CN103686187A - Estimation method for transformation domain overall high-precision motion vector - Google Patents

Abstract

The invention discloses an estimation method for a transformation domain overall high-precision motion vector. The method is mainly used for estimating the motion vector of a translation motion image sequence. The method comprises the steps of video extraction, multi-dimension matrix partitioning recombination, matrix conversion, coefficient dimensionality reduction, folding point extraction, window confirmation, windowing data screening, least square linear iterative fitting, and motion vector obtaining. An embodiment of the invention proves that the precision error of the method can be reduced to 10-4 magnitude orders. The higher the precision of the motion vector estimation is, the higher the precision of motion estimation is; when the difference value distribution is closer to zero, the energy of a difference value block is smaller; the less the bit rate of code streams generated finally is, the better the compression performance is. The method refers to a plurality of frames of images, has the advantages of being low in calculating complexity, high in motion estimation precision and consecutive in result, and effectively overcomes the defects that in an existing empty domain, a motion vector estimating method is optimized in local estimation result, disperse in estimation precision, and high in calculating complexity.

Description

A kind of transform domain overall situation high accuracy motion vectors method of estimation

Technical field

The present invention relates to digital image video coding techniques field, more specifically, a kind of overall high accuracy motion vectors method of estimation of transform domain in image/video compressed encoding.The estimation of motion vectors that it is characterized in that the overall continuous precise of transform domain, transform domain wherein, after referring to video matrix being carried out to the discrete cosine transform of multidimensional vector matrix, processes processing to the coefficient after its conversion, thereby realizes estimation of motion vectors.

Background technology

Video can be regarded as the gradual sequence that a series of static map picture frames form, therefore its time domain redundancy will be far longer than spatial domain redundancy, therefore, when the second two field picture is encoded on the basis of former frame, can first utilize the piece of the first frame to predict the second frame, find out the displacement of each piece in reference frame in the frame that will encode, be referred to as motion vector, and the residual error that motion vector and two pieces subtract each other is encoded, thereby removed to a great extent time domain redundancy.Wherein, in the process of coding side searching motion vector, be called estimation, and in decoding end, prediction piece and the superimposed process of prediction residual be called to motion compensation.

Estimation is the key technology of Video coding, its the most basic principle is the temporal correlation utilizing between consecutive frame, by prediction, reduce temporal redundancy, in actual coding, in order to save code check, do not transmit the total data of each frame, but utilize estimation to obtain the difference between each frame and its prediction reference frame.Estimation is more accurate, convergence and zero is got in the distribution of difference, and the energy of difference block is less, and the bit bit rate of the code stream producing after conversion, quantification and closely related coding is also fewer, therefore, the order of accuarcy of motion estimation search has directly had influence on the compression performance of coding.

On the other hand, estimation is a module the most consuming time in video compression coding system, can account for 60%～80% of coding computing total amount, it is the most intensive place of computing in cataloged procedure, want to improve the coding rate of video compression system, reach Real Time Compression, the spent time of estimation of necessary shortening time occupation proportion maximum.

Up to the present, all video encoding standards are all in spatial domain, to find locally best matching piece in reference frame to carry out estimation of motion vectors, up-to-date video encoding standard H.265 estimation of motion vectors is also to adopt local optimum block matching motion vector method of estimation, the method complexity is very high, and estimation of motion vectors precision is discretization, H.265 estimation of motion vectors precision is 1/4th pixels.

Summary of the invention

In order to overcome the deficiency of method for estimating motion vector local optimum in existing spatial domain, precision discretization, high complexity, the present invention is directed to translation video sequence, proposed transform domain method for estimating motion vector, the method is not only of overall importance, and complexity is low, precision is high and continuous, and estimation precision is higher, the distribution of difference more levels off to zero, and the energy of difference block is less, the last code stream bit bit rate producing is also fewer, and compression performance is better.

A transform domain overall situation high accuracy motion vectors method of estimation, at least comprises following step:

The first step: video selecting step

Choose yuv format video, for translation video, carry out estimation of motion vectors, Y is luminance component, and U, V are chromatic components, and human visual system is more responsive to brightness ratio colour, and therefore the main Y luminance component that extracts is as experimental data;

Second step: multidimensional piecemeal reconstitution steps

This step is mainly that source video data is carried out to piecemeal restructuring, is divided into several video submatrixs, and a minute block size generally has 64*64,32*32,16*16 etc.;

The 3rd step: matrixing step

Use multidimensional vector matrix discrete cosine transform matrix, three-dimensional submatrix is carried out to the discrete cosine transform of multidimensional vector matrix, calculate three-dimensional coefficient matrix, coefficient after conversion is also a three-dimensional matrice, the corresponding three-dimensional coordinate of coefficient of each in this three-dimensional matrice, the position that this coefficient just can be corresponding in three-dimensional system of coordinate represents with a dot;

The 4th step: coefficient dimensionality reduction step

In the three-dimensional coefficient matrix that previous step is calculated, be not that 0 coefficient shows with dot in three-dimensional system of coordinate, these dots mainly concentrate in a folding plane, two edges of folding plane are on three-dimensional coordinate left and right side, and these points concentrate on a folding straightway, extract this two sides of containing motion vector information, just can the processing of three-dimensional coefficient be dropped on two dimensional surface and be processed;

The 5th step: folding point extraction step

First, data point in extract in previous step two faces is carried out to longitudinal axis maximum and minimum value intercepting, then, calculate these points of intercepting to the distance of initial point, according to tell several folding points apart from size, finally calculate respectively near the abscissa average of point each folding point, using this abscissa as folding point;

The 6th step: window determining step

The folding point calculating according to previous step, further determines the scope of window by the abscissa of folding point;

The 7th step: windowing data screening step

If motion vector is integer pixel, each section of straight line is all complete, if but motion vector is non-integer pixel, and final stage is not complete, and for integer pixel, the window that adopts previous step to obtain intercepts out final stage data; For non-integer pixel, windowing intercepts segment data second from the bottom;

The 8th step: straight line iterative fitting step

The window obtaining by previous step carries out data point intercepting, more respectively remaining data is carried out to least-squares line iterative fitting, and the slope that calculates fitting a straight line is d ₁and d ₂;

The 9th step: obtain motion vector step

The straight slope d calculating according to previous step ₁and d ₂, therefore estimated motion vector is [± d ₁, ± d ₂].

Described second step multidimensional piecemeal restructuring mainly realizes according to the following steps:

(1) step, is defined as the origin of coordinates by the video Y component upper left corner, laterally scans M to the right ₁individual pixel, skips to the second row, laterally scans M to the right ₁individual pixel, repeats said process, until scanning M ₂oK, now obtain first M of Y the first frame ₁* M ₂piece;

(2) step, skips to Y the second frame, repeats first step process, obtains first M of Y the second frame ₁* M ₂piece, skips to Y the 3rd frame, the 4th frame successively until N ₁frame, obtains first M ₁* M ₂* N ₁video submatrix;

(3) step, skips back to Y the first frame, moves to right successively to second, realizes the piecemeal sampling of view picture Y frame, obtains all M of Y frame for the 3rd until last piece repeats first two steps ₁* M ₂* N ₁3 D video submatrix, M in the present invention ₁and M ₂equate.

The 3rd described step multidimensional vector matrix discrete cosine transform step, is that the 3 D video submatrix that restructuring obtains to multidimensional piecemeal carries out the discrete cosine transform of multidimensional vector matrix, and the formula of conversion is:

$B_{\mathrm{IJ}}=C_{\mathrm{II}}{A}_{\mathrm{IJ}}{C}_{\mathrm{JJ}}^T$

For image sequence after translation, vector I is two-dimensional vector, and J is one dimension scalar, and matrix A is 3-D view matrix, and B is the three-dimensional coefficient matrix having converted, and C is the nuclear matrix of multidimensional vector matrix discrete cosine transform, and T is multidimensional vector matrix transpose.

The 8th described step least-squares line iterative fitting step, mainly realizes by the following method:

(1) step, the data that windowing intercepting is obtained are carried out least squares line fitting, simulate straight line;

(2) step, calculates each point to the distance of this straight line, and obtains its average, according to this average, sets suitable step-length;

(3) step, setting threshold is that average deducts step-length k doubly, and k is since 0, and k of every iteration adds 1;

(4) step, by the threshold value of setting, reject the point that distance is greater than threshold value, remaining point is used least squares line fitting method again, simulates straight line, and the slope that the slope of this straight line and (1) step is simulated to straight line contrasts, if difference is less than set-point, stop iteration, derive the slope that (4) step is tried to achieve, if difference is very large, turn back to (3) step, down carry out successively.

Beneficial effect: the present invention is directed to translation image sequence, in this transform domain, overall motion estimation algorithm is not only with reference to multiple image, the more traditional spatial domain of estimation of motion vectors estimation of motion vectors is more accurate, and the precision of estimation of motion vectors of the present invention is high and continuous, and estimation precision is higher, the distribution of difference more levels off to zero, and the energy of difference block is less, the last code stream bit bit rate producing is also fewer, and compression performance is better.

Accompanying drawing explanation

Fig. 1 is all step block diagrams that technical solution of the present invention is implemented.

Fig. 2 is the flow chart that in Fig. 1, least-squares line iterative fitting step is implemented.

Fig. 3 is that video that gray level image of the present invention obtains after by vector [3,2] translation carries out three-dimensional coefficient figure after the discrete cosine transform of multidimensional vector matrix.

Fig. 4 is for extracting k ₁data point in=0 plane.

Fig. 5 is for extracting k ₂data point in=0 plane.

Fig. 6 is for extracting k ₁data point in=0 plane, through constantly rejecting, finally remaining point adopts least squares line fitting method, matching straight line out.

Fig. 7 is for extracting k ₂data point in=0 plane, through constantly rejecting, finally remaining point adopts least squares line fitting method, matching straight line out.

Embodiment:

Below in conjunction with the accompanying drawing in the embodiment of the present invention, the technical scheme in the embodiment of the present invention is carried out to clear, complete description, obviously, described embodiment is only the present invention embodiment ideally.

Multidimensional vector matrix discrete cosine transform theory and operational criterion that in the present invention, multidimensional vector matrix discrete cosine transform step adopts are prior art, and the multidimensional vector matrixing formula adopting in this cell matrix shift step is:

For the three-dimensional Y video submatrix after multidimensional piecemeal reconstitution steps

, by matrix dimension be divided into two groups, use respectively two vector I, J represents, wherein I=(k ₁, k ₂), J=(k ₃),

just can be expressed as A _iJ, then convert:

Multidimensional vector matrix direct transform formula is

$B_{\mathrm{IJ}}=C_{\mathrm{II}}{A}_{\mathrm{IJ}}{C}_{\mathrm{JJ}}^T$

In formula, B is the three-dimensional coefficient matrix having converted, and C is the nuclear matrix of multidimensional vector matrix discrete cosine transform, and T is multidimensional vector matrix transpose, according to multidimensional vector multiplication of matrices criterion, A _iJmiddle I is two-dimentional, so C _iImiddle I is also two-dimentional, C _iIbe four-matrix, J is one dimension, so

middle J is also one dimension, it is two-dimensional matrix.

The multidimensional vector orthogonal matrix that the multi-dimensional orthogonal matrix adopting in above-mentioned said matrixing step is discrete cosine transform, provides the concrete operations operator formula that this invention is used below:

The form of two-dimensional vector discrete cosine transform operation operator is

$C_{\mathrm{JJ}}=\left(c_{k_1{n}_1}\right)$

C wherein _jJin two J are one dimensions, use respectively k ₁, n ₁distinguish,

c _jJelement in two-dimensional matrix,

$c_{k_1{n}_1}={\left(\frac{2}{N_1}\right)}^{\frac{1}{2}}c\left(k_1\right)\cos \frac{({2n}_1+1)k_1\mathrm{\π}}{{2N}_1},c\left(k_1\right)=\left\{\begin{array}{cc}\frac{1}{\sqrt{2}}& k_1=0\\ 1& k_1=\mathrm{other}\end{array},c\left(n_1\right)=\left\{\begin{array}{cc}\frac{1}{\sqrt{2}}& n_1=0\\ 1& n_1=\mathrm{other}\end{array}\right.\right.,$

K ₁=0,1 ..., N ₁-1, n ₁=0,1 ..., N ₁-1, N ₁it is the frame number of video.

The form of four-vector discrete cosine transform operation operator is

$C_{\mathrm{II}}=\left(c_{k_1{k}_2{m}_1{m}_2}\right)$

C wherein _iIin two I be all two-dimentional, use respectively (k ₁k ₂) and (m ₁m ₂) distinguish,

c _iIelement in four-matrix,

$c\left(k_i\right)=\left\{\begin{array}{cc}\frac{1}{\sqrt{2}}& k_i=0\\ 1& k_i=\mathrm{othe}\end{array},c\left(m_i\right)=\left\{\begin{array}{cc}\frac{1}{\sqrt{2}}& m_i=0\\ 1& m_i=\mathrm{other}\end{array},\right.\right.$ K _i=0,1 ..., M _i-1, m _i=0,1 ..., M _i-1, i=1,2, M ₁, M ₂it is the size of piecemeal.

Least squares line fitting method principle and formula that the present invention uses are as follows:

The basic principle of least squares line fitting is exactly, and with discrete distribution and the coordinate points that totally presents straight path, obtains straight line parameter.If the band of straight line ginseng equation is:

y＝ax+b

Utilize the coordinate (x of certain Interval Discrete point on straight path _i, y _i) (i=1 ..., n), definition error term

principle of least square method requires Q to reach minimum value.The condition of Q minimum is: $\frac{\∂Q}{\∂}=\frac{\∂Q}{\∂b}=0,$ Can obtain:

$\left[\begin{array}{cc}\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{x}_i^2& -\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{x}_i\\ -\underset{i=1}{\overset{n}{\mathrm{\Σ}}}& n\end{array}\right]=\left[\begin{array}{c}a\\ b\end{array}\right]\left[\begin{array}{c}-\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{x}_i{y}_i\\ \underset{i=1}{\overset{n}{\mathrm{\Σ}}}{y}_i\end{array}\right]$

Matrix operation or direct solving equations method be can use, parameter a and b obtained

$a=\frac{n\underset{i=1}{\overset{n}{\mathrm{\Σ}}}\left(x_i{y}_i\right)-\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{x}_i\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{y}_i}{n\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{x}_1^2-{\left(\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{x}_i\right)}^2}$

$b=\frac{\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{x}_i^2\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{y}_i-\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{x}_i\underset{i=1}{\overset{n}{\mathrm{\Σ}}}\left(x_i{y}_i\right)}{n\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{x}_i^2-{\left(\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{x}_i\right)}^2}$

The theoretical foundation that technical solution of the present invention is implemented is: the method is that the video of yuv format is extracted to the three-dimensional sub-block of Y component according to above-mentioned multidimensional piecemeal reconstitution steps, then carry out the discrete cosine transform of multidimensional vector matrix, after conversion, coefficient produces special energy plane.The first situation, required translation vector [d ₁, d ₂] middle d ₁and d ₂all be less than or equal to 1 and be not at 0 o'clock, this time, plane can not produce foldingly, and this plane expression formula is k ₁d ₁+ k ₂d ₂+ k ₃=0, k wherein ₁, k ₂, k ₃for reference axis; The second situation, if required translation vector [d ₁, d ₂] in have at least one to be greater than 1, now plane can produce Fold, first's plane expression formula is still k ₁d ₁+ k ₂d ₂+ k ₃=0, the initial point of this folding plane by frequency domain and vertical with the direction of motion vector, afterwards folding plane and the previous plane vertical mirror that is connected.For both of these case, get respectively k ₁=0 and k ₂=0 two plane, at this moment can obtain k ₂d ₂+ k ₃=0 and k ₁d ₁+ k ₃=0, the plane by frequency domain initial point just becomes the straightway by initial point before, may occur Fold, occur when folding, folding straightway respectively with straight line vertical mirror the last period, then utilize the data point windowing fitting a straight line of these two planes to obtain d ₁and d ₂, because no matter image is toward which direction motion, first plane of folding plane is by the initial point of frequency domain, and d ₁and d ₂perseverance is more than or equal to 0, and the motion vector that estimate is [± d ₁, ± d ₂] these four kinds of situations.

In order to verify precision of the present invention and to describe specific embodiment of the invention flow process in detail, we take ideal situation as example, and the image sequence obtaining by a width gray level image translation, as video source, further illustrates technology contents of the present invention by reference to the accompanying drawings:

The first step, for the characteristic of verifying that the technical program precision is high and continuous, the present embodiment adopts is the image sequence that gray level image obtains after by known vector teranslation, the theoretical value of motion vector can give experimental result as reference, and what adopt in actual applications is the video of yuv format, extract Y luminance component, their effects are the same.First, the gray level image of a width 64*64 (being equivalent to Y luminance component in yuv format video) is moved 63 times by vector [3,2], obtain 64 frame image sequence.

Second step, the image sequence that step 1 is obtained carries out the restructuring of multidimensional piecemeal: (1) step, the video Y component upper left corner is defined as to the origin of coordinates, laterally scan 64 pixels to the right, skip to the second row, laterally scan 64 pixels to the right, repeat said process, until scan 64 row, now obtain first 64*64 piece of Y the first frame.(2) step, skips to Y the second frame, repeats first step process, obtains first 64*64 piece of Y the second frame, skips to successively Y the 3rd frame, the 4th frame until the 64th frame, obtains the video submatrix of first 64*64*64.(3) step, skips back to Y the first frame, moves to right successively to second, realizes the piecemeal sampling of view picture Y frame, obtains all 64*64*64 3 D video submatrixs of Y frame for the 3rd until last piece repeats first two steps.

The 3rd step, image sequence after piecemeal restructuring is carried out to the discrete cosine transform of multidimensional vector matrix, coefficient after conversion is also a three-dimensional matrice, the corresponding three-dimensional coordinate of coefficient of each in this three-dimensional matrice, and this coefficient just can be at three-dimensional system of coordinate k ₁k ₂k ₃the position of middle correspondence represents with a dot, by each in three-dimensional coefficient matrix, is not that 0 coefficient is at three-dimensional system of coordinate k ₁k ₂k ₃in show, these dots mainly concentrate in a folding plane, result as shown in Figure 3.

The 4th step, the 3 dimensional drawing by three-dimensional matrice coefficient forms after conversion, converts two-dimentional plane to, extracts k ₁dot in=0 plane, these dots concentrate on a folding straightway, as shown in Figure 4; Extract k ₂dot in=0 plane, these dots also concentrate on a folding straightway, as shown in Figure 5 equally.

The 5th step, asks folding point.In Fig. 4, first, ordinate is reached to peaked point and extract, the average that then calculates the abscissa of these points is 31.5372, and this average is folding point; In Fig. 5, the abscissa that calculates folding point that uses the same method is 21.2134 and 41.3573.

Six, seven steps, determine the scope of window, and carry out data cutout.Corresponding diagram 4, the scope of window is set as [32,63], extracts final stage data point; Corresponding diagram 5, now the scope of window is set as [41,63], intercepting final stage data.

Eight, nine steps, windowing straight line iterative fitting: (1) step, the data after windowing intercepting first adopt least squares line fitting method, simulate straight line, and the straight slope of corresponding diagram 4 matching is for the first time-1.7866.(2) step, the range averaging value that corresponding diagram 4 is obtained a little this straight line is 3.7812, and step-length is probably set as 1/10th of average, and the step-length that Fig. 4 is corresponding is set as 0.25; (3) step, rejects the point that distance is greater than 3.7812, and remaining point carries out least squares line fitting again, and the slope of the straight line of this time matching is 3.0615; (4) step, then rejects the point that is greater than 3.5312 to air line distance, then follows least squares line fitting with remaining point, and the fitting a straight line that so iterates, until the absolute value of the difference of the slope of trying to achieve for twice is less than at 0.0002 o'clock, stops iteration.Fig. 6 is the straight line of corresponding diagram 4 last matchings, and its slope is that-1.9994, Fig. 7 is the straight line of corresponding diagram 5 last matchings, slope is 3.0000, and the motion vector therefore estimating is [± 3.0000, ± 1.9994], theoretical value is [3,2], and the trueness error of estimation can drop to 10 ^-4magnitude.

The present invention carries out estimation of motion vectors in multidimensional transform territory.Due to traditional spatial domain estimation of motion vectors not only precision be discrete, H.265 middle brightness movement vector is estimated as 1/4th pixel precisions, and operand is also large especially while finding local optimum match block, the multidimensional vector matrix discrete cosine transform domain estimation that the present invention proposes, adopt dimensionality reduction least-squares line iterative fitting detected slope to carry out estimation, first coefficient is dropped to two dimension from three-dimensional, again two-dimentional coefficient is carried out to windowing least-squares line iterative fitting, the method trueness error can drop to 10-4 magnitude.

Although the preferred embodiment that has been shown specifically and has described in conjunction with figure discloses the present invention, but those skilled in the art are to be understood that, the estimation of motion vectors of the multidimensional vector matrix discrete cosine transform domain proposing for the invention described above can also be made various improvement on the basis that does not depart from content of the present invention.Therefore, protection scope of the present invention should be determined by the content of described claims.

One of ordinary skill in the art will appreciate that all or part of flow process realizing in above-described embodiment method, to come the hardware that instruction is relevant to complete by computer program, described program can be stored in a computer read/write memory medium, this program, when carrying out, can comprise the as above flow process of embodiment.Wherein said storage medium can be CD, read-only store-memory body or random store-memory body etc.

Claims (5)

1. a transform domain overall situation high accuracy motion vectors method of estimation, is characterized in that: at least comprise following step:

The first step: video selecting step

Choose yuv format video, Y is luminance component, and U, V are chromatic components, and human visual system is more responsive to brightness ratio colour, and therefore the main Y luminance component that extracts is as experimental data;

Second step: multidimensional piecemeal reconstitution steps

This step is mainly that source video data is carried out to piecemeal restructuring, is divided into several video submatrixs, and a minute block size generally has 64*64,32*32,16*16 etc.;

The 3rd step: matrixing step

Use multidimensional vector matrix discrete cosine transform matrix, three-dimensional submatrix is carried out to the discrete cosine transform of multidimensional vector matrix, calculate three-dimensional coefficient matrix, coefficient after conversion is also a three-dimensional matrice, the corresponding three-dimensional coordinate of coefficient of each in this three-dimensional matrice, the position that this coefficient just can be corresponding in three-dimensional system of coordinate represents with a dot;

The 4th step: coefficient dimensionality reduction step

In the three-dimensional coefficient matrix that previous step is calculated, be not that 0 coefficient shows with dot in three-dimensional system of coordinate, these dots mainly concentrate in a folding plane, two edges of folding plane are on three-dimensional coordinate left and right side, and these points concentrate on a folding straightway, extract this two sides of containing motion vector information, just can the processing of three-dimensional coefficient be dropped on two dimensional surface and be processed;

The 5th step: folding point extraction step

First, data point in extract in previous step two faces is carried out to longitudinal axis maximum and minimum value intercepting, then, calculate these points of intercepting to the distance of initial point, according to tell several folding points apart from size, finally calculate respectively near the abscissa average of point each folding point, using this abscissa as folding point;

The 6th step: window determining step

The folding point calculating according to previous step, further determines the scope of window by the abscissa of folding point;

The 7th step: windowing data screening step

If motion vector is integer pixel, each section of straight line is all complete, if but motion vector is non-integer pixel, and final stage is not complete, and for integer pixel, the window that adopts previous step to obtain intercepts out final stage data; For non-integer pixel, windowing intercepts segment data second from the bottom;

The 8th step: straight line iterative fitting step

The window obtaining by previous step carries out data point intercepting, more respectively remaining data is carried out to least squares line fitting, and the slope that calculates fitting a straight line is d ₁and d ₂;

The 9th step: obtain motion vector step

The straight slope d calculating according to previous step ₁and d ₂, therefore estimated motion vector is [± d ₁, ± d ₂].

2. a kind of transform domain overall situation high accuracy motion vectors method of estimation according to claim 1, is characterized in that, described second step multidimensional piecemeal restructuring mainly realizes according to the following steps:

(1) step, is defined as the origin of coordinates by the video Y component upper left corner, laterally scans M to the right ₁individual pixel, skips to the second row, laterally scans M to the right ₁individual pixel, repeats said process, until scanning M ₂oK, now obtain first M of Y the first frame ₁* M ₂piece;

(2) step, skips to Y the second frame, repeats first step process, obtains first M of Y the second frame ₁* M ₂piece, skips to Y the 3rd frame, the 4th frame successively until N ₁frame, obtains first M ₁* M ₂* N ₁video submatrix;

(3) step, skips back to Y the first frame, moves to right successively to second, realizes the piecemeal sampling of view picture Y frame, obtains all M of Y frame for the 3rd until last piece repeats first two steps ₁* M ₂* N ₁3 D video submatrix, M in the present invention ₁and M ₂equate.

3. a kind of transform domain overall situation high accuracy motion vectors method of estimation according to claim 1, it is characterized in that: the 3rd described step multidimensional vector matrix discrete cosine transform, be that the 3 D video submatrix that restructuring obtains to multidimensional piecemeal carries out the discrete cosine transform of multidimensional vector matrix, the formula of conversion is:

$B_{\mathrm{IJ}}=C_{\mathrm{II}}{A}_{\mathrm{IJ}}{C}_{\mathrm{JJ}}^T$

For image sequence after translation, vector I is two-dimensional vector, and J is one dimension scalar, and matrix A is 3-D view matrix, and B is the three-dimensional coefficient matrix having converted, and C is the nuclear matrix of multidimensional vector matrix discrete cosine transform, and T is multidimensional vector matrix transpose.

4. a kind of transform domain overall situation high accuracy motion vectors method of estimation according to claim 1, is characterized in that: described the 8th step straight line iterative fitting step, mainly realizes by the following method:

(1) step, obtains data acquisition least squares line fitting with windowing intercepting, simulates straight line;

(2) step, calculates each point to the distance of this straight line, and obtains its average, according to this average, sets suitable step-length;

(3) step, setting threshold is that average deducts step-length k doubly, and k is since 0, and k of every iteration adds 1;

(4) step, by the threshold value of setting, reject the point that distance is greater than threshold value, with remaining point, again adopt least squares line fitting method, simulate straight line, the slope that the slope of this straight line and (1) step is simulated to straight line contrasts, if difference is less than set-point, stop iteration, derive the slope that (4) step is tried to achieve, if difference is very large, turn back to (3) step, down carry out successively.

5. a kind of transform domain overall situation high accuracy motion vectors method of estimation according to claim 4, is characterized in that: described least squares line fitting principle and formula are:

With discrete distribution and the coordinate points that totally presents straight path, obtain straight line parameter, the band ginseng equation of establishing straight line is:

y＝ax+b

Utilize the coordinate (x of certain Interval Discrete point on straight path _i, y _i) (i=1 ..., n), definition error term

principle of least square method requires Q to reach minimum value, and the condition of Q minimum is $\frac{\∂Q}{\∂}=\frac{\∂Q}{\∂b}=0,$ Can obtain:

$\left[\begin{array}{cc}\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{x}_i^2& -\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{x}_i\\ -\underset{i=1}{\overset{n}{\mathrm{\Σ}}}& n\end{array}\right]=\left[\begin{array}{c}a\\ b\end{array}\right]\left[\begin{array}{c}-\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{x}_i{y}_i\\ \underset{i=1}{\overset{n}{\mathrm{\Σ}}}{y}_i\end{array}\right]$

Matrix operation or direct solving equations method be can use, parameter a and b obtained:

$a=\frac{n\underset{i=1}{\overset{n}{\mathrm{\Σ}}}\left(x_i{y}_i\right)-\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{x}_i\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{y}_i}{n\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{x}_1^2-{\left(\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{x}_i\right)}^2}$

$b=\frac{\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{x}_i^2\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{y}_i-\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{x}_i\underset{i=1}{\overset{n}{\mathrm{\Σ}}}\left(x_i{y}_i\right)}{n\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{x}_i^2-{\left(\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{x}_i\right)}^2}$

Images