JP2980810B2 - Motion vector search method and apparatus - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method and an apparatus for searching for a motion vector in moving picture coding. More specifically, an attempt is made to provide a novel method and apparatus for expressing a motion vector field with a smaller number of representative motion vectors by using a quad-tree, and reducing the information amount of the motion vector field. Is what you do.

[0002]

2. Description of the Related Art In moving picture coding, a method using a combination of inter-frame prediction by estimating motion information and coding of prediction error is often used. For encoding at a lower rate (less image data), it is important how motion information is estimated and used. In the motion vector estimation based on the block matching method, which has been frequently used in moving image coding, there is a problem that when there is a block having a different motion, the estimation accuracy is significantly deteriorated. There are several documents showing a method for solving such a problem.

Reference 1. JLBarron, DJFleet, SSBea
uchemin, TABurkitt: “Performance of Optical Flow
Techniques ", Proceeding CVPR June, 1992, pp. 236-
242

Reference 2. Ryoma Ohtsuna, Maki Sato, Aizawa
Kiyoharu, Hatori Mitsutoshi: "Basic Study of Motion Estimation Based on One Pixel Matching", Television Techniques, ICS94-38 (1994)

[0005] Reference 1 estimates a denser vector field for higher-order processing in the field of computer vision and the like. If a dense motion vector field can be used, the estimation accuracy will not be significantly degraded even in a boundary region between different motions.

[0006] In Reference 2, when a dense vector field is generated, a method of using error information by matching on a pixel-by-pixel basis is used. According to this method, an error plane provided for each pixel can be effectively used for calculation.

The contents of Reference 2 will be described below.

[0008] Motion estimation based on one-pixel matching

In general, when motion estimation is performed by the block matching method, matching is performed around the pixel x.
A window W _m and a search range W _s are set, and a motion vector Δx that minimizes the error function e _B (Δx) is obtained as shown in Expression (1). e _B (Δx) = Σ | I _{k + 1} (x + Δx) −I _k (x) | (1) where Σ is the sum of all the pixels x included in the matching window W _m . I _{k + 1} (x + Δx)
Denotes a reference image, and I _k (x) denotes an original image.

[0010] If the number of pixels included in the matching window W _m was reduced to one pixel, that is, when the matching is one pixel rather than block, Equation (1) becomes Equation (2). e _x (Δx) = | I _{k + 1} (x + Δx) −I _k (x) | (2) This is an error function of one pixel matching at the pixel x _ex
(Δx).

FIG. 7 shows an error function e _{x for} one-pixel matching.
In (Δx), for many pixels x included in a region Rx ₀ (solid black in the figure) including the pixel x ₀ , the operator F
Is applied to obtain the motion vector Δx ₀ at the pixel x ₀ , and examples of the assumed motion model include a parallel movement model and an affine model.
Here, in the basic framework of one-pixel matching, first, 1
Matching is performed at the pixels, and some operator F acts on the result to obtain a motion. At this time, an error function e _x (Δx) is calculated in advance for each pixel, and its value (distortion plane in a pixel unit: error plane) is held, so that the degree of freedom of processing increases, and the nonlinear function as the operator F is increased. And the like, more flexible motion estimation becomes possible.

Specifically, the motion vector Δx of the pixel x ₀
When seeking _0, the basic framework of a motion estimation based on 1 pixel matching, 1) First, with respect to some threshold T, motion e _x0 ([Delta] x ₀₎ becomes ≦ T [Delta] x ₀ selected as candidate vectors. This set of motion vectors, V _x0 = | represented by _{{Δx 0 e x0 (Δx 0} ) ≦ T}. For a region R _x0 including 2) x _0, under the assumption of some kind of motion model, e _x ([Delta] x) (where the operator x is an evaluation function in which) is included in the region R _x0 F Apply [•]. _{3) Δx 0 = arg min F} [e x (Δx 0), here, x is included in the region R _x0] obtaining a motion vector [Delta] x ₀ contained in the vector V _x0 where the.

In this framework, an operator F, an area R
_Since there is a degree of freedom in specifying _x0 , it is possible to perform motion estimation suitable for the properties of each region of the image.

As described above, various methods are conceivable for deriving the vector field by one-pixel matching, depending on the selection of the operator F and the control of the region R _x0 , but in order to finally extract the representative motion vector, It is desirable that the region is appropriately divided. Here, motion and region estimation are performed simultaneously by a method similar to distance conversion and clustering.

This method includes the steps of creating a binary label image labeled "1" or "0" and repeating distance conversion of the image. First, 1 pixel matching error function e _x a ([Delta] x) binarized by the threshold T _m, which represents in binary error function f ([Delta] x) of equation (3). When the _{e x (Δx) ≦ T m} , when _{f (Δx) = 1 e x} (Δx)> T m is, f (Δx) = 0 ( 3)

The distance conversion of the label image obtained by the binarization error function f (Δx) is performed on each motion vector Δx, and the distance conversion value d _x (Δx) is used as an index.

FIG. 8 shows distance conversion of label images classified according to each motion vector Δx. The black portions represent pixels x that could not be matched. The distance conversion value d from the point A is equivalent to the distance d from the pixel x to the first contact with a certain pixel 79 among several pixels that could not be matched when the value was increased, and the distance d from the pixel x. Only matched pixels exist within the range of the circle 71 of _x (Δx). Therefore, since it is considered that Δx in which the value of d _x (Δx) increases is likely to be a true motion vector, the distance x
Δx that gives the maximum value of d _x (Δx) is defined as a motion vector. That is, the motion vector [Delta] x to give _{d x max = maxd x (Δx} ) (4).

Specifically, the following procedure is taken in consideration of the influence of noise and the like. 1. Distance conversion is performed on the label image of each motion vector Δx. Distance The larger pixel than the maximum value d _{x max} is the value d ₀ of assigned as a motion vector Δx give d _{x max.} 2. For the pixel x for which the motion vector is determined, the binarization error function f (Δx) is obtained by calculating the motion vector Δx by the distance d _x
When giving the _max, when f (Δx) = 1 motion vector [Delta] x does not give the distance d _{x max,} and labeled as f (Δx) = 0 (5 ), and updates the label image. 3. The above is repeated while the distance d ₀ is gradually reduced.

The vector field obtained in this way is
The motion vector Δx becomes uniform for each region R _x0, which is suitable for interpolation approximation using the representative motion vector. The above is the outline disclosed in Reference 2.

A block matching method used for extracting a motion vector often includes a quad tree (Quad-Tre).
e) is used. If the prediction error of the block is larger than a certain threshold, this quad tree performs quadrant division, performs block matching of each of the four divided blocks, checks the prediction error, and if the prediction error is larger than a certain threshold, Is 4 more for blocks that exceed the threshold
The prediction error is examined by performing block matching by dividing. However, the threshold value depends on the image, and if the threshold value is small, there is a possibility that erroneous estimation of a motion may be caused due to an increase in additional information and an influence of noise due to a small block size if the threshold value is small. . If the threshold value is increased, it becomes difficult to extract a small area, and the reproducibility of the motion vector field and the prediction error deteriorate. For this reason, in the conventional quad tree division, an integration procedure of dividing an image into small sub-blocks in advance and integrating blocks having the same vector is adopted.

[0021]

The vector fields obtained by the methods disclosed in Documents 1 and 2 are dense on a pixel-by-pixel basis, and the accuracy in the boundary region of motion is ensured. However, when transmitting as inter-frame prediction information in ultra-low-rate coding, there is a problem that the amount of information is still large and needs to be solved. Also, when the block matching method is used, if the number of vectors is reduced by increasing the block size, additional information is unnecessary and the amount of information can be easily reduced. , The vector becomes uniform, the prediction error increases, and the reproducibility of the vector field deteriorates.

Therefore, when an image is divided into small sub-blocks in advance by the conventional quad tree division and an integration procedure of integrating sub-blocks having the same vector in the sub-block is adopted, this integration procedure is used. There remains an unsolved problem that the amount of processing increases accordingly.

[0023]

SUMMARY OF THE INVENTION The present invention has been made to solve such problems. That is, when a block is divided by a quad tree, first, an error function e _x (Δ
x) is calculated in advance, and by having a dense vector field on a pixel-by-pixel basis, a prediction error value calculated from the vector field is used as a threshold value for block division. For additional information accompanying the division, the division order is determined in advance. Thus, the vector field can be reproduced by transmitting only the hierarchical level of the division.

[0024]

[Action] Prior to division of the block, one pixel matching error function e _x a ([Delta] x) is calculated, because they give a dense vector field, the prediction error value calculated from the vector field, performs division It can be a threshold value for determining whether or not. For this reason, the threshold value changes adaptively according to the block to be divided, and optimal block division with the highest coding efficiency has become possible. Further, since the block division can be started from the entire image, the representative motion vector can be globally extracted, and the integration procedure which has been a problem in the past becomes unnecessary. The calculation of the threshold, in order to advance calculates an error function e _x 1 pixel ([Delta] x), there is no need for re-calculation for prediction error. By determining the quad tree generation procedure in this way, additional information for region description is required only at the hierarchical level, and the representative vector extraction order maintains positional continuity. Thus, the amount of information can be reduced by differential encoding using continuity of vectors.

[0025]

FIG. 1 shows a circuit configuration of an embodiment of the present invention. Where 11 is an error plane calculator, 12 is a memory,
13 is an original vector field generator, 14 is a threshold value generator, 15 is a division determiner, 16 is a block selector, 17 is a block matching unit, 18 is a quad tree generator, and 31 to 45 are This is a signal line for exchanging various information. When enumerating various kinds of information exchanged through the signal lines 31 to 44, the signal line 31 represents the original image I _k (x), the signal line 32 represents the reference image I _{k + 1} (x), and the signal line 33 represents 1 Error function e _x (Δx) of pixel matching
And the signal line 34 of 1 pixel matching read out from the memory 12 the error function e _x ([Delta] x), the signal line 35 is a vector field V (x), the average prediction error e _r of the signal line 36 is 1 pixel Threshold _Tq
Sum, the signal line 37 of 1 pixel matching read out from the memory 12 the error function e _x ([Delta] x), the signal lines 38 divide instructions, the signal line 39 is a block b _N, the _M, signal line 40 is a vector of V _bN, the _M, the average prediction error e _bN of the signal line 41 one pixel _unit, the _M, signal line 42 of 1 pixel matching read out from the memory 12 the error function e _x a ([Delta] x), the signal line 43 is blocked the position M (N), the signal line 44 representative vector components V _R, the signal lines 45 represent respectively the hierarchy level N.

FIG. 2 shows the principle of the operation of extracting a representative vector by the quad tree having the circuit configuration of FIG.
FIG. 8A shows an original vector field 80 including a circle 82 and a square 83.
84 is indicated by a broken line, and the divided block is represented by 85
Indicates an overlap portion between the divided block 84 and the square 83, and the arrow shown therein indicates the original vector field 8
It represents a part of a number of vectors existing in 0.

FIG. 2B shows the divided block 84 shown in FIG. 2A and an overlapped portion 85 indicated by diagonal lines, and the arrow shown therein indicates that the divided block 84 is subjected to block matching. Although the obtained vector is shown, it does not represent the vector of the overlap portion 85 in FIG. 9B, which indicates that the error is large at the hierarchical level of this division.

FIG. 9C shows a case where the divided block 84 shown in FIG. 9B is divided into four parts. In this four-part division, the vector of the upper right block including the overlap part 85 is still overlapped. Since the vector of the portion 85 is not accurately represented, the divided block 84
(C) indicates that the error is large at the hierarchical level of four divisions.

In FIG. 4D, the vector of the block including the overlapping portion 85 in FIG. 4C is not accurate, and the error (error function of one-pixel matching (signal line 3)
7) In the block including the overlapping portion 85,
Since the average error exceeds the threshold, the block is further divided into four, and the vector of each divided block is obtained. Here, it is shown for the first time that an accurate vector of the overlap portion 85 can be obtained.

FIG. 5E shows a final division result 90 by the vector field 80 shown in FIG. Here, division is performed until the vector in each block becomes substantially uniform, and a circle 92 and a square 93 corresponding to the circle 82 and the square 83 in FIG.

FIGS. 3 to 5 show the flow of operation of the circuit configuration shown in FIG. 1, and will be described with reference to FIG.

The original image I _k (x) is input from the signal line 31 and the reference image I _{k + 1} (x) is input from the signal line 32 to the error plane calculator 11. the error function e _x ([Delta] x) of the pixel _{matching, e x (Δx) = |} I k + 1 (x + Δx) -I k (x) | by seeking to store feed via a signal line 33 to the memory 12.

The original vector field generator 13 outputs an error function e _x (Δ) for one-pixel matching from the memory 12 through a signal line 34.
x) is read out, the vector field V (x) of the original image is obtained by one-pixel matching, and output to the signal line 35 (S1,
(Fig. 3).

The conditions of the components of each circuit are set. That is, in the quad tree generator 18, a block division procedure is defined.

FIG. 6 shows an example of a block dividing procedure by the quad tree. The hierarchical level of the quad tree is N, and the current block position in the generation order of the divided blocks is represented by M (N). N = 0 represents the entire undivided screen, and its block position is M (0) = 0. At the hierarchical level N = 1, the entire screen is divided into four parts, and M (1) = 0 at the upper left, M (1) = 1 at the upper right (not shown), and M (1) at the lower left as shown in FIG. = 2, lower right is M
(1) = 3 is set. Here, M (1) =
When the error of one divided block is equal to or larger than the threshold value, the hierarchical level N is set to N = 2, and the divided block is further divided into four, and M (2) =
0, M (2) = 1 at upper right, M (2) = 2 at lower left, and M (2) = 3 at lower right (S
2). The lowest hierarchical level is N = _NL . Further, the m-th representative vectors obtained as V _m, representing the representative vectors to obtain a V _m hierarchy N _Vm.

First, the hierarchical levels N = 0, M (N) = 0,
m = 0, the quad tree generator 18 places the current block position M (N) on the signal line 43 and the signal line 45
Is output to the layer N (N = 0) (S3). The block selector 16 that has received the instruction of the block position M (N) and the hierarchy N selects the block b from which the representative vector is to be extracted.
_{N and M} are determined and output to the signal line 39 (S4).

FIG. 5 shows the details of the procedure for determining the block b _{N, M.} When the flag fl (recursive processing with fl = 1) instructing the recursive processing of the block determination procedure is "1" at the block position M (N) = 3 (S21Y,
FIG. 5), the hierarchical level N is set to N = N−1 (S2
2) If N = 0 has not been reached (N in S23), the process returns to step S21, and if N = 0, it means that all divisions have been completed, and the process ends (S23).
Y).

If the block position M (N) is not 3 or the flag fl is not 1 (S21N), the flag fl is set to 0 (S25), and the block b to be processed is set.
_After setting _{N and M} , the block selector 16 outputs the block bN _{and M} to the signal line 39 and exits the subroutine for setting the block bN _{and M} (S26).

The signal line 39 by the block b _N to be _processed, the block matching unit 17 in line 42 that has received the instruction for _M reads the error function e _x from the memory 12 1 pixel matching ([Delta] x) block Perform matching, and output the obtained motion vector V _{bN, M} via the signal line 40 (S5). Further, the block matching unit 17
And the threshold generator 14, based on the vector V _{bN, M} and the vector V (x) (signal line 35) of the pixel x corresponding to the block b _{N, M} (signal line 39) of the original image I _k (x),
_Assuming that the total number of pixels included in the block b _{N, M} (signal line 39) is S (b _{N, M} ), the average prediction error e in pixel units
_bN, obtaining the _M and e _r from the following two equations. _{e bN, M = (1 /} S (b N, M)) Σe x (V bN, M) e r = (1 / S (b N, M)) Σe x (V (x)) where x block b _N, is in all respects the pixels included in _M (signal line _{39), e x (V bN} , M) is a vector V _bN, prediction error for _M at the point _x, e x (V ( x)) is the prediction error of the point x with respect to the vector V (x), and Σ represents the sum of all the points x included in the block b _{N, M.} Mean prediction error e _bN of one pixel _{unit, M} is output by the signal line 41 is calculated by the block matching unit 17, an average prediction error e _r relates the original image, processing
Of the block b _{N, M} (signal line 39) to be
Vector field V (x) (signal line 35) and 1 pixel matching
Error function e _x ([Delta] x) (signal line 37) and add it to the threshold T _q are calculated by the threshold generator 14 which has received the signal line 36
Is output (S6).

In the division judging unit 15, the average prediction error e _{b N, M} per pixel received by the signal line 41 and the average prediction error per pixel pertaining to the original image received by the signal line 36.
by comparing the e _r + T _q obtained by adding the threshold value T _q to e _r, e
_{bN, M> e r + T} q and, when the hierarchical level N is not in the lowest hierarchy level N _L and (S7: NO), increments the hierarchical level N (N = N + 1), the division instruction is output by the signal line 38 , The upper left block of the four-divided block is designated as M (N) = 0, and the block b _{N, M} is divided by the quad tree generator 18 to obtain a step S
The operation returns to operation 4 (S8).

In the quad tree generator 18,
_{e bN, M ≦ e r +} T when that is the _q or N = N _L (S7Y), a representative vector V _m V _bN, and _M, and the hierarchy level N _Vm and N, the signal representative vector component V _R Line 4
4 and outputs the hierarchical level N to the signal line 45. And
The number indicating the m-th representative vector is set to m + 1 (S9, FIG. 4), and the block position M (N) is advanced to M (N) +1 (S9).
10), if M (N) = 4 (S11Y), N = N−
The flag fl is set to 1 to indicate that recursive processing is necessary for block determination processing (S12). When M (N) is not 4 (S11N), the hierarchical level N is set to 0. It is checked whether there is any data (S13). If it is not 0, the process returns to the operation of step S4 (S13N). If it is 0, it means that all the divisions have been completed, and the process is terminated.

[0042]

As is clear from the above description, according to the present invention, by determining the generation order of quad trees, additional information for region description is required only at the hierarchical level, and the representative vector Since the extraction order retains the positional continuity, the amount of information can be reduced by differential encoding using the continuity of the vector, and the effect of the present invention is extremely large.

[Brief description of the drawings]

FIG. 1 is a circuit diagram showing an embodiment of the present invention.

FIG. 2 is a representative vector extraction diagram for explaining the principle of a representative vector extraction operation by a quad tree having the circuit configuration of FIG. 1;

FIG. 3 is a flowchart showing a flow of an operation of extracting a representative vector by a quad tree having the circuit configuration of FIG. 1;

4 is a flowchart showing a flow of an operation of extracting a representative vector by a quad tree having the circuit configuration of FIG. 1 together with FIG. 3;

FIG. 5 is a flowchart showing details of a part of the operation in FIG. 3;

FIG. 6 is a block diagram of a quad tree in the operation of FIG. 3;

FIG. 7 is a motion estimation diagram based on conventional one-pixel matching.

FIG. 8 is a distance conversion diagram of a conventional label image.

[Explanation of symbols]

Reference Signs List 11 error plane calculator 12 memory 13 original vector field generator 14 threshold value generator 15 division determiner 16 block selector 17 block matching unit 18 quad tree generator 31-45 signal line 70 screen 71 yen 79 Pixels not removed 80 Original vector field 82 Circle 83 Square 84 Divided block 85 Overlapping part 90 Final division result of original vector field 80 92 Circle corresponding to circle 82 93 Square corresponding to square 83 b _{N, M} block d distance e _{bN, M,} e _r 1 average prediction error of the pixel units e _x (Δx) 1 pixel error function F operator fl flag M (1) of the matching, M (2), M ( N) block position N, N _Vm hierarchy level N _L lowest hierarchy level T _q threshold V (x) vector field (or vector at position of point x) V _m m-th Representative vector x position

──────────────────────────────────────────────────続 き Continuation of the front page (72) Inventor Mitsutoshi Hatori 1-6-24-808 Sengoku, Bunkyo-ku, Tokyo (56) References JP-A-1-69182 (JP, A) (58) Fields surveyed ( Int.Cl. ⁶ , DB name) H04N ^7/ 24-7/68

Claims (6)

(57) [Claims] 1. A motion vector generation process for obtaining a motion vector field (35) in units of one pixel between frames of a moving image (11, 12, 13), and the one pixel in a block (39) to be processed. The processing to be performed using the prediction error obtained from the unit motion vector field (35) and the error function (37) of one-pixel matching.
One pixel of the original image calculated in each block
A threshold generation process for obtaining an average prediction error in units and generating a threshold (36) that changes according to the average prediction error is performed (1
4), the position (43) of the block to be processed and the hierarchical level (4
5) to select the block (39) to be processed.
(16), and the block (39) to be processed is designated, and
One-pixel matching error function corresponding to the indicated block
According to the number (42), the motion vector (40) and the
Block matching to find the average prediction error (41)
Processing (17), and obtained in the block matching processing (17).
When the average prediction error (41) exceeds the threshold (36), the representative motion vector (44) is selected and the hierarchical level (45) of the division is set while the motion vector field is divided by the quad tree. Representative vector (4
Perform quad tree processing as additional information in 4) ( 1)
5,18 ) A motion vector search method. 2. An inter-frame prediction of a moving image encoding,
Error function of pixel matching (33, 34, 37, 42)
And a motion vector generation process for obtaining a motion vector field (35) in units of one pixel (11, 12, 13), the position (43) of the block to be processed and the hierarchical level (4)
5) performing a block selection process for selecting a block (39) to be processed from (16), and instructing the block (39) to be processed, the one-pixel matching corresponding to the designated block; A block matching process for obtaining a motion vector (40) and an average prediction error (41) in units of one pixel by an error function (42) of (17) is performed. A threshold generation process is performed to calculate an average of prediction errors from the error function (37) of pixel matching and the motion vector field (35) in units of one pixel and to generate an adaptive threshold (36) according to the average prediction error (14). ), Comparing the average prediction error (41) for each pixel with the adaptive threshold (36) to determine whether or not to perform quadtree division, and issue a division command (38). (15) receiving the division instruction (38) and the motion vector (40), the position (43) of the block to be processed for quad tree generation and the hierarchy level ( 45) and a quadtree generation process for outputting the motion vector (40) received at the end of the quadtree division as a representative motion vector (44) (18). 3. The threshold generation process (14) sets the block (39) to be processed to b _{N, M} ,
Error function of pixel matching (37) and e _x ([Delta] x), the one pixel of the motion vector field a (35) V
And (x), the prediction error of the motion vector field V (x) and _{e x (V (x))} , any threshold constant and T _q, x
Is the position of a pixel included in the block b _{N, M} to be processed _, the total number of pixels included in the block b _{N, M} to be processed is S (b _{N, M} ), and the adaptive threshold (36) is e _r
+ When the _{T q, e r = (1} / S (b N, M)) Σe x (V (x)) where Σ is for the position x of the pixel included in the block b _{N, M} be the processing 3. The motion vector search method according to claim 2, including a calculation process represented as a cumulative sum. 4. A motion vector means (1) for obtaining a motion vector field (35) in units of one pixel between frames of a moving image.
And 1,12,13), to processing using the prediction error obtained from the error function of the one pixel unit of the motion vector field (35) and 1 pixel matching in the block (39) to be treated (37) Bekiso
One pixel of the original image calculated in each block
Threshold generating means for generating a threshold value (36) that varies according to the average prediction error calculating an average prediction error of a unit (14)
, The position of the block to be processed (43) and the hierarchical level (4
5) to select the block (39) to be processed.
For selecting a block (39) for processing and a block (39) to be processed.
One-pixel matching error function corresponding to the indicated block
According to the number (42), the motion vector (40) and the
Block matching to find the average prediction error (41)
Means (17) and the block matching means (17)
The average prediction error (41) exceeds the threshold (36)
Sometimes , while dividing the motion vector field by the quad tree, the representative motion vector (44) is selected and the hierarchical level (45) of the division is set to the representative vector (4).
Quad tree means ( 15, 1 ) as additional information of 4)
8 ) A motion vector search device comprising: 5. A motion vector generating means (11, 12) for obtaining an error function (33, 34, 37, 42) of one-pixel matching between video encoded frames and a motion vector field (35) in pixel units. , 13), the position of the block to be processed (43) and the hierarchical level (4
5), a block selecting means (16) for selecting a block (39) to be processed, and a block (39) to be processed is specified, and the one-pixel matching corresponding to the specified block is performed. A block matching means (17) for calculating a motion vector (40) and an average prediction error (41) in units of one pixel by an error function (42); A threshold generation means (14) for calculating an average of prediction errors from the error function (37) of the above and the motion vector field (35) in units of one pixel and generating an adaptive threshold (36) according to the average prediction error; A comparison is made between the average prediction error (41) for each pixel and the adaptive threshold (36) to determine whether or not quadtree division is to be performed, and to issue a division instruction (38). The position (43) of the block to be processed for quad tree generation in response to the determination means (15), the division instruction (38), and the motion vector (40), and the hierarchy level (45) A quad tree generating unit (18) for outputting the motion vector (40) received at the end of the quad tree division as a representative motion vector (44). 6. The threshold generation means (14) sets the block (39) to be processed to b _{N, M} ,
Error function of pixel matching (37) and e _x ([Delta] x), the one pixel of the motion vector field a (35) V
And (x), the prediction error of the motion vector field V (x) and _{e x (V (x))} , any threshold constant and T _q, x
Is the position of a pixel included in the block b _{N, M} to be processed _, the total number of pixels included in the block b _{N, M} to be processed is S (b _{N, M} ), and the adaptive threshold (36) is e _r
+ When the _{T q, e r = (1} / S (b N, M)) Σe x (V (x)) where Σ is for the position x of the pixel included in the block b _{N, M} be the processing 6. The motion vector search device according to claim 5, which performs a calculation operation represented as a cumulative sum.