KR20040084327A - Method for motion estimation based on adaptive search range - Google Patents

Abstract

PURPOSE: A motion prediction method using an adaptive search region is provided to efficiently perform motion prediction by adaptively determining a search region with reference to the characteristic of an input moving picture. CONSTITUTION: The first macro block group formed as certain macro blocks among a plurality of macro blocks for motion prediction is generated(220). Motion information about the macro blocks included in the generated first macro block group is calculated(240). A search region is determined using calculated motion information of the macro blocks included in the first macro block group to calculate motion information of macro blocks included in the second macro block group(260).

Description

Method for motion estimation based on adaptive search range

The present invention relates to a motion prediction method for a video compression coding standard in a low bit rate environment. In particular, an adaptive search region determination method for adaptively determining a search region in consideration of motion vectors of adjacent macroblocks and motion prediction using the same It is about a method.

The rapid development of communication technology is increasing the demand for multimedia services in mobile communication. However, it is very difficult to stably and quickly serve a clear image in a mobile communication channel environment where only a relatively small amount of data can be transmitted. In video coding, motion compensation coding that removes temporal redundancy is widely used to obtain high data compression rate, which is an important part of international video coding standards such as MPEG and H.263. The motion compensation encoding technique is a method of taking a macroblock most similar to a current macroblock from a previous frame through motion prediction, and transcoding the difference between the predicted image and the input image.

The present invention proposes an algorithm that improves the performance of the motion vector estimation algorithm in the MPEG encoder as shown in FIG.

In the encoder shown in Fig. 1, firstly input image data is decomposed into blocks of 8x8 pixels. The DCT (Discrete Cosine Transform) unit 110 performs a DCT operation on image data input in units of 8 × 8 pixel blocks to remove spatial correlation. The quantization unit (Q) 120 performs quantization on the DCT coefficients obtained by the DCT unit 120 and expresses them as some representative values, thereby performing high efficiency lossy compression. The variable length coding unit (VLC) 130 performs entropy encoding on the quantized DCT transform coefficients and outputs an entropy-coded data stream.

An inverse quantization unit (IQ) 140 inversely quantizes the image data quantized by the quantization unit 120. The inverse DCT unit 150 performs IDCT transformation on the image data dequantized by the inverse quantization unit 140. The frame memory unit 160 stores image data converted by the IDCT unit 150 in units of frames. The motion estimation unit (ME) 170 is used to remove temporal correlation by using the image data of the input current frame and the image data of the previous frame stored in the frame memory unit 160.

The motion prediction technique used in the motion estimation unit 170 of FIG. 1 includes a block matching algorithm (BMA). The block matching algorithm is a method of estimating block-by-block motion from a previous frame on the assumption that all pixels in a block have the same motion vector. The most widely used block matching algorithm is the Full Search BMA (FS-BMA). The global search technique finds the best motion vector because it finds the most similar block in the search area of the previous frame, but the search area needs to calculate Mean Absolute Difference (MAD) and Mean Square Error (MSE) for each search position. As the size of blocks used for searching and searching increases, the amount of calculation increases exponentially. Many high speed motion prediction techniques have been proposed since such large computations make it difficult to process and transmit images in real time.

Representative fast motion prediction algorithms include three-step hierarchical search (3SHS), two-dimensional logarithmic search (LOGS), and one-dimensional full search (ODFS). ), And the fast motion prediction method (MVFAST) adopted as the fast motion prediction technique in MPEG-4 part 7 are widely known. In addition, there are various high speed motion prediction techniques as a variant of the fast motion prediction technique. However, these high-speed motion prediction techniques must consider factors such as bit rate, speed, performance, algorithm complexity, memory size and bandwidth, and must select appropriate elements among these high-speed motion prediction techniques.

However, these fast search methods are algorithms assuming that frames are unimodal, so there is an error that is likely to fall into local minimum in real images, and is significantly better than global search techniques in large motion images. There was a problem of this deterioration.

The technical problem to be achieved by the present invention is to provide an improved motion prediction method capable of efficient motion prediction by adaptively determining the search area in consideration of the characteristics of the input video to solve the problems in the prior art will be.

1 is a block diagram illustrating a conventional video encoding system.

2 is a flowchart for explaining a method for determining an adaptive search region according to an embodiment of the present invention.

3 is a diagram illustrating a macroblock motion prediction sequence in an adaptive search region determination method according to an embodiment of the present invention.

4 illustrates a surrounding macroblock for determining an adaptive search region according to an embodiment of the present invention.

5 illustrates an adaptive search region determined in accordance with an embodiment of the present invention.

6 is a view for explaining a method of determining an initial motion vector used in an embodiment of the present invention.

7 is a view for explaining a method for determining an initial motion vector used in an embodiment of the present invention.

8 is a diagram for explaining a motion prediction search algorithm used in an embodiment of the present invention.

The technical problem is a motion prediction method using an adaptive search region according to the present invention, generating a first macroblock group consisting of a predetermined macroblock of a plurality of macroblocks for motion prediction, and the generation Calculating motion information for the macrogroups belonging to the first macroblock group, and using the calculated motion information of the macroblock belonging to the first macroblock group, And determining a search area for calculating motion information, wherein the second macroblock group includes remaining macroblocks except macroblocks belonging to a first macroblock group among a plurality of macroblocks for motion prediction. It is achieved by the way it is done.

Strong correlation exists between frames immediately adjacent in time. There is also a significant spatial correlation between the motion vector of the current macroblock in the current frame and the motion vector of the surrounding macroblocks. Using the correlation between the motion vector of the current macroblock and the motion vector of the macroblock at the same position of the previous frame and the motion vectors of the surrounding macroblocks having a spatial correlation with the motion vector of the current macroblock The motion vector may be estimated.

The present invention is characterized by determining an adaptive search region for performing motion prediction using the temporal and spatial correlation of such a motion vector, and performing a fast motion prediction method based on the determined adaptive search region.

In the motion prediction method according to the present invention, the search area for performing the motion prediction of the macroblock is newly defined after the search area is newly defined using the motion vector of the surrounding macroblock instead of the search area in the conventional global search technique. The motion prediction is performed only in the region.

The search area creates a minimal rectangle that contains the motion vectors of the surrounding macroblocks. Since the motion vectors of the surrounding macroblocks all have similar values due to their spatial correlation, the search area can be very small. Therefore, the search area needs to be adjusted to compensate for the risk of falling into the local minimum. . The motion prediction is performed only in the newly created search area because the motion correlation of the current macroblock is similar to the motion vector of the surrounding macroblock because the motion of the object is continuous because the spatial correlation in the frame is used.

Also, in determining the initial motion vector of the current macroblock, the motion vector of the macroblock at the same position of the previous frame and the motion vector of the surrounding macroblock are used.

Unlike the global search method, the search method reduces the amount of computation compared to the global search method by applying the diamond search method centered on the initial motion vector.

Hereinafter, a fast motion prediction method using an adaptive search region according to an embodiment of the present invention will be described with reference to FIGS. 2 to 6.

2 is a flowchart illustrating a motion prediction method according to an embodiment of the present invention.

In operation 220, a first macroblock group consisting of predetermined macroblocks among a plurality of macroblocks for motion prediction is generated. The first macroblock group according to an embodiment of the present invention is composed of macroblocks 1 to 50 of the macroblocks shown in FIG. 3. However, it is also possible to optionally form the first macroblock group from other macroblocks.

In operation 240, a motion vector is calculated for macroblocks belonging to the first macroblock group among the macroblocks shown in FIG. 3, that is, macroblocks 1 to 50. In the present embodiment, a motion vector is calculated for blocks belonging to the first macroblock group by using a search area having the same size as the search area in the global search technique.

In operation 260, a motion vector is calculated for macroblocks that do not belong to the first macroblock group among the macroblocks shown in FIG. 3, that is, macroblocks 51 to 99. In the present embodiment, the motion vector is obtained by adjusting the size of the search area by the motion vectors of the four surrounding macroblocks obtained in operation 240.

For example, the search region of macroblock 57 uses motion vectors of four neighboring macroblocks, as shown in FIG.

Among the macroblocks shown in FIG. 3, motion vectors of macroblocks ₂ , ₇ , ₈ , ₁₃ are ( x ₂ , y ₂ ), ( x ₇ , y ₇ ), ( x ₈ , y ₈ ), and ( x ₁₃ , y ₁₃ ), the search region of the macroblock 57 is calculated using Equations 1 to 4 below.

[Equation 1]

maximum x = MAX ( x ₂ , x ₇ , x ₈ , x ₁₃ ) + 2

[Equation 2]

minimum x = MIN ( x ₂ , x ₇ , x ₈ , x ₁₃ )-2

[Equation 3]

maximum y = MAX ( y ₂ , y ₇ , y ₈ , y ₁₃ ) + 2

[Equation 4]

minimum y = MIN ( y ₂ , y ₇ , y ₈ , y ₁₃ )-2

In Equation 1 to 4, the reason why ± 2 at each value is that the local motions in the frame are almost the same, so that when the motion vectors have similar values, the search area is too small to prevent the local minimum from falling. .

Optionally, Equations 5-8 are used to ensure that the search region obtained primarily by Equations 1-4 does not deviate from the search region of the global search technique, for example (-16, 15).

[Equation 5]

search_ x _max = MIN (15, maximum x )

[Equation 6]

search_ x _min = MAX (-16, minimum x )

[Equation 7]

search_ y _max = MIN (15, maximum y )

[Equation 8]

search_ y _min = MAX (-16, minimum y)

In the method for determining an adaptive search region according to the present invention, four values obtained by using the results of applying Equations 5 to 8 to the results of Equations 1 to 4 or 1 to 4 are used. We create the smallest rectangle containing these four points, which is the new search area. 5 shows a new search region obtained using the adaptive search region determination method according to the present invention.

In operation 280, motion prediction is performed only within the new search region obtained in operation 260 for macroblocks belonging to the second macroblock group. Due to the spatial correlation of macroblocks in a frame, since the motion vectors of neighboring macroblocks are the same, since the motion prediction is performed only in a certain region including the motion vectors of neighboring macroblocks, an existing search method, for example, Compared to the global search method, it is possible to reduce the amount of computation for motion prediction.

As described above, the motion estimation algorithm using the adaptive search region according to the present invention is an algorithm that effectively reduces the search point by using the fast sum of absolute algorithm and variably adjusting the search region by the surrounding motion vector. The adaptive search range (ASR) algorithm utilizes a property in which a motion vector of one macroblock has much similarity with a motion vector of a neighboring macroblock. That is, local movement in one frame has the same movement.

The difference between the ASR algorithm and the existing global search or other fast algorithms is the order in which the motion of macroblocks in a frame is predicted. In the global search technique, motion prediction is performed in the order from left to right and top to bottom from the macroblock located at the upper left of the frame. In the ASR algorithm, the motion prediction is performed in the same order as in FIG. 3.

Optionally, the motion prediction method according to the present invention detects a static macroblock in a corresponding frame and does not perform motion prediction on the detected static macroblock. This is done through the following process.

These macroblocks only need to calculate the SAD at (0,0) because the motion vector of many macroblocks on the video sequence that is not motiony has (0,0). If the SAD at (0,0) is smaller than the threshold T , then the motion vector of this macroblock is (0,0) and no further motion prediction is performed. That is, the macroblock is determined as a static macroblock, and motion prediction is not performed on the macroblock. In one embodiment, the threshold T is defined as QP × 20, QP is fixed at 6 for I frames, and P frames vary from 6-31.

In addition, as in the conventional fast search algorithm, the selection of the initial motion vector is very important in the adaptive search region algorithm. In the adaptive search region algorithm, the initial motion vector can be largely divided into two types.

First, in the case of macroblocks 1 to 50, since there is no motion vector to refer to since the motion prediction of the surrounding macroblocks has not been performed yet, as shown in FIG. Let the motion vector of the block be the initial motion vector.

In addition, the initial motion vectors of macroblocks 51 to 99 use the motion vectors of the surrounding macroblocks, as shown in FIG. 7. For example, the initial motion vectors of the macroblocks 57 and 57 are represented by Equations 9 and 10. You can get it.

[Equation 9]

initial MV_ x = MEDIAN ( x ₇ , x ₂ , x ₅₂ )

[Equation 10]

initial MV_ y = MEDIAN ( y ₇ , y ₂ , y ₅₂ )

As described above, by adopting the method according to the present invention, the initial motion vector in the macroblock having the adaptive search region can be approached to the motion vector to obtain the maximum, thereby reducing the amount of computation, thereby reducing the time. If the motion of the surrounding macroblock and the current macroblock in the large image is different, the global search method may be performed to improve the image quality.

A search method in the ASR algorithm according to an embodiment of the present invention can be divided into two types. For example, macroblocks 1 to 50 perform the diamond search method shown in FIG. 8. The diamond search method is performed in the following order. Hereinafter, two types of diamond search methods used in the present invention, namely, a small diamond search (SDS) method and a large diamond search (LDS) method will be described.

Small Diamond Search (SDS)

Step 1: Compute the SAD of five search points including the initial motion vector. If the SAD of the point at the center is the smallest, that point is the motion vector, otherwise switch to step 2.

Step 2: The center point of the SDSP is moved to the smallest point of the SAD obtained in step 1, and the SAD of all points on the SDSP is calculated. If the SAD of the centered point is the smallest, that point is the motion vector to be found, otherwise continue with step 2.

Large Diamond Search (LDS)

Step 1: Compute the SAD of nine search points including the initial motion vector. If the SAD of the centered point is the smallest, switch to step 3, otherwise switch to step 2.

Step 2: The center point of the LDSP is moved to the smallest point of the SAD obtained in Step 1, and the SAD of all points on the LDSP is calculated. If the SAD of the center point is the smallest, switch to step 3, otherwise continue with step 2.

Step 3: Switch the search method from LDSP to SDSP and the smallest SAD of the five points is the motion vector to be obtained.

In the following, the diamond search method according to the present invention will be described using the LDS method and the SDS method described above.

First, according to the LDSP, SADs of nine search points including an initial motion vector are calculated.

As a result of the LDSP search, when the SAD of the center point is the smallest, the search method is switched from LDSP to SDSP, and the point whose SAD is the smallest among the five points is the motion vector.

As a result of the LDSP search, if the SAD of the point at a location other than the center is the smallest, move the center point of the LDS to the point with the smallest SAD, calculate the SAD of all points on the LDSP, and, as a result, the SAD of the centered point If is smallest, switch the search method from LDSP to SDSP, and the smallest SAD of the five points is the motion vector to be obtained. As a result, if the SAD of the point at a position other than the center is the smallest, the process is repeated.

The present invention is not limited to the above-described embodiment, and of course, modifications may be made by those skilled in the art within the spirit of the present invention.

The invention can also be embodied as computer readable code on a computer readable recording medium. The computer-readable recording medium includes all kinds of recording devices that store data that can be read by a computer system. Examples of computer-readable recording media include ROM, RAM, CD-ROM, magnetic tape, hard disk, floppy disk, flash memory, optical data storage device, and also carrier waves (for example, transmission over the Internet). It also includes the implementation in the form of. The computer readable recording medium can also be distributed over network coupled computer systems so that the computer readable code is stored and executed in a distributed fashion.

As described above, according to the present invention, instead of using a search region that performs motion prediction in an existing algorithm, a new search region is created using a motion vector of surrounding macroblocks, and then motion prediction is performed only within the new search region. Compared to the conventional search method, the computational amount can be significantly reduced, and the time can be shortened by reducing the computational amount by approaching the motion vector to obtain the maximum initial motion vector in the macroblock with the adaptive search region. For example, if the motion of the surrounding macroblocks and the current macroblocks is different in a large motion image, global search is performed to improve image quality.

Claims (7)

In the motion prediction method using the adaptive search region, (a) generating a first macroblock group consisting of predetermined macroblocks among a plurality of macroblocks for motion prediction; (b) calculating motion information on macro groups belonging to the generated first macroblock group; (c) determining a search region for calculating motion information of a macroblock belonging to a second macroblock group by using motion information of the macroblock belonging to the first macroblock group; The second macroblock group is composed of the remaining macroblocks except for the macroblock belonging to the first macroblock group of the plurality of macroblocks for the motion prediction. The method of claim 1, further comprising: (d) calculating motion information of a macroblock belonging to a second macroblock group by using the search region determined in step (c). The method of claim 1, wherein the 4-connected neighborhood blocks of the macroblock belonging to the first macroblock group are macroblocks belonging to the second macroblock group. The method of claim 1, wherein the calculating of motion information of the macroblock of the first macroblock group is performed using a global search method. The method of claim 1, wherein the four access neighboring blocks of the macroblock belonging to the second macroblock group are macroblocks belonging to the first macroblock group. The method of claim 1, wherein the motion information of the macroblock belonging to the second macroblock group is calculated based on the motion information of the macroblocks belonging to the first macroblock group adjacent to the macroblock belonging to the second macroblock group. A motion prediction method characterized by the above. The method of claim 1, wherein the motion information of the second macroblock group macroblock is calculated based on motion information of four adjacent blocks adjacent to the macroblock belonging to the second macroblock group.