Search Range Adjustment and Motion Vector Prediction for Fast

Motion Estimation

Using Neighbouring Motion Vectors and Distortions for Adjustment of a Search

Range and a Starting Point

Do-Kyung Lee and Je-Chang Jeong

Department of Electronics and Computer Engineering, Hanyang University, Seoul, Korea

Keywords: Fast Motion Estimation, Search Range Adjustment, Motion Vector Prediction, Partial Distortion Search,

Block Matching Algorithm.

Abstract: The block matching algorithm has been widely used for motion estimation, but it suffers from heavy

computational complexity. Many researchers in video processing area have proposed fast motion estimation

algorithms by adopting various ways to reduce its complexity. In this paper, we introduce a conventional

method controlling a search range and defining a new starting point, and also discuss our proposed method

which supplements previous work by using neighbouring block’s motion vectors and distortions.

Consequently, we obtained remarkable performance about 80 times faster than full search and 2.6 times

faster than conventional algorithm with small video quality improvement in terms of PSNR. Therefore, the

proposed method can be applied for real-time video processing applications.

1 INTRODUCTION

Motion Estimation (ME) and Motion Compensation

(MC) are important parts for video coding standards,

such as MPEG-1/2/3, H.261/263/264 and HEVC.

The purpose of ME and MC is to remove temporal

redundancy between successive frames in video

sequences. The most popular method of ME

algorithms is the Block Matching Algorithm (BMA)

which is utilized by many video coding standards.

BMA divides frames up into rectangular blocks and

match up with other blocks in reference frames.

BMAs are the most effective way because it has

very simple structure and also convenient to

implement. They find the optimal motion vector

from the minimum distortion value in terms of Sum

of Absolute Difference (SAD), sum of squared

difference (SSD), and Mean Squared Error (MSE).

The Full Search (FS) is the most straightforward

BMA; it calculates distortions and finds an optimal

motion vector between the current block and all

possible locations in pre-defined search window.

However, FS suffers from computational complexity

and it is heavy burden to implement real-time video

processing. Therefore, many fast motion estimation

(FME) algorithms to reduce complexity of FS have

been proposed for last few decades without peak-to-

noise ratio (PSNR) loss. For example, Three-Step

Search (TSS), Four-Step Search (FSS), Hexagonal

Search (HS), and Enhanced Hexagonal Search

(EHS) are taken the coarse search using geometric

search patterns that evaluates only few points in the

search window. Successive Elimination Algorithm

(SEA), Multilevel Successive Elimination Algorithm

(MSEA), Normalized Partial Distortion Search

(NPDS), Adjustable Partial Distortion Search

(APDS) and Search Range Adjustment and

Matching Point Decimation (SRAMPD) are also

proposed for FME.

In this paper, we briefly review about SRAMPD

in Section 2, SRAMPD predicts motion vector and

improves search speed by reducing search widow

size. Proposed methods that compensate defects of

SRAMPD are introduced in Section 3. Experimental

results compared with other algorithm are illustrated

in Section 4. Finally, we conclude our method with

some discussions in Section 5.

2 CONVENTIONAL METHOD

In this Section, we introduce a search range

130

Lee D. and Jeong J..

Search Range Adjustment and Motion Vector Prediction for Fast Motion Estimation - Using Neighbouring Motion Vectors and Distortions for Adjustment

of a Search Range and a Starting Point.

DOI: 10.5220/0004066001300133

In Proceedings of the International Conference on Signal Processing and Multimedia Applications and Wireless Information Networks and Systems

(SIGMAP-2012), pages 130-133

ISBN: 978-989-8565-25-9

 2012 SCITEPRESS (Science and Technology Publications, Lda.)

adjustment method in SRAMPD since it is a basic

idea of proposed algorithm. For reducing the search

range without noticeable video quality loss,

SRAMPD proposed a starting point decision and an

adaptive search range adjustment using a spatial

correlation between adjacently located motion

vectors. Since we cannot obtain the motion vector of

the current block until ME of the current block is

accomplished, SRAMPD firstly predict the motion

vector (mv

) by applying median function to

neighbouring motion vectors (left, upper and upper

left) as follows,

(, , ).

pluul

median=mv mv mv mv

(1)

For an adaptive search range, SRAMPD calculates

SAD(mv

) of new starting point which is located at

and also evaluates SADs of its neighbouring

blocks which are located left, upper and upper left

position relative to new starting point is collected in

C = [SAD(mv

) SAD

SAD

], respectively.

Finally, a search range is determined by,

max(

max 0, min , .

SAD

D R round

⎛⎞

⋅

⎜⎟

⎝⎠

(2)

The function round( ) gives the nearest integer value

of input data, R is a maximum value of given search

range and T

SAD

is a threshold value which is set to 8

to control search range. Also, R·max(C) is

normalized by pixel numbers in the block.

3 PROPOSED ALGORITHM

3.1 Motion Vector Prediction

SRAMPD use the median function for prediction but

it cannot get accurate motion vector because

neighbouring motion vectors are not always reliable

and there are many outliers in the motion vector

field. These error vectors can be propagated spatially

and it leads video quality degradation, as well. If we

find the correct starting point ideally, ME is not

necessary because the predicted motion vector will

be a final motion vector. For this reason, we

proposed a more accurate motion vector prediction

method using distortion values corresponded to

neighbouring motion vectors.

First, collect SADs include zero vector and

neighbouring MVs like, S = [SAD(mv

) SAD(mv

SAD(mv

), SAD(mv

)] and evaluate a motion

vector of the minimum SAD of S as:

arg min( ),

pnew

mv S

(3)

means the zero vector and mv

is the motion

vector of the block located upper right position

relative to current block. mv

is the co-located

motion vector in the reference frame and it can be

substituted for mv

when the previous motion

vector field is not exist.

Additionally, we consider the consistency of the

reference motion vectors. If two motion vectors have

same magnitude and direction, proposed method

considers the other motion vector as unreliable

motion vector. This motion vector is skipped when

proposed algorithm is calculating the predicted

motion vector.

D_mv

P_mv

F_mv

Figure 1: The description of P_mv, D_mv and F_mv.

Figure 1 shows the relationship of motion

vectors resulted by motion vector prediction and

motion estimation. P_mv is the predicted motion

vector, F_mv is the final motion vector and we

define the difference motion vector that is D_mv

which is the result of ME for measuring the accuracy

of prediction methods.

Table 1: The Comparison about the average length (pixels)

of D_mv.

Sequence SRAMPD Proposed

Akiyo 0.03 0.01

Coastguard 0.29 0.04

Container 0.16 0.03

Flower 0.61 0.13

Foreman 1.25 0.43

Mobile 0.39 0.05

Stefan 2.08 0.30

Mother 0.98 0.28

Table 0.80 0.18

Average 0.73 0.16

Table 1 presents the accuracy of motion vector

prediction method in SRAMPD and proposed

algorithm. The prediction accuracy is significantly

increased since the average length of D_mv is

decreased about 4.6 times. It means the predicted

motion vector is more similar to the final motion

vector. Therefore, the search range can be reduced

more than conventional method for speed-up

improvement.

SearchRangeAdjustmentandMotionVectorPredictionforFastMotionEstimation-UsingNeighbouringMotionVectors

andDistortionsforAdjustmentofaSearchRangeandaStartingPoint

131

Table 2: The results of reference methods and proposed algorithm in terms of time per frame and PSNR.

FS NPDS APDS SRAMPD Proposed

Sequences Time PSNR Time PSNR Time PSNR Time PSNR Time PSNR

Akiyo 0.401 42.34 0.039 42.23 0.019 42.32 0.012 42.32 0.0030 42.32

Coastguard 0.413 30.43 0.045 30.32 0.029 30.44 0.013 30.42 0.0053 30.43

Container 0.438 38.17 0.044 38.15 0.026 38.16 0.012 38.03 0.0038 37.99

Flower 0.452 25.89 0.046 25.82 0.033 25.90 0.014 25.85 0.0076 25.87

Foreman 0.407 31.81 0.044 31.69 0.036 31.82 0.016 31.88 0.0062 31.84

Mobile 0.467 25.04 0.046 24.84 0.028 25.02 0.018 24.97 0.0071 25.01

Stefan 0.419 23.90 0.048 23.73 0.049 23.89 0.017 24.71 0.0064 25.00

Mother &daughter 0.415 40.05 0.044 39.94 0.036 40.00 0.012 39.75 0.0047 39.87

Table 0.463 31.46 0.045 31.23 0.029 31.42 0.012 31.47 0.0042 31.27

average 0.430 32.12 0.045 31.99 0.032 32.11 0.014 32.16 0.0054 32.18

Speed-up ratio 1.000 9.659 13.590 30.764 80.190

PSNR difference 0.000 -0.127 -0.014 0.035 0.057

3.2 Search Range Adjustment

Upper

Block

Left

Block

Starting

Point

Right

Block

Bottom

Block

Figure 2: The location of reference blocks.

In SRAMPD, the author refers to left, upper and

upper-left block of initial point for the search range

adjustment. However, in reference frame, there also

exist right and bottom block, and it is more

correlated with initial point than upper-left position

since, in video sequence, the dominant edges are

horizontal and vertical direction. Therefore, in

proposed algorithm, we refer to left, upper, right and

bottom block’s distortion value for the search range

control. The modified collection of SADs follows,

'[ ( ) ].

lurb

C SAD SAD SAD SAD SAD= mv

(4)

The equation (2) is used in proposed method to

obtain a search range; however, we changed the

threshold value experimentally that is much larger

than SRAMPD about 4 times and we utilized

average of C’ instead of maximum SAD because the

proposed method of motion vector prediction is

more accurate than conventional algorithm. Also,

4 EXPERIMENTAL RESULTS

To evaluate performance of the proposed algorithms,

we compare them with FS, NPDS, APDS and

SRAMPD in terms of time per frame and PSNR. We

use speed measurement as time per frame because

each sequences have the different number of frames.

The experimental setup is as follows: the block size

of 16 x 16, the search window size of ±16. Eight

CIF (352 x 288 pixels) video sequences, “Akiyo”,

“Coastguard”, “Container”, “Flower”, “Mobile”,

“Stefan”, “Mother & daughter” and “Table”, are

used.

Table 2 shows the performance of the proposed

algorithms. The proposed algorithms are about 2.6

times faster than SRAMPD, averagely. Especially,

in “Flower” sequence, 5.31 times faster than

conventional method. The proposed method obtains

good performance for almost every sequence even if

the sequence has the fast motion characteristic. In

“Table” sequence, we cannot get a better result in

terms of PSNR because there is a scene change at

131th frame. When the scene is completely shifts,

the common characteristics include the motion

vector filed is not ordinary. Proposed method use the

co-located motion vector in reference frame, it can

be a weak-point if there is breakpoint caused by

scene change or fast and unpredictable motions.

From Table 3, we can observe the search

window size that is almost 2.5 times smaller than

SRAMPD; therefore, our method for search range

adjustment yields speed-up improvement in terms of

time. Especially, for “Akiyo” and “Container”,

proposed algorithm achieves remarkable results

SIGMAP2012-InternationalConferenceonSignalProcessingandMultimediaApplications

132

since they have slow motion characteristic. Even the

object in the sequence moves fast, we still obtain the

gain about both PSNR and speed up improvement.

Table 3: The Comparison about the average search

window size.

Sequence SRAMPD Proposed

Akiyo 11.51 2.81

Coastguard 12.55 5.35

Container 11.22 2.98

Flower 12.80 7.51

Foreman 13.36 5.73

Mobile 15.24 7.58

Stefan 13.83 5.91

Mother 10.98 3.52

Table 11.18 3.98

Average 12.52 5.04

5 CONCLUSIONS

In this paper, we proposed a new search range

adjustment and a motion vector prediction method

by adopting adjacently located block’s motion

vector and SADs. Using the proposed motion vector

prediction, we find the almost correct initial point.

Therefore, we additionally reduced the search range

and also computational complexity. Comparing with

conventional algorithm, experimental results

confirmed that our method achieves about 2.6 times

faster speed with small PSNR improvement.

ACKNOWLEDGEMENTS

This work was supported by the Industrial Strategic

technology development program(10039886,

Combined Modulator and Demodulator with Real

Time Multi Codec Transcoder based on Hardware

Processing) funded by the Ministry of Knowledge

Economy(MKE, Korea).

REFERENCES

T. Koga, K. Iinuma, A. hirano, Y. Iijima, and T. Ishiguro,

1981. Motion compensated interframe coding for

video conferencing, In Proc. Nat. Telecommun. Conf.,

New Orleans, L.A.

R. Li, B. Zeng, and M. L. Liou, 1994. A new three-step

search algorithm for block motion estimation, IEEE

Trans. Circuits Syst. Video Technol.

Li W., Salari E, 1995. Successive elimination algorithm

for motion estimation, IEEE Trans. Image Process.

Gao X. O., Duanmu C. J., Zou C. R., 2000. A multilevel

successive elimination algorithm for block motion

estimation, IEEE Trans. Image Process.

Ahn T. G., Y. H. M., Kim J. H., 2004. Fast full-search

motion estimation based on multilevel successive

elimination algorithm, IEEE Trans. CSVT: Trans.

Letter.

S. Zhu and K. K. Ma, 2000. A new diamond search

algorithm for fast block matching motion estimation.

IEEE Transactions on Image Processing.

C. Zhu, X. Lin, and L. P. Chau, 2000. Hexagon-based

search pattern for fast block motion estimation. IEEE

Trans. Circuits Syst. Video Technol.

C. Zhu, X. Lin, and L. P. Chau, 2004. Enhanced

Hexagonal search for fast block motion estimation.

IEEE Trans. Circuits Syst. Video Technol.

C. K. Cheung and L. M. Po, 2000. Normalized partial

distortion search algorithm for block motion

estimation, IEEE Trans. Circuits Syst. Video Technol.

Cheung C. H., Po L. M, 2003. Adjustable partial distortion

search algorithm for fast block motion estimation,

IEEE Trans. Circuits Syst. Video Technol.

Jing X., Chau L. P, 2007. Partial distortion search

algorithm using predictive search area for fast full-

search motion estimation, IEEE Signal Process. Letter.

S. Lee, 2010. Fast motion estimation based on search

range adjustment and matching point decimation, IET

Image process.

SearchRangeAdjustmentandMotionVectorPredictionforFastMotionEstimation-UsingNeighbouringMotionVectors

andDistortionsforAdjustmentofaSearchRangeandaStartingPoint

133