IE20020426A1 - Video data half-pel motion estimation - Google Patents

Abstract

Half-pel motion estimation for video data is achieved very effectively for low variation of moving average estimates (MAEs). In each of the horizontal and vertical direction a half-pel motion vector (MV) candidate set is selected with reference to MAEs on the integer pel grid. The candidate set is rejected if the difference between the integer MV MAE and an integer grid MAE is greater that a threshold fraction of the integer MV MAE. <Figure 1>

Description

The invention relates to video data half-pel motion estimation.

Prior Art Discussion Video standards such a H.263 and MPEG 4 support use of half-pel motion vectors (MVs) for predicting a current frame from a previous frame with sub-pixel accuracy. Heretofore the MVs have been determined by motion estimation (ME) in which the reference picture is interpolated by a factor of two as shown in Fig. 1. This, however, involves high computational complexity and large storage requirements. Another approach has been to use moving average estimation (MAEs) around the current integer MV. This approach, however, is ineffective under low variation of MAEs where there is low motion and spatial detail.

British Patent Specification No. GB2309135A describes a system for estimating image motion, in which image segments of adjacent frames are compared to generate error values. First and second motion vectors are generated and are summed.

The invention is directed towards providing a method and video data processor for improved motion estimation with lower processing requirements.

SUMMARY OF THE INVENTION According to the invention, there is provided a method for motion estimation in video data processing, the method comprising the steps of:- X -2(a) finding an integer-pel, and (b) determining a half-pel motion vector to yield the smallest moving average estimate on a sub-pixel grid in each of the vertical and horizontal axes, in which the half-pel motion vector is determined by comparison of moving average estimates on the integer-pel.

In one embodiment, the determining step (b) comprises identifying a candidate set of half-pel motion vectors, and subsequently accepting or rejecting the candidate set.

In another embodiment, a candidate set is rejected if the difference between an MAE on the integer grid on the candidate set side and the integer pel MV MAE is less than a threshold value multiplied by the integer pel MV MAE.

In a further embodiment, for each axis, a first candidate set is identified; the first candidate set is accepted or rejected, and if rejected a second candidate set is identified; and said second candidate set is accepted or rejected.

In one embodiment, a default candidate set is identified if the second candidate set is rejected.

In another embodiment, the default candidate set is a set of half-pel motion vectors on an axis extending through a central integer pel.

In a further embodiment, the threshold value is approximately 0.25.

In one embodiment, the threshold selection is made dynamically in response to past variation of MAEs. -3In a further embodiment, a decision is taken before half-pel determination as to whether there should be a single motion vector or four motion vectors per block.

According to another aspect, the invention provides a video data processor 5 comprising means for performing any motion estimation method as defined above.

DETAILED DESCRIPTION QF THE INVENTION Brief Description of the Drawings The invention will be more clearly understood from the following description of some embodiments thereof, given by way of example only with reference to the accompanying drawings in which:15 Fig. 1 is a diagram showing an integer pel grid and a half-pel grid and associated MVs of the prior art; and Figs. 2 to 4 are diagrams showing a method for finding half-pel MVs.

Description of the Embodiments Referring to Figs. 2 to 4, in a motion estimation method of the invention the integer pel MV is located. The MAE corresponding to the integer MV is denoted mO. The pixel values on the integer-pel grid are known, and MAEs corresponding to four neighbouring locations on the integer pel grid as shown in Fig. 2 are denoted as ml, m2, m3, and m4.

Horizontal Half-pel MV selection -4As shown in Fig 3, the following conditions should be met for finding the half-pel MV to yield the smallest MAE on the sub-pixel grid in the horizontal direction. 1. If 2(m3-m0)<(m4-m0), the half-pel MV in the horizontal direction (axis) is one of 5 the three valueson the xl line shown in Fig 3. 2. If((m3-md)>TH*md) ,the selection made in step 1 is wrong due to less variation of MAEs, then the half-pel MV in the horizontal direction is one of the three values on the xO line as shown in Fig 3. “TH” means threshold. 3. If (m3-m0)>2(m4-m0) ,the half-pel MV in the horizontal direction is one of the 10 three values on the x2 line as shown in Fig 3. 4. If ((w4-«z0) . If neither of the conditions (of steps 1 and 3) is met, the half-pel MV in horizontal 15 direction is one of the three values on the xO line as shown in Fig 3. x0,xl,x2 correspond to 0,-1/2,+1/2 MV values respectively.

Vertical Half-pel MV Selection Similarly, to find half pel MVs in the vertical direction (axis) as shown in Fig. 4 the following condition should be met. 1. If 2(ml-mO)<(m2-mO), the half-pel MV in the vertical direction is one of the three 25 values on the yl line as shown in Fig. 4. 2. If ((m/-wd)>TH*OT0, the selection made in step 1 is wrong due to less variation of MAEs. Then the half-pel MV in the vertical direction is one of the three values on the yO line as shown in Fig. 4. -53. If (ml-mOy>2(m2-mO), the half-pel MV in the vertical direction is one of the three values on the y2 line as shown in Fig. 4. 4. If((w2-md) . If neither of the conditions (of steps 1 and 3) is met, the half-pel MV in the vertical direction is one of the three values on the yO line as shown in Fig. 4. y0,yl,y2 correspond to 0,-1/2,+1/2 MV values respectively.

Threshold selection Proper threshold (TH) selection is important to achieve good motion estimation performance. It has been found that TH=0.25 is a good choice for video with low motion and low spatial details. This threshold selection also can be made dynamically depending on the past variation of MAEs. The threshold value depends upon the input video data. Based upon past frame MAEs the threshold value can be selected dynamically.

Single MV/4MVDecision before Half-Pel In the current video standards such as H.263, MPEG2, and MPEG4 there can be a single MV or four MVs for a macro block. Normally the single MV/4MV decision depends on the single MV/4MV MAEs after half-pel calculations. Simplified half-pel ME (prior art) processing does not give correct MAEs after half-pel. Making the single MV/4MV decision after half-pel calculation in case of simplified half-pel ME may result in poor performance because of wrong MAEs. If the single MV /4MV decision is made before half-pel, this yields correct results in the case of simplified -6half-pel ME processing and saves 50% of the computation, because it is sufficient to calculate half pel MVs for single or 4MVs.

It will be appreciated that the invention provides very effective motion estimation 5 even where there is low variation of MAEs i.e. less spatial and low motion video sequences. Steps 2 and 4 for the horizontal and vertical directions are particularly effective to achieve this benefit. It will also be appreciated that the invention provides for determining the half-pel MV without the need to compute all of the halfpel values and store them in memory.

The invention is not limited to the embodiments described but may be varied in construction and detail.

Claims (11)

1. A method for motion estimation in video data processing, the method comprising fhe steps of:(a) finding an integer-pel, and (b) determining a half-pel motion vector to yield the smallest moving average estimate on a sub-pixel grid in each of the vertical and horizontal axes, in which the half-pel motion vector is determined by comparison of moving average estimates on the integer-pel.

2. A method as claimed in claim 1, wherein the determining step (b) comprises identifying a candidate set of half-pel motion vectors, and subsequently accepting or rejecting the candidate set.

3. A method as claimed in claim 2, wherein a candidate set is rejected if the difference between an MAE on the integer grid on the candidate set side and the integer pel MV MAE is less than a threshold value multiplied by the integer pel MV MAE.

4. A method as claimed in claim 2 or 3, wherein, for each axis, a first candidate set is identified; the first candidate set is accepted or rejected, and if rejected a second candidate set is identified; and said second candidate set is accepted or rejected.

5. A method as claimed in claim 4, wherein a default candidate set is identified if the second candidate set is rejected.

6. A method as claimed in claim 5, wherein the default candidate set is a set of half-pel motion vectors on an axis extending through a central integer pel.

7. A method as claimed in claims 3 to 6, wherein the threshold value is 5 approximately 0.25.

8. A method as claimed in claims 3 or 7, wherein threshold selection is made dynamically in response to past variation of MAEs.

9. 10 9. A method as claimed in any preceding claim, wherein a decision is taken before half-pel determination as to whether there should be a single motion vector or four motion vectors per block. 10. A video data processor comprising means for performing any motion 15 estimation method as claimed in any preceding claim.

10.

11. A computer program product comprising software code for performing a method as claimed in any of claims 1 to 6 when executing on a processor. ji B. 1/1 X: half-pel MV □ : integer-pel MV xl x2 ........7 z.......: r.......7k....... .......7 T, t F 1 KA. 1 V F h m3 1 ! J 7 V U Γ - π m4 Fig. 3 Fig. 4