A FAST SEARCH ALGORITHM FOR SUB-PIXEL MOTION

ESTIMATION IN H.264

Dong-kyun Park, Hyo-moon Cho

School of Electrical, University of Ulsan,7-524 san29 Muger-dong Nam-gu ,Ulsan, South Korea

Jong-Hwa Lee

School of Electrical, University of Ulsan,7-524 san29 Muger-dong Nam-gu ,Ulsan, South Korea

Keywords: SAD, H.264, Motion Estimation, Sub-pixel.

Abstract: We propose the advanced sub-pixel block matching algorithm to reduce the computational complexity by

using a statistical characteristic of SAD (Sum of Absolute Difference). Generally, the probability of the

minimum SAD values is highest when a searching point has one pixel distance from the reference point.

Thus, we can overcome high computational complexity problem by reducing the searching area. The main

concept of proposed algorithm is one of the fast searching algorithms based on TSS (Three Step Search)

method. First, we find three minimal SAD points of all nine searching points on integer distance and then

we obtain two minimal SAD points on 1/2-pixel between the second minimal SAD point and the first, and

between the third and first, respectively. Finally, we can find matching point by comparing the SAD values

among six points which are on triangle from the first minimal SAD point and two 1/2-pixel points including

1/4-pixels. The proposed algorithm in this paper needs only 14 searching points in sub-pixel mode, whereas

the conventional TSS method needs total 25 searching points. Therefore, this algorithm improves the

processing speed as 51%.

1 INTRODUCTION

The image compression technologies have been

researched on the removing the temporal and spatial

redundancies. The H.264 video standard has higher

computational complexity for traditional video

compression methods, though it gives an efficient

compression ratio. Thus, the fast and efficient

methods of H.264 compression processing are

requested for the real-time processing. Specially, the

researches for the motion estimation (ME) and

motion compensation (MC) in H.264 are carried out

actively and widely since those took more than

about 70% from the whole H.264 operation [4].

Generally, motion estimation in integer unit cannot

obtain accurate motion estimation because the

motion of object on video image does not move in

integer unit, always. Thus, sub-pixel units such as

1/2-pixel and 1/4-pixel are used in order to accurate

motion estimation in advanced video compression

standard. For higher accuracy, H.264 standard uses

1/4-pixel ME/MC, whereas MPEG-4 part2 uses only

1/2-pixel. As this reason, the computational

complexity of H.264 is remarkably increased by

contrast with H.263 or MPEG-4 part 2.

In this paper, we propose the efficient and fast

sub-pixel motion-search algorithm reducing the

search area and a computational complexity. To

evaluate our proposed algorithm, we compared with

the test result of JM11.0.

2 MOTION ESTIMATION

ALGORITHMS

2.1 Error Value

and SAD Characteristic

The SAD is widely used for block-match since it is

simple and easy computation method to evaluate the

motion estimate without multiplication and division.

Its equation is shown in eq. (1).

135

Park D., Cho H. and Lee J. (2007).

A FAST SEARCH ALGORITHM FOR SUB-PIXEL MOTION ESTIMATION IN H.264.

In Proceedings of the Second International Conference on Signal Processing and Multimedia Applications, pages 135-138

DOI: 10.5220/0002139801350138

 SciTePress

∑∑

−

−=

ijij

rcSAD

(1)

where, N x N is the block size, and c

and r

are the

sample image signal at position (i, j)of the current

frame image and the reference frame image,

respectively.

))()(1,1(

))(1)(1,(

)1)()(,1(

)1)(1)(,(),(

yyxxyxI

yyxxyxIyxI

ssss

−−+++

−+−++

−+−

(2)

I(x, y) represents the brightness at integer-pixel (x, y)

or sub-pixel (x

, y

Fig. 1 shows the SAD-error distribution in the

range of

1± from the reference point. The SAD

error value is increased as far off from the reference

as shown in Fig. 1. The fact that The SAD error-

surface graph as shown Fig. 1 has high probability at

quarter-pixel location is known from Table 1 which

is obtained by experiments. Therefore, the 8 scores

within ±1-pixel locationis have higher value and we

could assume it include sub-pixel motion vector.

Figure 1: A typical SAD error surface.

Table 1 shows the property of SAD error-surface

for several sequences. The findings of experiment

indicated that more than 96% had a parabolic

distribution of adjacent SAD values around the

smallest SAD values. Therefore, pixel point which

has the smallest SAD in integer pixel can be found

using the parabolic distribution between neighboring

pixels of SAD error surface. Then, the pixel point

with the smallest SAD in integer pixel can be found

using the distribution of SAD values after computing

the SAD of values around pixels within the range of

1 based on that position. On the basis of that position,

SAD of neighboring pixels within the range of 1 can

be computed. Using the distribution property of

computed SAD values, pixel position which has the

smallest SAD in half pixel and quarter pixel can be

forecasted.

Table 1: Probability of finding the optimal quarter-pixel

location.

Sequence # of blocks # of single valley blocks Ratio

Claire 48807 47140 96.58%

News 29601 28908 97.66%

Salesman 44352 44106 99.45%

Trevor 14751 14199 96.26%

2.2 Ordinary Motion Estimation

Ordinary motion estimation used FS (Full Search).

FS is the algorithm to search all (2R+1)

number of

candidates. FS method has problems in applying to

real time video clip compression system because

heavy computation load is required, though it

enables the most optimal motion vector by searching

for all positions within the search range. Therefore,

a lot of high speed algorithms have been proposed to

resolve those problems.

2.3 Three Step Search

This algorithm is most widely known in its three

step form, the ‘three step search (TSS)’, but it can be

carried out with other numbers of steps. For a search

window of ±(2

-1) pixels, the TSS algorithm is as

follows;

1. Search location (0,0)

2. Set S =

)12( −

3. Search eight locations

S−+ /

pixels around

(0,0)

4. From the nine locations searched so far, pick

the location with the smallest SAD and make

this the new search origin.

5. Set S = S/2

6. Repeat stages 3-5 until S=1.

Figure 2: Three Step Search (TSS).

SIGMAP 2007 - International Conference on Signal Processing and Multimedia Applications

136

Figure 2 illustrates the procedure for a search

window of

± 1.75(i.e. N=3). The first ‘step’

involves searching location (0, 0) and eight locations

1±

pixels around the origin. The second ‘step’

searches

5.0±

pixel around the best match from the

first step(highlighted in bold) and the third step

searches

25.0±

pixels around the best match from

the second step(again highlighted). The best match

from this third step in chosen as the result of the

search algorithm. With a search window of

1.75,

three repetitions (steps) are required to find the best

match. A total of (9+8+8) = 25, search comparisons

are required for the TSS, compared with (15

15) =

225 comparisons for the equivalent full search. In

general, (8N+1) comparisons are required for a

search area of

)12( −±

pixels.

3 PROPOSED ALGORITHM

In this paper, the second smallest SAD and the third

smallest SAD can be found in integer pixel 8 points

around the point which possess the smallest SAD in

integer pixel to identify the sub pixel motion vector

in that direction, using distribution property of SAD

error surface, as shown in Fig. 1. Therefore, 1/2-

pixel and 1/4-pixel motion vector can be obtained

through integer pixel SAD. In this regard, 1/2-pixel

and 1/4-pixel motion vector can be obtained through

the position of integer pixel SAD value.

As shown in Equation (3), it is found that both

side points next to pixel have the biggest effect

because of the characteristics of H.264, which

interpolates 1/2-pixel with 6-tap FIR filter.

Therefore, 1/2-pixel and 1/4-pixel motion estimation

at a position between the second smallest SAD and

the third smallest SAD where SAD was the smallest-

was proposed, using the following property.

Figure 3: Proposed motion estimation in half pixel.

If SAD is minimum at (x,y), 1/2-pixel motion

vector can be obtained in the following way:

Step 1: Find (x,y) with the smallest SAD in integer

pixel, and find the point which has the second

smallest SAD and the third smallest SAD in the

neighboring integer pixel point within

Step 2: Find 1/2-pixel SAD point between the

second smallest SAD point and the third smallest

SAD point.

Considering that the 1/2-pixel searching point has

the motion estimation in the direction from the

minimum SAD to the second largest and the third

largest SAD, two searching points are enough.

In addition, 1/4-pixel searching point can be

found using 1/2-pixel searching point, as shown in

Fig. 4.

Figure 4: Proposed motion estimation in quarter pixel.

Find 1/4-pixel SAD value can be found by using

two points and minimum integer SAD found in 1/2-

pixel as shown in Fig 4. Equation (3) is an example

in which 1/4-pixel motion estimation in Fig 4 was

used.

2/)(3

2/)(2

2/)(1

+∂=

∂

(3)

Therefore, the strength of the proposed method is

that the computation was made simpler and the

processing time was shortened because only 2

searching points were sought in 1/2-pixel and 3

searching points in 1/4-pixel, unlike the previous

way which required 8 searching points for 1/2-pixel

and 1/4-pixel.

A FAST SEARCH ALGORITHM FOR SUB-PIXEL MOTION ESTIMATION IN H.264

137

4 IMPLEMENTATION BASED

ON THE PROPOSED

ALGORITHM

The test was conducted using JM11.0 open source.

Fast search algorithm embodied in JM11.0 were

used exactly for integer pixel, and a comparison was

made between the method which reduces the number

of searching points for half and quarter pixel and the

method which searches for in whole under existing

method, and the result was summarized. The

proposed algorithm is considered to embody the

function nearly on a par with PSNR, compared with

previous algorithm, to evaluate the image quality.

A comparison was made between the method

which searches 5 points under existing method and

the proposed method, in order to validate the

efficiency of the proposed method for sub pixel

motion estimation.

Table2: Performance comparison about propose algorithm.

5 CONCLUSIONS

This paper proposed the method of motion

estimation in less complicated sub-pixel to ensure

the efficient motion estimation process which

comprises the largest portion of calculation in

H.264, the next generation image compression codec

with superb compression efficiency. The proposed

method is to forecast the direction which has less

error range in sub pixel to search for motion vector,

using the error surface property and pixel

interpolation property of SAD. Searching only in the

direction forecast on the basis of the difference in

SAD value, the searching point was remarkably

reduced to two for half pixel and five for quarter

pixel respectively, unlike existing method which

required 8 searching points.

The result of test based on the proposed method

indicated that the time that it took to calculate was

reduced by approximately 24% for encoder

processing and by approximately 51% for sub pixel

motion estimation, compared to the function of

existing JM 11.0. Therefore, the efficiency of

proposed method for motion estimation was

validated.

ACKNOWLEDGEMENTS

The work was supported by the Korea Research

Foundation Grant funded by the Korean

Government and 2 level BK21 (EVERDEC (e-

Vehicle Research & human Resource Development

Center)) in MOE (Ministry of Education & Human

Resources Development). (KRF-2006-D00292)

REFERENCES

Cheng Du, Yun He, and Junli Zheng, “PPHS: A Parabolic

Prediction-Based, Fast Half-Pixel Search Algorithm

for Very Low Bit-Rate Moving-Picture Coding”, IEEE

Transactions on Circuits and Systems for Video

Technology, Vol. 13, No.6, June 2003.

Bo Zhou, Jian Chen, “A Fast Two-Step Search Algorithm

for Half-Pixel Motion Estimation”, ICECS 2003,

Proceedings of the 2003 10th IEEE International

Conference, Vol 2, pp 14-17, Dec 2003.

H. M. Wong, O. C. Au, J. Huang, S. Zhang and W. Z.

Yan, “Sub-Optimal Quarter- Pixel Inter-Prediction

Algorithm (SQIA)," ICASSP '05. IEEE International

Conference on Volume 2, pp. 921-924, March 18-23,

2005.

Iain E, G. Richardson, “Video Codec Design” pp102~ 10

5, willey

Richardson, “H.264 and MPEG-4 Video compression”,

Willey

W Zhenyu, J Baochen, Z Xudong, C Yu, “A New Full-

pixel and Sub-pixel Motion Vector Search Algorithm

for Fast Block-matching Motion Estimation”, Image

and Graphics, 2004. Proceedings. Third International

Conference , ICIG’04, pp18~20, Dec 2004 .

Algorithm

PSNR

(dB)

Bit-rate

(Kbit/s)

Encoder

processing

time (sec)

Sub-pixel

processing

time (sec)

JM 11.0 38.7 85.2 48.2 7.68

foreman

Proposed 37.6 80.6 36.3 3.70

JM 11.0 40.5 44.4 44.2 7.91

akiyo

Proposed 39.2 37.6 33.3 3.86

JM 11.0 40.2 87.8 46.9 8.12

news

Proposed 39.8 79.2 36.5 3.96

JM 11.0 39.2 87.8 47.8 8.49

table

tennis

Proposed 39.0 77.2 36.2 4.99

JM 11.0 38.2 105.1 49.9 8.52

Mobile

Proposed 37.8 91.2 38.0 4.23

SIGMAP 2007 - International Conference on Signal Processing and Multimedia Applications

138