KR19990041461A - Sub-pixel-based motion estimation method and apparatus therefor - Google Patents

Abstract

The present invention discloses a sub-pixel-based motion estimation method and an apparatus therefor for a moving picture encoding apparatus for encoding a moving picture. A first estimator for searching a block of a predetermined size of a previous or subsequent reference frame for each block of a current frame divided into blocks of a predetermined size on a pixel-by-pixel basis, And a second estimator for estimating an optimal subpixel motion vector using a motion vector of the refinement unit and an inter pixel error value generated upon motion vector estimation of the refinement unit, Maintain good performance without doing so dramatically reduces computation time, resulting in reduced computation time and hardware complexity.

Description

Sub-pixel-based motion estimation method and apparatus therefor

The present invention relates to a field of image coding, and more particularly, to a simple sub-pixel-based motion estimation method that does not require interpolation between pixels and an apparatus therefor.

Digital image data used in a personal computer, a game machine, a terrestrial digital broadcast receiver, a digital satellite broadcast receiver, and a cable television (CATV) supporting a video conference, a high-definition television, a VOD receiver, a Moving Picture Experts Group Since the amount of data is greatly increased in the process of digitizing analog signals and the characteristics of the video itself, it is compressed by an efficient compression method rather than being used as it is.

Compression of digital image data largely uses three methods. A method of reducing temporal redundancy, a method of reducing spatial redundancy, and a method of reducing by using the statistical characteristics of generated codes are mainly used. A typical method for reducing temporal redundancy is motion estimation and compensation.

The motion estimation and compensation method for finding the most similar portion from the previous or subsequent reference screen for a certain portion of the current screen and transmitting only the difference components of the two portions is as follows. As the motion vector is searched as precisely as possible, Can be reduced more effectively, but it takes a considerable amount of time and computation to find the most similar part of the screen before or after. Therefore, efforts to reduce the motion estimation time, which takes the longest time in coding the moving picture, are continuing.

On the other hand, the motion estimation method is largely a pixel-by-pixel basis estimation method and a block-by-block basis estimation method.

Among the pixel-based estimation methods, a two-dimensional logarithmic search and a three-step search method have been proposed as a method of finding a motion vector in an integer pixel unit, but a local minimum value, MPEG-2 and the like generally use the full search method because of the possibility of falling into a sub-pixel and the necessity of finding a motion vector more finely in a sub-pixel unit.

Next, a sub-pixel-by-pixel motion estimation method for more precisely estimating between a refinement center and a refinement point by interpolation so as to correspond to a possibility that a motion vector is not a pixel unit is adopted and used in MPEG-2 and the like at present, The increase in the amount of computation and the encoding time due to the increase in the number of frames has a lot of room for improvement.

In the case of the half-pixel-based motion estimation used in MPEG as an example of the sub-pixel-based motion estimation method, 16 In order to estimate and compensate for the motion of a luminance component in units of macroblocks having a pixel size of 16 pixels, an optimal motion-vector unit is searched by a global search method in an adjacent macroblock. The motion vector is estimated again in the half pixel unit near the motion vector of the found video unit. The eight half pixels which are candidates of the optimal motion vector of the half pixel unit to be found at this time are all interpolated (rounding average) .

Therefore, 16 In a 16-pixel-size macroblock, 8 16 16 half-pixels have to be calculated by interpolation, so that there is a problem that much more computation time and calculation time are required than motion estimation in the cleansing area.

It is an object of the present invention to provide a method of estimating a motion vector of a sub-pixel in which a calculation time and a calculation amount are reduced.

It is another object of the present invention to provide an apparatus for estimating a sub-pixel-based motion vector that reduces hardware complexity.

According to another aspect of the present invention, there is provided a motion estimation method for a moving picture encoding method, the method comprising the steps of: estimating, for each block of a current frame divided into blocks of a predetermined size, Estimating a motion vector in a cleansing area unit in which a pixel difference between two blocks is minimized by searching a block of a predetermined size on a pixel-by-pixel basis, and calculating an error value between pixels generated in a motion- And estimating the optimal sub-pixel-based motion vector using the motion vector.

According to another aspect of the present invention, there is provided a motion estimator for a moving picture encoding apparatus for encoding a moving picture, the apparatus comprising: A first estimator for searching a block of a predetermined size of pixels in a pixel-by-pixel basis and estimating a motion vector in a cleansing-area unit in which a pixel difference between the two blocks is minimized, and a first estimator for estimating a motion vector in a cleansing- And a second estimator for estimating an optimal sub-pixel-based motion vector using the motion vector.

1 is a block diagram of a moving picture encoding apparatus to which the present invention is applied.

2 is a schematic block diagram of a motion estimator according to the present invention.

FIG. 3 is a relative coordinate of a refinement center and a half-pixel for facilitating understanding of the present invention.

4 is a diagram showing an error value generated in the motion estimation method according to an embodiment of the present invention in a plan view.

5 is a diagram showing an error value generated in a motion estimation method according to another embodiment of the present invention in a plan view.

FIG. 6 is a diagram illustrating weights for a pixel of a rectifying unit and neighboring upper, lower, left, and right pixels for explaining a motion estimation method according to another embodiment of the present invention.

7 is a diagram showing an error value generated in a motion estimation method according to another embodiment of the present invention in a plan view.

8A to 8D are diagrams showing motion compensation error curves estimated by the motion estimation method applied in FIG.

9A to 9C are views for explaining the number of interpolation times required for the motion estimation method according to another embodiment of the present invention.

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS Hereinafter, preferred embodiments of a motion estimation method and a motion estimation method therefor according to the present invention will be described with reference to the accompanying drawings.

1, which is a block diagram of a moving picture encoding apparatus to which the present invention is applied, the moving picture encoding apparatus performs intra-frame compression and inter-frame compression, and compresses I, .

In general, the I frame pixel data is output to the discrete cosine transform (DCT) calculator 104 as it is without being affected by the subtractor 102. The DCT operator 104 multiplies the DCT by 8 The quantizer 106 quantizes the DCT data. The variable length encoder 108 performs run-length encoding of the quantized data to perform statistical encoding. The variable length encoder 108 performs statistical encoding on the statistically encoded image data, the motion estimator 122 and the additional information such as the macroblock type. The buffer 110 temporarily stores the bit rate of the multiplexed image data.

The compressed I frame output from the quantizer 106 is inversely quantized by the inverse quantizer 112 and the inverse DCT operator 114 performs inverse DCT on the inversely quantized I frame to output the decompressed I frame to the adder 106. [ Lt; / RTI > The decompressed I frame is transmitted to the frame memory 118, which functions as a buffer, without being affected by the adder 114, and is then stored in the frame memory 150 to predict and compress the P and B frames. Predictive coding of P and B frames is also similar, and P frame compression is described. The image frames stored in the frame memory 118 are supplied to the motion estimator 120. [

The motion estimator 120 calculates a motion vector using the current input image data and the image data of the previous or subsequent reference frame stored in the frame memory 118 and outputs the motion vector to the variable length encoder 108 and the motion compensator 122 do.

The motion compensator 122 reads a block corresponding to the motion vector estimated by the motion estimator 120 from the frame memory 118 and supplies the block to the subtractor 102. The subtractor 102 subtracts the predicted block from the frame memory 118 through the motion compensator 122 from the block corresponding to the current frame to be decompressed, in pixel-by-pixel units. The difference or residue obtained by the subtraction of the subtractor 102 is applied to the DCT performer 104.

On the other hand, the compressed P frame is decoded by the inverse quantizer 112 and the inverse DCT calculator 114, and the decoded data is applied to the first input of the adder 116. At the same time, each block of the reference image frame stored in the frame memory 118 is accessed to predict the current frame, and the accessed block is applied to the second input of the adder 116 via the motion compensator 122. The adder 116 adds the encoded residual or difference and the data output from the motion compensator 122 to reconstruct the actual image. The recovered P frame from the adder 116 is then stored in the frame memory 118 to predict and encode / decode the P and B frames.

FIG. 2 is a schematic block diagram of a motion estimator according to the present invention, and the motion estimator 200 can be applied to FIG. 2, the motion estimator 200 includes a first estimator 210 for estimating a motion vector MV on a per-picture-unit basis, and a motion vector MV ' And a second estimator 220 for estimating the second estimator 220.

The first estimator 210 divides the current frame to be compressed into macroblocks and searches for the divided current frame in the search area of the previous or subsequent reference frame on a pixel-by-pixel basis. Estimates and outputs a motion vector (MV) in a cleansing area by finding a position at which an error is minimized. The second estimator 220 estimates the optimal sub-pixel-based motion vector MV 'based on the motion vector of each cleansing area by the sub-pixel-based motion estimation method to be described below.

Hereinafter, a sub-pixel-based motion estimation method proposed by the present invention will be described.

3, Marked locations are the locations of the purification plant, The locations shown are the locations of the half pixels. The dual (0, 0) coordinates indicate where the motion vector of the cleansing unit is already pointed by the global search method. Therefore, the motion vector in the half-pixel unit may be the (0.0) coordinate itself or one of the coordinates of the eight half-pixels around the (0, 0) coordinate.

(I) a generalized inverse matrix solution,

3, A plane composed of three pixels can be modeled by a quadratic equation as shown in Equation (1), and is composed of nine unknown coefficients.

The equation shown in Equation 1 becomes a quadratic equation of the point when x or y is fixed. Each time the fixed x or y changes, the slope of the graph will be a quadratic curve that changes. Since the sum of the error difference at (0, 0), which is the motion vector of the refinement unit already found in the global search process, and the sum of the error differences at the surrounding eight purification stations are 9, 9 coefficients can be obtained. (1, 1) to the lower right square (coordinates (1, 1)) in the upper left corner,

(1, -1), (-1, -1), (-1, -1), and the coordinates , N = 2, ..., 9 in order from 0, 0, 0, 0, 1, 1, 1, When a motion vector is obtained, the sum can be calculated using the difference between two images as an absolute value, or the sum of squares of differences between two images can be calculated. In the present invention, both are generally expressed as a sum of error differences.

In order to solve Equation (1), Equation (2) The inverse of the matrix is to be found. Since this matrix is non-singular, the inverse matrix is obtained as in Equation (3).

Since every motion block in a macroblock is always set to (0,0) and the surrounding is set to its relative value (-1, 0, or 1) as in FIG. 3, the inverse matrix is not necessarily obtained every time, A fixed value such as 3 can be used. The increase of the calculation amount due to the inverse matrix of the matrix 9 is not necessary to be considered at all. Further, in Equation 3, The 9 matrix is a sparse matrix and can be advantageous in terms of computational complexity.

The nine coefficients c ₁ < RTI ID = 0.0 _> c ₉ to complete Equation 1 and substitute the respective coordinates of the half pixels of FIG. 3 to estimate the sum of the error differences in the coordinates. As shown in Equation (4), at this time, 9 9 matrix is a fixed and sparse matrix and has some degree of regularity.

=

(Ii) least square solution

(I) method uses the sum of the error differences of nine coordinates to solve the equation with nine unknowns. As a simpler method, we use x ² y ² , x ² y, xy ² in The equation (5) is given.

When the sum of the three-dimensional error differences obtained by the above Equation 1 is represented as a contour line as shown in FIG. 4, it can be represented as an ellipse obliquely inclined in the diagonal direction with respect to the x and y axes, The sum of the three-dimensional error differences calculated by Equation (5) is expressed by only the ellipses that are distorted in the horizontal direction in the x-axis or the y-axis as shown in Fig. 5, can do.

Since Equation 5 has six unknowns, the number of required coordinates is six, but it is not known which one of the nine given coordinates should be selected to produce a more accurate result. Therefore, as can be seen from Equation (6), when the number of unknowns is 6 and the given coordinate value is 9, the least squares method shown in Equation (7) is used in the present invention. The result of the pseudo inverse matrix at that time is given by Equation (8).

Here, f is the sum of the error differences of each coordinate, and A is the sum of error differences of 6 9 matrix, c is the unknown coefficient c ₁ c ₆ .

With the coefficients thus obtained, the equation (5) is completed in the same manner as in the method (i), and the coordinates of the half pixels are substituted as in the equation (9) to estimate the sum of error differences in the coordinates, Pixel as a sub-pixel-based motion vector.

=

(Iii) weighted least squares method

In order to obtain the solution of Equation (5), a coefficient is obtained by weighting coordinates closer to (0, 0) rather than just using the nine coordinates in Fig. This method is added to the step (ii) of multiplying Equation (5) of the method by a weighting matrix (W).

As shown in FIG. 6, since the error is smaller as the distance from the coordinates of the cleansing site that has already been found by the global search method is smaller, the influence by the closer coordinates is increased to improve the accuracy. In Fig. 6, the part (2) is the coordinate with which the weight is given. That is, a weight of s (side) is applied to (0, -1), (-1,0), (1,0), (0,1) The weight matrix W is given by Equation 10. < EMI ID = 10.0 > The values of c 'and s are both smaller than 1, and the value of c' is smaller than the value of s. Here, the weighted least squares solution is expressed by Equation (11).

(Iv) a simplified fast algorithm,

This method uses Equation (12) by eliminating the xy term more simply than (ii) method.

The sum of the error differences obtained by the equation (12) becomes a shape of a circle contour line as shown in Fig. In addition, this means that there are four points (0, -1), (-1,0), (1,0), and (0,0) , 1). Therefore, a square matrix is obtained. Assuming that x and y are separable, the motion compensation error E (x) can be given by Equation (13).

here, = 2, a > 0, b <1 and c 0. b is an accurate motion vector, and x is one of the vectors around b, it can be represented as shown in FIG.

In FIG. 8A, P ₀ is a square mean of the motion compensation error of the refinement motion vector (m ₀ , n ₀ ) represented by Equation (14) below, and Y (m, n) Y '(m, n) are the luminance values of the current frame and the previous frame, respectively. Only the horizontal direction (i.e., the x-axis direction) of the motion vector will be described as an example, and the vertical direction (i.e., the y-axis direction) can be obtained in the same way because x and y are separable.

8, when the square mean of the motion compensation error when shifting by one pixel to the left and the right from the motion vector in the unit of the cleansing center is defined as P ₁ and P _-1 , P ₁ and P _-1 can be calculated by the above calculation method have.

On the other hand, P ₀ , P ₁ , and P _-1 , which are motion compensation prediction errors, can be used as follows.

Since (P ₁ -P ₀ ) / (P _-1 -P ₀ ) in Equation (15) is determined by the value of the correct motion vector b, it can be written as a function g (b) for the variable b. In this case, (P ₁ -P ₀₎ and (P _-1 -P ₀₎ x-axis component of the optimal motion vector in half pixel units becomes searching, by the relative dimensions of (MV _x ') yi ₀ -0.5 m, m ₀ , and m ₀ +0.5.

That is, (P ₁ -P ₀₎ is as shown in full (b) of Figure 8 is greater than (P _-1 -P _0), and the MV _x 'is a m-0.5, opposed to (P _-1 - P ₀ ) is sufficiently larger than (P ₁ -P ₀ ), MV _x 'becomes m + 0.5 as shown in FIG. 8 (d). Except for the above two cases, MV _x 'becomes m ₀ as shown in FIG. 8 (c). This can be expressed by Equation (16).

Here, MV _x 'is a new motion vector in the horizontal direction (x-axis direction), and a new motion vector in the vertical direction (y-axis direction) can be obtained in the same way using two motion compensation errors in another vertical direction . Similarly, subpixels of other precision, such as quarter-pixel units, can be easily derived.

In this simplified high-speed algorithm, the error value for the difference between pixels is generally modeled as shown in Equation (13), and instead of calculating the error value at the half pixel position directly using the error value, Value of the pixel value is used to estimate the error value at the half-pixel position.

= 2, Equation (16) can be simply expressed as Equation (17).

Also, = 1, Equation (16) is simply expressed by Equation (18).

The multiplication by 2 in Equation (18) can be easily implemented by a 1-bit shift operation, and the multiplication by Equation (17) to (3) is also performed by 1-bit multiplication by a 1-bit shift operation The multiplication by this method is not performed once.

(V) modified weighted least squares method

(0, -1), (- 1,0), (0,0), and (0,0), which were used in the method ), (1,0), (0,1) are weighted and least squares solution is used. The process of obtaining is the same as (iii).

(Vi) partial interpolation solution

In order to increase the accuracy after finding the half-pixel first by the method of (iv) having no complexity due to the multiplication, the half-pixel nearest to the nearest half-pixel and the half-pixel found by the method (iv) The sum of error differences is again obtained by interpolation, and the values are compared.

Therefore, there are three cases as shown in Figs. 9 (a) to 9 (c). In FIG. 9, the sum of the error differences at the position (0, 0) of the refinement site is used when it is already found in finding the motion in the refinement unit. Is the half-pixel obtained by the method (iv) Is a half pixel for which the sum of error differences must be found again by interpolation.

9 (a) and 9 (b), the interpolation required by the proposed method is three, while the interpolation required by the proposed method is three, and the interpolation necessary for full- (C), the computation amount is less than half that of global motion estimation using interpolation.

In other words, if the half-pixel found is the left or right side of the motion vector of the cleansing unit, there are three half-pixels to be interpolated, and if the half-pixel is the diagonal direction of the motion- If the half-pixel is the motion vector of the refinement unit, or if it is itself, there are four half-pixels to be interpolated.

Implementations of the invention may be implemented in software, depending on the application, or in hardware for high-speed real-time processing.

The present invention is a sub-pixel-based motion estimation method which has good performance without interpolation between pixels, and has the effect of drastically reducing the amount of computation, thereby reducing computation time and hardware complexity.

Claims (22)

A motion estimation method for moving picture encoding, comprising: (a) For each block of a current frame divided into blocks of a predetermined size, a block of a predetermined size of a previous frame or a subsequent reference frame is searched for every pixel to estimate a motion vector in a cleansing area where the pixel difference between the two blocks is the smallest step; And (b) estimating a motion vector of an optimal sub-pixel unit using the motion vector of the refinement unit and the inter-pixel error value generated when the motion vector estimating unit of the refinement unit is used. . The method of claim 1, wherein the step (b) (b1) calculating coefficients for the following equation representing a plane composed of the motion vector of the cleansing area unit and the surrounding eight pixels thereof using pixel error values for the pixels; --(One) (b2) substituting the calculated coefficients and the respective coordinates of the sub-pixels around the motion vector of the refinement unit into the equation (1) to estimate the sum of error differences in each coordinate of the sub-pixels; And (b3) determining a position having a smallest sum of error differences in each coordinate as an optimal sub-pixel unit motion vector. The method of claim 1, wherein the step (b) (b11) selecting a first predetermined number of purification sites using the least squares method among the motion vectors of the purification unit unit and the surrounding purification units; (b12) calculating the first predetermined number of coefficients for the following equation using the pixel error values for the first predetermined number of purification points; ---(2) (b13) substituting the first predetermined number of coefficients and the respective coordinates of the sub-pixels around the motion vector in the reflex center into the equation (2) to estimate the sum of error differences in each coordinate of the sub-pixels ; And (b14) determining a position having a smallest sum of error differences in each coordinate as an optimal sub-pixel unit motion vector. 4. The method of claim 3, wherein in step (b11), a first predetermined number of purifiers is selected using a least squares method, by weighting the purifiers of the purifier-unit motion vectors and some of the surrounding purifiers. / RTI &gt; 5. The method of claim 4, wherein a weight of a first predetermined value is applied to the coordinates of the motion vector of the refinement unit, and a weight of a second predetermined value is applied to the coordinates of the upper, lower, left, Wherein the first weight is smaller than the second weight, and the first and second weights are smaller than " 1 ". The method of claim 1, wherein the step (b) (b21) calculating motion compensation error values at neighboring pixel positions in the horizontal and vertical directions from the motion vector of each refinement unit; And (b22) comparing the calculated motion compensation error values based on the motion compensation error value of the motion-picture-unit motion vector unit to estimate an optimal sub-pixel motion vector. 7. The method of claim 6, wherein, in the step (b22) If the motion compensation error value between the block of the reference frame (the previous block) and the block of the current frame (the current block) is larger than the motion compensation error value between the block of the previous reference frame and the current block, The motion compensation error value between the previous block and the current block is sufficiently larger than the motion compensation error value between the current block and the current block. If the motion compensation error value between the previous block and the current block is sufficiently larger than the motion compensation error value between the current block and the current block, And an upper half pixel is estimated as an x component and a y component of an optimal subpixel motion vector, and otherwise, the refinement portion is estimated as a subpixel motion vector. 7. The method of claim 6, wherein step (b) (b31) Weights are applied to the coordinate values of the motion vectors of the refinement unit and the surrounding pixels of the motion vectors of the refinement unit, and the second predetermined number of refinement points is calculated using the least squares method Selecting; (b32) calculating the second predetermined number of coefficients for the following equation using the inter pixel error value for the second predetermined number of pixels; --- (3) (b33) substituting the second predetermined number of coefficients and the respective coordinates of the sub-pixels around the motion vector of the refinement unit into the equation (3) to estimate a sum of error differences in each coordinate of the sub-pixel; And (b34) determining a position having a smallest sum of error differences in each coordinate as an optimal sub-pixel unit motion vector. The method of claim 8, wherein a weight of a first predetermined value is applied to the coordinates of the motion vector of the refinement unit, and a weight of a second predetermined value is applied to the coordinates of the upper, lower, left, Wherein the first weight is smaller than the second weight, and the first and second weights are smaller than " 1 ". 9. The method of claim 8, wherein step (b) (b35) obtaining a sum of error differences by interpolating subpixels nearest to the nearest distance from the searched subpixel in step (b34); And (b36) comparing the sum of the error differences obtained in the step (b35), and determining a position having the smallest value as a motion vector of an optimal sub-pixel unit. 11. The method of claim 10, wherein, in step (b35), when the subpixel found in step (b34) is located in the left, right, or diagonal direction of the motion vector of the refinement unit, the number of subpixels to be interpolated is three, Wherein the number of sub-pixels to be interpolated is four when the searched sub-pixel is up, down, or itself of a motion vector of a refinement unit. 1. A motion estimator for a moving picture encoder for encoding a moving picture, the motion estimator comprising: A first estimator for estimating a motion vector in a cleansing area in which a pixel difference between two blocks is minimized by searching for each block of a current frame divided into blocks of a predetermined size, ; And And a second estimator for estimating an optimal sub-pixel-by-pixel motion vector using the motion vector of each of the refinement units and the inter-pixel error value generated when estimating the motion vector of each refinement unit. 13. The apparatus of claim 12, wherein the second estimator comprises: A first calculator for calculating coefficients for the following equation expressing a plane composed of the motion vector of the refinement unit and eight surrounding pixels thereof using the inter pixel error values in the pixels; And A position having the smallest value of the sum of error differences at each coordinate of the subpixels is calculated by using the calculated coefficients and the coordinates of the subpixel, which is a motion vector of the refinement unit, and surrounding subpixels, And a determiner for determining the motion vector as a motion vector. 13. The apparatus of claim 12, wherein the second estimator comprises: A selector for selecting a first predetermined number of cleansing sites by using a least squares method among the motion vectors of the cleansing site unit and surrounding cleansing sites; A calculator for calculating the first predetermined number of coefficients for the following equation using the inter pixel error value for the first predetermined number of pixels; And Using the first predetermined number of coefficients and the coordinates of the subpixel, which is the motion vector of the refinement unit, and the surrounding subpixels, a position having the smallest value of the error difference in each coordinate of the subpixels is referred to as an optimal And determining a motion vector of the subpixel unit as a motion vector. 15. The motion estimator of claim 14, wherein the selector selects a first predetermined number of refinement points using a least squares method by weighting the refinement units of the refinement-area-based motion vector and a part of the refiners of the surrounding refinement units. . The method of claim 15, wherein a weight of a first predetermined value is applied to the coordinates of the motion vector of the refinement unit, and a weight of a second predetermined value is applied to the coordinates of the upper, Wherein the first weight is smaller than the second weight and the first and second weights are smaller than " 1 ". 13. The apparatus of claim 12, wherein the second estimator comprises: A calculator for calculating motion compensation error values at neighboring pixel positions in the horizontal and vertical directions from the motion vector of each refinement unit; And And a determiner for determining an optimal sub-pixel-based motion vector by comparing the calculated motion compensation error values on the basis of the motion compensation error value of the motion-vector-based motion vector. 18. The apparatus of claim 17, If the motion compensation error value between the block of the reference frame (the previous block) and the block of the current frame (the current block) is larger than the motion compensation error value between the block of the previous reference frame and the current block, The motion compensation error value between the previous block and the current block is sufficiently larger than the motion compensation error value between the current block and the subsequent block. If the motion compensation error value between the previous block and the current block is sufficiently larger than the motion compensation error value between the current block and the current block, And an upper half pixel is determined as an x component and a y component of an optimal sub-pixel motion vector, and otherwise the motion vector is determined as a sub-pixel motion vector. 18. The apparatus of claim 17, wherein the second estimator comprises: A selector for weighting the coordinate values of the motion vectors of the refinement unit and the refinement centers of the surrounding refinement sites and selecting the second predetermined number of refinement sites using the least squares method; A calculator to calculate the second predetermined number of coefficients for the following equation using the inter pixel error value in the second predetermined number of pixels; And Using the second predetermined number of coefficients and the coordinates of the subpixel, which is a motion vector of the refinement unit, and the surrounding subpixels, a position having the smallest value among the sum of error differences in each coordinate of the subpixel is referred to as an optimal And determining a motion vector of the subpixel unit as a motion vector. The method of claim 19, wherein a weight of a first predetermined value is applied to the coordinates of the motion vector of the refinement unit, and a weight of a second predetermined value is applied to the coordinates of the upper, Wherein the first weight is smaller than the second weight and the first and second weights are smaller than " 1 ". 20. The apparatus of claim 19, wherein the second estimator comprises: An interpolator for obtaining a sum of error differences by interpolating subpixels nearest to the closest distance from the subpixel found in the determiner; And Further comprising a comparator for comparing the sums of error differences obtained in the interpolator and determining a position having the smallest value as an optimal sub-pixel unit motion vector. The apparatus of claim 21, wherein the interpolator has three interpolators to interpolate when the subpixel searched in the determiner is located in the left, right, or diagonal direction of the motion vector of the refinement unit, Wherein the number of macroblocks to be interpolated when the motion vector is up, down, or itself is four.