponding points images match precision and stereo image correction, comprise the following steps: step 1, improve the matching precision of sparse corresponding points: local image blocking is amplified, utilizes block SSD matching process to find out minimum SSD value and corresponding center point coordinate; Step 2, stereo image correction: the preliminary treatment of corresponding points coordinate, SVD method is utilized to solve rear 6 coefficients of Hg matrix and insert front 3 coefficients, carry out anti-preliminary treatment to Hg matrix and try to achieve H matrix, recycling H matrix will wherein be corrected into and only there is horizontal parallax with another figure by a figure.Have the while of improving sparse corresponding points images match precision and obviously can not increase search time, three-dimensional image correction method implementing procedure is simple, without the need to demarcate and corresponding points do not require at grade and decrease the advantages such as the search time in dense corresponding point matching process.

Description

A kind of method improving sparse corresponding points images match precision and stereo image correction

Technical field

The present invention relates to a kind of image procossing and computer vision technique, particularly a kind of method improving sparse corresponding points images match precision and stereo image correction.

Background technology

For the sparse features Point matching of two images, mainly contain two kinds of solutions: 1, manually find corresponding points and mate; 2, the automated characterization point matching algorithms such as Harris, SUSAN isocenter detection algorithm or SIFT, SURF are utilized to find corresponding points and mate.

At computer vision field, stereo image correction corrects also referred to as polar curve, all corresponding points of two images are made only to there is horizontal parallax to a wherein width (or two width) image rectification, reduce the search time in dense corresponding point matching process, mainly contain three kinds of solutions: 1, video camera is demarcated, the outer ginseng R of the internal reference K of video camera and diverse location, T-phase is utilized to combine, ask for required homography matrix H to correct (the method needs to demarcate video camera, and corresponding points do not require at grade) picture; 2, solving of homography matrix H is carried out to the image that only there is single plane or plane at infinity, and correct the image (the method without the need to demarcating, but can not ensure that out-of-plane image is corrected equally, and corresponding points need to limit at grade) of this plane; 3, set up the majorized function of fundamental matrix F and corresponding points, utilize the method for nonlinear optimization to carry out parameter optimization, finally solve homography matrix H(the method without the need to demarcating, corresponding points do not require at grade yet).

In fact, original sparse corresponding point matching method exists artificial coupling may have +the not high enough problem with there is error hiding during automated characterization Point matching of precision of 2 pixel error.Original part method for correcting polar line exists to be needed camera calibration or corresponding points requirement problem at grade.

Summary of the invention

The object of the invention is to overcome the shortcoming of prior art and deficiency, a kind of method improving sparse corresponding points images match precision and stereo image correction is provided, the method is that a kind of parallel optical axis structure camera is under operational environment, improve sparse corresponding point matching precision and without the need to demarcate, and corresponding points do not require in conplane three-dimensional image correction method, the camera of the method is for the formation of parallel optical axis structure.

Object of the present invention is achieved through the following technical solutions: a kind of method improving sparse corresponding points images match precision and stereo image correction, comprises the following steps:

Binocular camera shooting at least two images of step 1, use parallel optical axis structure;

Step 2, automated characterization point is utilized to detect and matching algorithm or manually choose the sparse corresponding points of at least 5 group, record corresponding points coordinate;

Step 3, improve the matching precision of sparse corresponding points: local image blocking is amplified, utilizes block SSD matching process to find out minimum SSD value and corresponding center point coordinate;

Step 4, stereo image correction: the preliminary treatment of corresponding points coordinate, utilize SVD method to solve H _grear 6 coefficients of matrix also insert front 3 coefficients, to H _gmatrix carries out anti-preliminary treatment and tries to achieve H matrix, and recycling H matrix will wherein be corrected into and only there is horizontal parallax with another figure by a figure.

Described step 3 comprises the following steps:

The amplification of A, topography's square: utilize the brightness of bilinear interpolation image blocking centered by corresponding points by left figure and right two figure or color value to carry out the partial enlargement of image;

B, block SSD mate: by the right figure square after interpolation, travel through SSD coupling, try to achieve the minimum value of SSD in the left figure after interpolation, and records center point coordinates; Left figure and right figure interchangeable, without loss of generality, with the corresponding points of right figure for ideal value, in left figure, improve corresponding point matching precision is example.

Described steps A comprises and comprising the following steps:

A1, by the i*i(of right figure centered by corresponding points as 11*11) image brightness of size or color square, utilize bilinear interpolation to obtain (2i ?1) * (2i ?1) (as 21*21) image blocking that wide and height respectively amplifies 2 times,

A2, (i+2j) * (i+2j) (as the 21*21) image blocking centered by selected corresponding points is extracted to left figure; And utilize bilinear interpolation to obtain (2i+4j ?1) * (2i+4j ?1) (as 41*41) image blocking that wide and height respectively amplifies 2 times, wherein, the span of i is 5≤i≤15, and the span of j is: 2≤j≤10.

Described step B comprises the following steps:

B1, by the right figure square after interpolation, the order traversal carrying out onesize (2i ?1) * (2i ?1) in the large square of left figure after interpolation compares, and tries to achieve the minimum value of SSD;

B2, records center point coordinates.

Described step 4 comprises the following steps:

(1) corresponding points coordinate preliminary treatment;

(2) SVD method is utilized to solve H _grear 6 coefficients of matrix also insert front 3 coefficients;

(3) to H _gmatrix carries out anti-preliminary treatment and tries to achieve H matrix;

(4) all corresponding points coordinates of image in public domain are calculated according to H matrix bilinear interpolation;

(5) image after utilizing corresponding points coordinate and bilinear interpolation to obtain a wherein figure correction.

In described step (1), the preliminary treatment of described corresponding points coordinate has following three kinds of selections:

A () does not process;

(b) center translation x _zl=x _l-E (x _l), y _zl=y _l-E (y _l), x _zr=x _r-E (x _r), y _zr=y _r-E (y _r);

(c) normalization x _gl=(x _l-E (x _l))/D (x _l), y _gl=(y _l-E (y _l))/D (y _l), x _gr=(x _r-E (x _r))/D (x _r), y _gr=(y _r-E (y _r))/D (y _r);

Wherein, subscript l represents Zuo Tu, subscript r represents right figure, and subscript z represents center translation process, and subscript g represents normalized, x is the abscissa of point, y is the ordinate of point, and E (x) is the average of abscissa before preliminary treatment, and D (x) is the variance of abscissa before preliminary treatment, E (y) is the average of ordinate before preliminary treatment, and D (y) is the variance of ordinate before preliminary treatment.

In described step (2), to a correct image any in left figure and right figure, if correct left figure, then according to the image y coordinate identical structure equation Ah=0 after correction namely:

$[x_{\mathrm{gl}},y_{\mathrm{gl}},1,-x_{\mathrm{gl}}{y}_{\mathrm{gr}},-y_{\mathrm{gl}}{y}_{\mathrm{gr}},-y_{\mathrm{gr}}]\left[\begin{array}{c}{h}_{g4}\\ h_{g5}\\ h_{g6}\\ h_{g7}\\ h_{g8}\\ h_{g9}\end{array}\right]=0,$

A matrix is carried out SVD and decompose A=UDV ^t, try to achieve pretreated homography matrix:

$H_g=\left[\begin{array}{ccc}{h}_{g1}& h_{g2}& h_{g3}\\ h_{g4}& h_{g5}& h_{g6}\\ h_{g7}& h_{g8}& h_{g9}\end{array}\right]$ Rear 6 coefficients be last row of V, and all coefficient will amplify simultaneously makes h _g9=1; Add front 3 coefficient [h _g1h _g2h _g3]=[1 0 0] or [h _g1h _g2h _g3]=[h _g5-h _g40].

In described step (3), for the preliminary treatment carried out, to H _gmatrix carries out anti-preliminary treatment, tries to achieve required homography matrix H:

(I) does not process;

The translation of (II) anticentre $T_{\mathrm{zl}}=\left[\begin{array}{ccc}1& 0& -E\left(x_l\right)\\ 0& 1& -E\left(y_l\right)\\ 0& 0& 1\end{array}\right],T_{\mathrm{zr}}=\left[\begin{array}{ccc}1& 0& -E\left(x_r\right)\\ 0& 1& -E\left(y_r\right)\\ 0& 0& 1\end{array}\right],H=T_{\mathrm{zr}}^{-1}{H}_g{T}_{\mathrm{zl}};$

(III) renormalization: $T_{\mathrm{gl}}=\left[\begin{array}{ccc}1/D\left(x_l\right)& 0& -E\left(x_l\right)\\ 0& 1/D\left(y_l\right)& -E\left(y_l\right)\\ 0& 0& 1\end{array}\right],$

$T_{\mathrm{gr}}=\left[\begin{array}{ccc}1/D\left(x_r\right)& 0& -E\left(x_r\right)\\ 0& 1/D\left(y_r\right)& -E\left(y_r\right)\\ 0& 0& 1\end{array}\right],H=T_{\mathrm{gr}}^{-1}{H}_g{T}_{\mathrm{gl}},$

If correct right figure, its way is with to the correction of left figure in like manner;

Described step (4) comprises the following steps:

(4-1) bilinear interpolation is carried out one by one according to the square frame of 16*16, f (1), f (2), f (3), f (4) refers to the mapping point of square frame four vertex positions respectively, inner at each square frame, utilizes formula:

F ₁₂(k, 1)=(f (1) * (16-k)+f (2) * k) >>2, f ₂₃(16, k)=(f (2) * (16-k)+f (3) * k) >>2, f ₃₄(k, 16)=(f (3) * (16-k)+f (4) * k) >>2, f ₄₁(1, k)=(f (4) * (16-k)+f (1) * k) point coordinates of the first interpolation four edges of >>2;

(4-2) formula is utilized:

The point coordinates of the interior zone of f (x, y)=(f (1) * (16-x) * (16-y)+f (2) * (x) * (16-y)+f (3) * (16-x) * (y)+f (4) * x*y) >>3 interpolation square frame;

Wherein, subscript represents first summit on interpolation limit and the mark on the second summit, and k representative to be positioned on interpolation limit and to be the coordinate of k-1 pixel with first vertex distance on this interpolation limit, and the span of k is 2 ~ 15; X, y represent the coordinate in square frame, and the span of x and y value is 2 ~ 15.The implication of formula is according to all corresponding points coordinates in H matrix interpolation computed image public domain.

In described step (5), the image after utilizing corresponding points coordinate and bilinear interpolation to obtain left figure correction; At brightness or the color value of the most contiguous 4 the corresponding points coordinates of impact point, obtain brightness under required rounded coordinate or color value with bilinear interpolation method; Utilize formula:

f(x,y)=(f(1)*(64-x)*(64-y)+f(2)*(x)*(64-y)+f(3)*(64-x)*(y)+f(4)*x*y)>>12，

Wherein, f (1), f (2), f (3) and f (4) refers to brightness or the color value of 4 the most contiguous coordinates respectively.

This programme specifically can be described below:

First two width images are manually looked for a little or automated characterization point detect and coupling, obtain being greater than 5 groups of sparse corresponding points; Then interpolation is carried out to topography, carry out the essence coupling of corresponding points; Finally carry out not requiring at conplane stereo image correction without the need to demarcation, corresponding points.

Specifically comprise following key step:

(1) manually or automatically choose image and be no less than arbitrarily 5 groups of corresponding points, obtain the coordinate of corresponding points;

(2) bilinear interpolation is carried out to the brightness of topography or color value, carry out the essence coupling of corresponding points.Think that right figure corresponding points position is standard, the corresponding points that left figure finds must with real left figure corresponding points very close to (as error exists +in 2 pixels), and due to camera be parallel optical axis structure, block SSD can be utilized to mate;

(2.1) first the i*i(of right figure centered by corresponding points as 11*11) image brightness of size or color square, utilize bilinear interpolation to obtain (2i-1) * (2i-1) (as 21*21) image blocking that wide and height respectively amplifies 2 times.Search at the original left figure image blocking of onesize i*i, hunting zone be near corresponding points j*j(as 5*5) square in, namely amplify after 2j*2j(as 10*10) square in;

(2.2) then (i+2j) * (i+2j) (as the 21*21) image blocking centered by selected corresponding points is extracted to left figure; And utilize bilinear interpolation to obtain (2i+4j-1) * (2i+4j-1) (as 41*41) image blocking that wide and height respectively amplifies 2 times;

(2.3) by the right figure square after interpolation, the order traversal carrying out onesize (2i-1) * (2i-1) in the large square of left figure after interpolation compares, and tries to achieve the minimum value of SSD, and records center point coordinates, then now the precision of corresponding points is +0.5 pixel.The comparison formula of SSD is $s=\underset{1}{\overset{2i}{\mathrm{\Σ}}}\underset{1}{\overset{2i}{\mathrm{\Σ}}}{(f(m,n)-g(m,n))}^2;$

(3) carry out not requiring at conplane stereo image correction without the need to demarcation, corresponding points.To any correct image of left and right two figure, might as well think to correct left figure, if correct way in like manner to right figure.Left figure corresponding points coordinate [x ₁, y ₁], right figure corresponding points coordinate [x _r, y _r],

$H_g=\left[\begin{array}{ccc}{h}_{g1}& h_{g2}& h_{g3}\\ h_{g4}& h_{g5}& h_{g6}\\ h_{g7}& h_{g8}& h_{g9}\end{array}\right];$

(3.1) preliminary treatment of corresponding points coordinate: (a) does not process; (b) center translation x _zl=x _l-E (x _l), y _zl=y _l-E (y _l), x _zr=x _r-E (x _r), y _zr=y _r-E (y _r); (c) normalization x _gl=(x _l-E (x _l))/D (x _l), y _gl=(y _l-E (y _l))/D (y _l), x _gr=(x _r-E (x _r))/D (x _r), y _gr=(y _r-E (y _r))/D (y _r).Wherein, E (x) is the average of coordinate before normalization, and D (x) is the variance of coordinate before normalization;

(3.2) method of SVD is utilized to solve rear 6 coefficients of H matrix, according to the image y coordinate identical structure equation Ah=0 after correction namely:

$[x_{\mathrm{gl}},y_{\mathrm{gl}},1,-x_{\mathrm{gl}}{y}_{\mathrm{gr}},-y_{\mathrm{gl}}{y}_{\mathrm{gr}},-y_{\mathrm{gr}}]\left[\begin{array}{c}{h}_{g4}\\ h_{g5}\\ h_{g6}\\ h_{g7}\\ h_{g8}\\ h_{g9}\end{array}\right]=0;$

A matrix carries out SVD and decomposes A=UDV ^t, try to achieve homography matrix H _grear 6 coefficients be V last row, adding front 3 coefficients is [h _g1h _g2h _g3]=[1 0 0] or [h _g1h _g2h _g3]=[h _g5-h _g40];

(3.3) for the preliminary treatment carried out, to H _gmatrix carries out anti-preliminary treatment, tries to achieve homography matrix H;

In described step (1), the preliminary treatment of described corresponding points coordinate has following three kinds of selections:

A () does not process;

The translation of (b) anticentre: $T_{\mathrm{zl}}=\left[\begin{array}{ccc}1& 0& -E\left(x_l\right)\\ 0& 1& -E\left(y_l\right)\\ 0& 0& 1\end{array}\right],T_{\mathrm{zr}}=\left[\begin{array}{ccc}1& 0& -E\left(x_r\right)\\ 0& 1& -E\left(y_r\right)\\ 0& 0& 1\end{array}\right],H=T_{\mathrm{zr}}^{-1}{H}_g{T}_{\mathrm{zl}};$

(c) renormalization: $T_{\mathrm{gl}}=\left[\begin{array}{ccc}1/D\left(x_l\right)& 0& -E\left(x_l\right)\\ 0& 1/D\left(y_l\right)& -E\left(y_l\right)\\ 0& 0& 1\end{array}\right],$

$T_{\mathrm{gr}}=\left[\begin{array}{ccc}1/D\left(x_r\right)& 0& -E\left(x_r\right)\\ 0& 1/D\left(y_r\right)& -E\left(y_r\right)\\ 0& 0& 1\end{array}\right],H=T_{\mathrm{gr}}^{-1}{H}_g{T}_{\mathrm{gl}},$

(3.4) all corresponding points of image in public domain are calculated according to H matrix interpolation.For improving arithmetic speed, the square frame of every 16*16 just uses homography matrix to obtain 4 groups of corresponding points coordinates, and method and 4 groups of corresponding points coordinates of all the other corresponding points coordinate bilinear interpolations obtain.Bilinear interpolation method carries out one by one according to the square frame of 16*16, f (1), f (2), and f (3), f (4) refer to the mapping point of square frame four vertex positions respectively.Inner at each square frame, utilize formula f ₁₂(k, 1)=(f (1) * (16-k)+f (2) * k) >>2, f ₂₃(16, k)=(f (2) * (16-k)+f (3) * k) >>2, f ₃₄(k, 16)=(f (3) * (16-k)+f (4) * k) >>2, f ₄₁(1, k)=(f (4) * (16-k)+f (1) * k) the first interpolation four edges of >>2, then formula is utilized: the interior zone of f (x, y)=(f (1) * (16-x) * (16-y)+f (2) * (x) * (16-y)+f (3) * (16-x) * (y)+f (4) * x*y) >>3 interpolation square frame.Wherein, subscript represents first summit on interpolation limit and the mark on the second summit, and k representative to be positioned on interpolation limit and to be the coordinate of k-1 pixel with first vertex distance on this interpolation limit, and k value is 2 ~ 15; X, y represent the coordinate in square frame, and value is 2 ~ 15;

(3.5) another image conversion is become the image having corresponding element with draft in public domain.The brightness of new images element or color value will be obtained at former figure by corresponding points coordinate, because corresponding points coordinate may have decimal, so need to utilize the pixel value of bilinear interpolation method under the pixel value of the most contiguous 4 coordinates of impact point obtains required rounded coordinate.Utilize the pixel value of bilinear interpolation method under the pixel value of the most contiguous 4 the corresponding points coordinates of impact point obtains required rounded coordinate.Utilize formula:

f(x,y)=(f(1)*(64-x)*(64-y)+f(2)*(x)*(64-y)+f(3)*(64-x)*(y)+f(4)*x*y)>>12，

Wherein, f (1), f (2), f (3) and f (4) refers to the pixel value of 4 the most contiguous coordinates respectively.

Operation principle of the present invention:

1, for the sparse corresponding points of having mated, near corresponding points, utilize bilinear interpolation to obtain the image of partial enlargement, then the method utilizing Block-matching to search for selects brightness or the most similar square of color value in the image of partial enlargement, thus matching precision is brought up to+0.5 pixel.

2, stereo image correction utilizes the corresponding points of coupling to solve homography matrix H, and solving of this matrix makes improvements normalization direct linear transformation, introduces the constraint that y coordinate is constant, then utilizes this matrix to carry out a wherein figure and carry out mapping and resampling.

The present invention has following advantage and effect relative to prior art:

1, the meticulous matching process of corresponding points for parallel optical axis structure camera image of the present invention's employing, can improve the matching precision of corresponding points, will manually coupling may have +the matching precision of 2 pixels is brought up to +the matching precision of 0.5 pixel.

2, the method that topography amplifies and Block-matching is searched near corresponding points is utilized, instead of coupling after entire image is amplified, significantly can reduce search time, avoid producing error hiding simultaneously.

3, the three-dimensional image correction method implementing procedure that adopts of the present invention is simple, compared with the bearing calibration of part isometric image, without the need to carrying out camera calibration and corresponding points do not require at grade.

4, the three-dimensional image correction method of the present invention's employing, keep the feature of original three-dimensional image correction method, make all corresponding points of image after correcting only there is horizontal parallax, decrease the search time in dense corresponding point matching process, there is actual engineering significance.

Accompanying drawing explanation

Fig. 1 is the overall structure schematic diagram of an embodiment of the present invention.

Fig. 2 is the general flow chart of the specific embodiment of the invention.

Fig. 3 is the flow chart of the raising corresponding point matching precision of the specific embodiment of the invention.

Fig. 4 is the flow chart of the stereo image correction of the specific embodiment of the invention.

Embodiment

Below in conjunction with embodiment and accompanying drawing, the present invention is described in further detail, but embodiments of the present invention are not limited thereto.

Embodiment

The parallel optical axis structure of the present embodiment is binocular camera, as shown in Figure 1, a kind of overall structure figure of embodiment comprises binocular camera shooting camera lens 1, described binocular camera has the first pick-up lens and the second pick-up lens, first pick-up lens and the second pick-up lens are that parallel mode arranges, to ensure that the optical axis 4 of two cameras is parallel constructions; The corresponding transducer 3 of each camera; Light source 2 has enough illuminations in the process of guarantee collection image.

The general flow chart of raising corresponding points precision of the present invention and three-dimensional image correction method as shown in Figure 2; Concrete steps are: first take two pictures with parallel optical axis structure camera; Recycle the detection of automated characterization point and matching algorithm or artificial searching and be no less than 5 groups of corresponding points, corresponding points preferably can be evenly distributed on image each several part; Then to topography's bilinear interpolation, the essence coupling of corresponding points is carried out; Finally carry out without the need to demarcating, corresponding points are without the need at conplane stereo image correction.

As shown in Figure 3, for improving the flow chart of corresponding point matching precision; The first step be the image blocking bilinear interpolation of right figure centered by corresponding points is become wide, the high each image blocking amplifying 2 times; Second step the image blocking of left figure centered by corresponding points is added the scope bilinear interpolation of needs search becomes wide, high each image blocking amplifying 2 times; 3rd step right figure square and left figure square is carried out onesize order traversal compare, and tries to achieve minimum SSD numerical value, the coordinate of records center point; 4th step is repetition aforesaid operations, until all corresponding points all complete essence coupling.

As shown in Figure 4, be the flow chart of stereo image correction; The first step carries out the preliminary treatment of corresponding points, can select not process, center translation and normalization three kinds of modes, with normalized best results; Second step utilizes SVD method to solve H after preliminary treatment _grear 6 coefficients [h4, h5, h6, h7, h8, h9] of matrix, whole coefficient to amplify together and make h9=1, and add front 3 coefficients for [1,0,0] or [h5 ,-h4,0]; 3rd step is to H _gmatrix carries out anti-preliminary treatment (do not process, center translation, renormalization), tries to achieve required homography matrix H; 4th step is square frame image being divided into multiple 16*16, utilizes homography matrix H to calculate the corresponding points coordinate at all square frame 4 angles, and utilizes the method for bilinear interpolation to ask for all corresponding points coordinates in public domain; 5th step be utilize corresponding points coordinate and bilinear interpolation obtain a wherein figure carry out three-dimensional correction after image.

Above-described embodiment is the present invention's preferably execution mode; but embodiments of the present invention are not restricted to the described embodiments; change, the modification done under other any does not deviate from Spirit Essence of the present invention and principle, substitute, combine, simplify; all should be the substitute mode of equivalence, be included within protection scope of the present invention.

Claims (6)

1. improve a method for sparse corresponding points images match precision and stereo image correction, it is characterized in that, comprise the following steps:

Binocular camera shooting at least two images of step 1, use parallel optical axis structure;

Step 2, automated characterization point is utilized to detect and matching algorithm or manually choose the sparse corresponding points of at least 5 group, record corresponding points coordinate;

Step 3, improve the matching precision of sparse corresponding points: local image blocking is amplified, utilizes block SSD matching process to find out minimum SSD value and corresponding center point coordinate;

Step 4, stereo image correction: the preliminary treatment of corresponding points coordinate, utilize SVD method to solve H _grear 6 coefficients of matrix also insert front 3 coefficients, to H _gmatrix carries out anti-preliminary treatment and tries to achieve H matrix, and recycling H matrix will wherein be corrected into and only there is horizontal parallax with another figure by a figure;

Described step 3 comprises the following steps:

The amplification of A, topography's square: utilize the brightness of bilinear interpolation image blocking centered by corresponding points by left figure and right figure or color value to carry out the partial enlargement of image;

B, block SSD mate: by the right figure square after interpolation, travel through SSD coupling, try to achieve the minimum value of SSD in the left figure after interpolation, and records center point coordinates;

Described step 4 comprises the following steps:

(1) corresponding points coordinate preliminary treatment;

(2) SVD method is utilized to solve H _grear 6 coefficients of matrix also insert front 3 coefficients;

(3) to H _gmatrix carries out anti-preliminary treatment and tries to achieve H matrix;

(4) image is divided into the square frame of multiple 16*16, utilizes homography matrix H to calculate the corresponding points coordinate at all square frame 4 angles, and utilize the method for bilinear interpolation to ask for all corresponding points coordinates in public domain;

(5) utilize corresponding points coordinate and bilinear interpolation obtain a wherein figure carry out three-dimensional correction after image;

In described step (1), the preliminary treatment of described corresponding points coordinate has following three kinds of selections:

A () does not process;

(b) center translation: x _zl=x _l-E (x _l), y _zl=y _l-E (y _l), x _zr=x _r-E (x _r), y _zr=y _r-E (y _r);

(c) normalization: x _gl=(x _l-E (x _l))/D (x _l), y _gl=(y _l-E (y _l))/D (y _l), x _gr=(x _r-E (x _r))/D (x _r), y _gr=(y _r-E (y _r))/D (y _r);

Wherein, under be designated as l symbology data used belong to left figure, under be designated as r symbology data used belong to right figure, under be designated as the symbology center translation process of z, under be designated as the symbology normalized of g, x is the abscissa of point, y is the ordinate of point, E (x) is the average of abscissa before preliminary treatment, D (x) is the variance of abscissa before preliminary treatment, E (y) is the average of ordinate before preliminary treatment, and D (y) is the variance of ordinate before preliminary treatment;

In described step (2), to a correct image any in left figure and right figure, if correct left figure, then according to the image identical structure equation Ah=0 after correction, that is:

$[x_{\mathrm{gl}},y_{\mathrm{gl}},1,{-x}_{\mathrm{gl}}{y}_{\mathrm{gr}},{-y}_{\mathrm{gl}}{y}_{\mathrm{gr}},{-y}_{\mathrm{gr}}]\left[\begin{array}{c}{h}_{g4}\\ h_{g5}\\ h_{g6}\\ h_{g7}\\ h_{g8}\\ h_{g9}\end{array}\right]=0,$

A matrix is carried out SVD and decompose A=UDV ^t, try to achieve pretreated homography matrix:

$H_g=\left[\begin{array}{ccc}{h}_{g1}& h_{g2}& h_{g3}\\ h_{g4}& h_{g5}& h_{g6}\\ h_{g7}& h_{g8}& h_{g9}\end{array}\right]$ Rear 6 coefficients be last row of the Section 3 V that SVD decomposes, and all coefficient will amplify simultaneously makes h _g9=1; Add front 3 coefficient [h _g1h _g2h _g3]=[1 0 0] or [h _g1h _g2h _g3]=[h _g5-h _g40];

Wherein, under be designated as the symbology of g through pretreated data, under be designated as l symbology data used belong to left figure, under be designated as r symbology data used belong to right figure, H _gpretreated homography matrix is tried to achieve in representative; Formula Ah=0 is $y_{\mathrm{gr}}=H_g\left[\begin{array}{c}{x}_{\mathrm{gl}}\\ y_{\mathrm{gl}}\\ 1\end{array}\right]$ Expansion, it is equal with the y coordinate of right figure to be that order corrects rear left figure y coordinate.

2. the method for raising according to claim 1 sparse corresponding points images match precision and stereo image correction, is characterized in that, described steps A comprises the following steps:

A1, by the image brightness of the i × i size of right figure centered by corresponding points or color square, utilize bilinear interpolation to obtain (2i ?1) × (2i ?1) image blocking that wide and height respectively amplifies 2 times;

A2, (i+2j) × (i+2j) image blocking centered by selected corresponding points is extracted to left figure, and utilize bilinear interpolation to obtain (2i+4j ?1) × (2i+4j ?1) image blocking that wide and height respectively amplifies 2 times, wherein, the span of i is 5≤i≤15, and the span of j is: 2≤j≤10.

3. the method for raising according to claim 1 sparse corresponding points images match precision and stereo image correction, is characterized in that, described step B comprises the following steps:

B1, by the right figure square after interpolation, the order traversal carrying out onesize (2i ?1) × (2i ?1) in the large square of left figure after interpolation compares, and tries to achieve the minimum value of SSD;

B2, records center point coordinates.

4. the method for raising according to claim 1 sparse corresponding points images match precision and stereo image correction, is characterized in that, in described step (3), for the preliminary treatment carried out, to H _gmatrix carries out anti-preliminary treatment, tries to achieve required homography matrix H:

(I) does not process;

The translation of (II) anticentre: $T_{\mathrm{zl}}=\left[\begin{array}{ccc}1& 0& -E\left(x_l\right)\\ 0& 1& -E\left(y_l\right)\\ 0& 0& 1\end{array}\right],T_{\mathrm{zr}}=\left[\begin{array}{ccc}1& 0& -E\left(x_r\right)\\ 0& 1& -E\left(y_r\right)\\ 0& 0& 1\end{array}\right],H=T_{\mathrm{zr}}^{-1}{H}_g{T}_{\mathrm{zl}};$

(III) renormalization: $T_{\mathrm{gl}}=\left[\begin{array}{ccc}1/D\left(x_l\right)& 0& -E\left(x_l\right)\\ 0& 1/D\left(y_l\right)& -E\left(y_l\right)\\ 0& 0& 1\end{array}\right],$

$T_{\mathrm{gr}}=\left[\begin{array}{ccc}1/D\left(x_r\right)& 0& -E\left(x_r\right)\\ 0& 1/D\left(y_r\right)& -E\left(y_r\right)\\ 0& 0& 1\end{array}\right],H=T_{\mathrm{gr}}^{-1}{H}_g{T}_{\mathrm{gl}},$

Wherein, under be designated as l symbology data used belong to left figure, under be designated as r symbology data used belong to right figure, under be designated as the symbology center translation process of z, under be designated as the symbology normalized of g, x is the abscissa of point, y is the ordinate of point, E (x) is the average of abscissa before preliminary treatment, D (x) is the variance of abscissa before preliminary treatment, E (y) is the average of ordinate before preliminary treatment, and D (y) is the variance of ordinate before preliminary treatment.

5. the method for raising according to claim 1 sparse corresponding points images match precision and stereo image correction, is characterized in that, described step (4) comprises the following steps:

(4-1) bilinear interpolation according to 16 × 16 square frame carry out one by one, inner at each square frame, utilize formula:

F ₁₂(k, 1)=(f (1) * (16-k)+f (2) * k) >>2, f ₂₃(16, k)=(f (2) * (16-k)+f (3) * k) >>2, f ₃₄(k, 16)=(f (3) * (16-k)+f (4) * k) >>2, f ₄₁(1, k)=(f (4) * (16-k)+f (1) * k) >>2, the point coordinates of first interpolation four edges;

Wherein, k representative to be positioned on interpolation limit and to be the coordinate of k-1 pixel with first vertex distance on this interpolation limit, and the span of k is 2 ~ 15; X represents the abscissa in square frame, and y represents the ordinate in square frame, and the span of x and y is 2 ~ 15, and the implication of formula used is with all corresponding points coordinates in H matrix interpolation computed image public domain;

(4-2) formula is utilized:

F (x, y)=(f (1) * (16-x) * (16-y)+f (2) * (x) * (16-y)+f (3) * (16-x) * (y)+f (4) * x*y) >>3, the point coordinates of the interior zone of interpolation square frame.

6. the method for raising according to claim 1 sparse corresponding points images match precision and stereo image correction, is characterized in that, in described step (5), and the image after utilizing corresponding points coordinate and bilinear interpolation to obtain left figure correction; At brightness or the color value of the most contiguous 4 the corresponding points coordinates of impact point, the formula by bilinear interpolation:

F (x, y)=(f (1) * (64-x) * (64-y)+f (2) * (x) * (64-y)+f (3) * (64-x) * (y)+f (4) * x*y) >>12, obtains the brightness under required rounded coordinate or color value;

Wherein, f (1), f (2), f (3) and f (4) refers to brightness or the color value of 4 the most contiguous coordinates respectively.