KR100431075B1 - A fast method of estimating block motion vectors - Google Patents

Abstract

The present invention relates to a fast block matched motion estimation method for a video encoder. More specifically, the present invention relates to a fast motion estimation technique that is excellent in motion estimation performance and is applicable to real-time applications as well as low motion images.

The present invention provides a one-dimensional feature matching technique (1DFM) using spatial correlation of motion vectors, comprising: a first step of obtaining an additive projection of horizontal or vertical components with respect to a macroblock of a current frame; A second step of predicting a horizontal or vertical component of the motion vector of the current macroblock using already calculated surrounding motion vectors; Obtaining an additional projection of the horizontal or vertical component with respect to the pixels of the candidate macroblocks corresponding to the position indicated by the horizontal or vertical component predicted in the second step with respect to the search region of the previous frame; A fourth step of obtaining a horizontal or vertical component of a motion vector having a minimum matching error by using addition projections of the horizontal or vertical components obtained in the first and third steps; A fifth step of obtaining an additive projection of a horizontal or vertical component only of pixels of a part of the search area by using the horizontal or vertical component of the motion vector obtained in the fourth step; A sixth step of obtaining a horizontal or vertical component of a motion vector having a minimum matching error by using addition projections of the horizontal or vertical components obtained in the first and fifth steps; And a seventh step of obtaining a final motion vector using the horizontal or vertical components obtained in the fourth and sixth steps.

Description

A fast method of estimating block motion vectors for video encoder

The present invention relates to a fast block matched motion estimation method for a video encoder. More specifically, the present invention relates to a fast motion estimation technique that is excellent in motion estimation performance and is applicable to real-time applications as well as low motion images.

Block-matched motion estimation is an important technique used to remove spatial correlation in today's representative video coding standards such as H.263 and MPEG-1 / 2/4. Full search (FS), a typical block-matched motion estimation technique, is optimal in terms of motion estimation performance, but due to the large amount of computation, it is inevitable to develop fast motion estimation techniques for real-time applications. Many high speed motion estimation techniques have been developed.

However, the conventional techniques show satisfactory performance for moving pictures with little or no motion, but have a disadvantage in that the performance is remarkably degraded for moving or moving pictures.

Since the early 80's, many fast motion estimation techniques have been developed to introduce various techniques to reduce the enormous computational amount of global search technique. Conventional fast motion estimation techniques can be classified into several categories according to the approach to reduce the amount of computation. In order to greatly reduce the amount of computation, the method of reducing the number of search points is assumed to be the most effective, assuming a uniform error surface model having only one minimum point, and the present invention will be applied to this technique.

Since three-step search (TSS) was developed in 1981, a variety of similar high-speed motion estimation techniques have been developed. In order of development age, they are as follows.

TSS (three-step search)

OTS (one-at-a-time)

NTSS (new three-step search)

1DFS (one-dimensional full search)

4SS (four-step search)

Block-based gradient descent search (BBGDS)

SES (simple and efficient search)

UCBDS (unrestricted center-biased diamond search)

AMT (adaptive motion tracking BMA)

MIBAS (center-biased minima bounded area search)

1DGDS (one-dimensional gradient descent search)

All of the above techniques have a procedure of gradually finding a point where the matching error is minimal by repeatedly searching for a specific search pattern. Two useful characteristics used for effective search are: first, statistical characteristics that the distribution of motion vectors in the real image is concentrated at (0, 0) points, and second, the motion vectors between adjacent blocks are The direction is similar to that of.

The TSS is a technique that does not consider the distribution characteristics of such a motion vector at all, and the technique using the first characteristic is OTS, NTSS, 4SS, BBGDS, UCBDS, MIBAS, etc., and the technique using both the first and second characteristics AMT, 1DGDS, and the like. However, the above techniques have satisfactory performance for images with little or no motion, but have a performance that is far less than global search for large or complex images.

On the other hand, feature matching has been considered to extract the feature of the block and use it for motion estimation. There are various features such as contour information of an object, but among the techniques related to the present invention, there are techniques using integrated projection as follows.

PTSS (projection three-step search)

Two-stage BMA using integral projections

FBMA (feature BMA) and NFBMA (new FBMA)

HPM (efficient BMA using integral projections)

These techniques reduce the amount of computation by using integral projections instead of the sum of absolute differences (SAD), which are commonly used in block-matched motion estimation. Here, the cumulative projection used refers to creating a one-dimensional array by adding all the pixel values that meet and progress in the two-dimensional image in a given direction.

In PTSS, addition projection is introduced for motion estimation for the first time and the computation is reduced by changing the matching criteria to addition projection in the existing TSS technique. Two-stage BMA using integral projections is a method of finding the final motion vector by performing SEAR over nine search points around the determined motion vector after performing global search using additive projection.

In addition, FBMA (feature BMA) and NFBMA (new FBMA) techniques, like two-stage BMA using integral projections, perform global search using additive projection, select multiple search points in the order of least matching error, and then select them. SAD is obtained for each and finds the minimum point. HPM has three hierarchical search steps, and each step adds all the pixel values of the search area to convert the two-dimensional pixel area search into two one-dimensional horizontal and vertical addition projection areas. This is how it is done.

However, since the PTSS is based on the TSS, the performance is greatly reduced, and the FBMA (feature BMA) and NFBMA (new FBMA) and the two-stage BMA using integral projections perform a global search using additive projection, so The decrease is minimal. In addition, the HPM has a disadvantage in that the motion estimation is inaccurate by adding and projecting all pixel values in the search area. As a result, none of the above techniques could fully utilize the addition projection to reduce the computational efficiency and increase the performance.

Accordingly, an object of the present invention is to propose a technique for performing feature matching in a one-dimensional addition projection region by selectively adding and projecting pixels in a search region using spatial correlation of motion vectors. In addition, this method is combined with gradient descent search using surrounding motion vectors to improve the speed and performance of motion estimation.

As a technical idea for achieving the above object of the present invention

A motion estimation method of fast block matching using 1-dimensional feature matching (1DFM) and two-step plus pattern searcher (2SPPS), comprising: a first step of obtaining additive projection of horizontal or vertical components of a macroblock of a current frame; ; A second step of predicting a horizontal or vertical component of the motion vector of the current macroblock using surrounding motion vectors; A third step of obtaining addition projections for pixels of candidate macroblocks corresponding to the horizontal or vertical components predicted in the second step with respect to the search area of the previous frame; A fourth step of obtaining a horizontal or vertical component of a motion vector having a minimum matching error by using the addition projections obtained in the first and third steps; A fifth step of obtaining an additive projection of horizontal or vertical components of only a portion of pixels in the search area by using the motion vector obtained in the fourth step; A sixth step of obtaining a horizontal or vertical component of a motion vector having a minimum matching error by using the addition projections obtained in the first and fifth steps; And a seventh step of obtaining a final motion vector using the horizontal or vertical components obtained in the fourth and sixth steps.

In addition, the two-step plus pattern search technique includes: a first step of obtaining SAD values at respective points with respect to a (+)-shaped search pattern composed of a search initial point and four surrounding points from a surrounding motion vector; Stopping the search if the location of the point where the SAD is minimum is the center of the search pattern; And a third step of performing a search by moving the center of the search pattern to the point where the SAD is the minimum when the position of the point having the smallest SAD is at least one of four surrounding points. Provide a method.

1 is a schematic diagram showing peripheral motion vectors used in a fast block matched motion estimation technique according to the present invention.

2A and 2B are schematic diagrams showing a selective addition projection method in a one-dimensional feature matching technique according to the present invention.

3A and 3B are schematic diagrams illustrating a two-step plus pattern search (2SPPS) technique in accordance with the present invention.

Hereinafter, with reference to the accompanying drawings, the configuration and operation of the embodiment of the present invention will be described in detail.

1 is a schematic diagram showing peripheral motion vectors used in a fast block matched motion estimation technique according to the present invention. 2A and 2B are schematic diagrams showing a selective addition projection method in a one-dimensional feature matching technique according to the present invention. 3A and 3B are schematic diagrams illustrating a two-step plus pattern search (2SPPS) technique in accordance with the present invention.

In the above, the conventional techniques for reducing the amount of motion estimation are discussed. As in the techniques described above, the method of finding the matching error at the search points forming a specific pattern from the search initial point, and performing the search by gradually moving the search pattern in the direction of minimizing the matching error, selects the initial search point. It has a big influence on the performance and computation of motion estimation.

The present invention describes a motion estimation method that can improve performance and reduce computation by providing a search initial point closer to the global minimum. That is, the present technique consists of two main parts: one-dimensional feature matching (1DFM) and two-stage plus pattern search (2SPPS) using selective addition projection using the correlation of motion vectors.

First, a 1D feature matching technique (1DFM) using selective addition projection using a correlation of a motion vector will be described.

In the present invention, in order to significantly reduce the amount of computation and to maintain the accuracy of the motion estimation, compared to the conventional techniques using addition projection, only pixels corresponding to a region which is likely to be determined as a motion vector among the pixels in the search region of the previous frame are maintained. We propose a selective addition projection. The 1DFM procedure of the present invention is as follows. In the present invention, N is the size of the macroblock which is the basic unit for which motion estimation is performed, and p is the size of the search region.

<First step>

; Arrangement of size N × 1 by finding addition projections for the horizontal and vertical directions of the macroblock (size N × N) at the position ( i , j ) of the current frame. Get

<Second Step>

; The vertical (horizontal) component of the motion vector of the current macroblock using the already calculated neighboring motion vectors. To predict. The motion vectors of the current macroblock can be predicted using the methods of Equations 1 to 3 using the already obtained peripheral motion vectors defined in FIG. 1. Mean means mean value and Median means mean value. In addition, the width of the area used for selective addition projection may be used in a variable size other than the width of the macroblock.

<Third Step>

; Navigation area of previous frame (size Pixels corresponding to candidate blocks corresponding to positions indicated by the vertical (horizontal) component predicted in the second step with respect to (In the vertical direction; see FIG. 2A) or (In the horizontal direction; see Fig. 2b) only by adding in the horizontal (vertical) direction Array of sizes Get

The position where the matching error defined by Equation 4 and Equation 6 (or Equation 5 and Equation 7) is minimized by using the vertical (horizontal) direction addition projections obtained in the first and third steps. , The horizontal (vertical) component of the motion vector, Obtain

<5th Step>

; The horizontal (vertical) component of the motion vector obtained in the fourth step, As in the third step, only pixels of a part of the search area are added and projected in the vertical (horizontal) direction using Array of sizes Get

<The sixth step>

; A position where the matching error defined by Equation 5) and Equation 7 (or Equation 4 and Equation 6) is minimized using the horizontal (vertical) addition projection obtained in the first and fifth steps, That is, the vertical (horizontal) component of the motion vector Obtain

<Step 7>

; The final motion vector is obtained from steps 4 and 6 Is given by

Next, a two-step plus pattern search (2SPPS) will be described.

Starting from the initial search point obtained from 1DFM or (0,0) and the surrounding motion vectors, a search for the surrounding search points is performed to obtain a more accurate motion vector. Unlike the 1DFM process, the matching criteria uses a general SAD for motion estimation. The procedure of the 2SPPS technique is as follows.

<First step>

; As shown in FIG. 3A (the basic pattern of 2SPPS) and FIG. 3B (the pattern of the next step according to the minimum SAD point of the previous step), the (+) search pattern of the initial search point and four surrounding points is shown. Find the SAD value at each point.

<Second Step>

; If the position of the point where the result SAD is the minimum in the first step is the center of the search pattern, the search is stopped. In this case, the final motion vector is the same as the search initial point. If the position of the point with the smallest SAD is one of the four surrounding points, go to step 3.

<Third Step>

; Search again by moving the center of the search pattern to the point where SAD is minimum. In this case, you only need to find the SAD for the three new points that do not overlap the previous search pattern, no matter where the point with the smallest SAD is. The point where SAD is minimum becomes the final motion vector.

Finally, a fast motion estimation technique using the 1DFM and 2SPPS techniques will be described. The motion estimation is performed by combining the 1DFM and 2SPPS techniques described above. The performance of the motion estimation is improved by additionally using the points obtained by the 1DFM method as the initial point of search for the 2SPPS method.

As described above, according to the fast block matched motion estimation method for a video encoder according to the present invention, it is possible to apply a motion estimation technique with excellent performance and high speed to not only images having low motion but also large and complex images. Thus, speed and performance of various real-time video compression applications can be improved.

Claims (6)

In the motion estimation method of fast block matching using one-dimensional feature matching (1DFM) and two-step plus pattern searcher (2SPPS), Obtaining addition projections of horizontal or vertical components with respect to the macroblock of the current frame; A second step of predicting a horizontal or vertical component of the motion vector of the current macroblock using surrounding motion vectors; A third step of obtaining addition projections for pixels of candidate macroblocks corresponding to the horizontal or vertical components predicted in the second step with respect to the search area of the previous frame; A fourth step of obtaining a horizontal or vertical component of a motion vector having a minimum matching error by using the addition projections obtained in the first and third steps; A fifth step of obtaining an additive projection of horizontal or vertical components of only a portion of pixels in the search area by using the motion vector obtained in the fourth step; A sixth step of obtaining a horizontal or vertical component of a motion vector having a minimum matching error by using the addition projections obtained in the first and fifth steps; And a seventh step of obtaining a final motion vector using the horizontal or vertical components obtained in the fourth and sixth steps. The method according to claim 1, wherein in the second step, additive projection is selectively used using an average value or a median value of horizontal or vertical components of motion vectors of neighboring macroblocks of the current macroblock. . The fast block matched motion estimation method of claim 2, wherein the width of the area used for the selective addition projection is changed to a size other than the width of the macroblock. The method of claim 1, wherein the two-stage plus search method (2SPPS) A first step of obtaining a SAD value at each point with respect to a (+) shape search pattern consisting of a search initial point and four surrounding points from the surrounding motion vector; Stopping the search if the location of the point where the SAD is the minimum is the center of the search pattern; And a third step of performing a search by moving the center of the search pattern to the point where the SAD is the minimum when the position of the point having the smallest SAD is at least one of four surrounding points. Way. 5. The method of claim 4, wherein the SAD is obtained for only three new points that do not overlap with the previous search pattern, regardless of which position the point has the smallest SAD. 6. The method of claim 1 or 4, wherein motion estimation is performed by combining the one-dimensional feature matching method and the two-step plus pattern searcher method.