Random Initial Search Points Prediction for Content Aware Motion

Estimation in H.264

Vidya N. More, Ajinkya Deshmukh, Dhiraj More and M. S. Sutaone

College of Engineering Pune, Wellesly Road, Shivaji Nagar, Pune, India

Keywords:

H.264, Content Aware Motion Estimation, Search Point Prediction.

Abstract:

Motion estimation algorithms used in video encoders are based on three important issues: selection of good

initial search points, choice of appropriate search pattern and effective early termination criteria at different

stages in algorithm.Motion vector prediction is also treated as initial search point prediction, in which pos-

sibility of good match block is predicted. Prediction is based on prior data from co-located and/or adjacent

macroblocks from reference frame or current video frame respectively.Different search patterns contribute in

achieving near accurate motion estimation. Different types of motion in real time videos can be tracked using

different types of patterns. Early termination criteria at different stages in algorithm, avoid search at further

possible locations which are pre-decided by pattern of search. This in turns reduces computations and motion

estimation time.Proposed algorithm is combination of two concepts, content awareness and initial point pre-

diction.Contents of video data is in terms of homogeneity coefﬁcients. Initial search point prediction is used to

avoid the search trapping into local minima.The algorithm is implemented on Reference Software of JM18.4

of H.264/AVC revised on 5th May 2011. The results of the implemented algorithm show that the total time

taken for encoding and motion estimation time are less as compared with other algorithms for the videos of

different resolutions.

1 INTRODUCTION

Video technology has been an inevitable and ubiqui-

tous part of our daily lives. Right from mobile till

the satellite surveillance systems, wide range of ap-

plications of the video technology have been imple-

mented. The need for factors like higher quality, low

bit rate and lesser disk space has triggered the devel-

opment in this ﬁeld. Even the hand-held devices now

are able to play the high deﬁnition videos due to ef-

ﬁcient encoding and decoding algorithms on the real

time platforms.

Presently, H.264/MPEG-4 v10 is used widely as

a standard video codec for all applications. H.264

applications mainly include satellite HDTV , high

capacity storage devices such as Blu-ray discs, in-

ternet protocol television, video over internet, etc.

The video coding standards are mainly divided into

two main classes, namely, MPEG-x and H.26x.

The MPEG-x codecs are developed ISO/IEC JTC1,

whereas the H.26x are developed by ITU-T. However

their joint work has resulted into standards such as

H.262/MPEG-2 and H.264/MPEG-4 part 10. Cur-

rent video standard H.265 is also published in 2013

which aims at higher throughput in all respects of

video compression but at the cost of complicated al-

gorithms and more hardware demand.It is observed

that the motion estimation time takes more than 60%

out of total time (refer to Table 1)thus there is need of

optimum algorithms.Distortion measurement criteria,

Sum of Absolute Difference (SAD)is mostly used in

all versions of JM code. Table 1 show results of Total

Time , ME Time , PSNR of Y and Bit Rate for a video

sequence Foreman.yuv of qcif resolution on Simpli-

ﬁed Unsymmetrical Hexagonal Search algorithm.

Motion estimation (ME)algorithms are quantiﬁed

on the parameters of Total encoding time, ME time

, Peak signal to noise ratio (PSNR) and Bitrate in-

crease/decrease. These parameters are mainly out-

come of three strategies of ME algorithm i.e. initial

search point prediction, different search patterns of

searching and early termination. Ample literature is

available on search pattern selection and correspond-

ing early termination threshold calculations.In this

proposed algorithm we are focusing on initial point

search on random basis and homogeneity analysis of

video contents.

Heuristic search algorithm of ME is full search.

159

More V., Deshmukh A., More D. and Sutaone M..

Random Initial Search Points Prediction for Content Aware Motion Estimation in H.264.

DOI: 10.5220/0005311401590165

In Proceedings of the 10th International Conference on Computer Vision Theory and Applications (VISAPP-2015), pages 159-165

ISBN: 978-989-758-089-5

 2015 SCITEPRESS (Science and Technology Publications, Lda.)

Table 1: More than 50% Time Taken by ME Compared with

Full Search.

ME Algorithm

Full Search UMHEX UMHEXS

Total Time 1142.068 318.435 293.341

ME Time 995.574 169.757 146.809

PSNR 35.805 35.76 35.78

Bitrate 1115.5 1168.99 1123.43

It searches the macroblock for each possibility of

match in reference video frame and thus it is the

optimum algorithm.Different block based ME algo-

rithms are developed to minimize the search points

in search range so as to reduce the ME Time. The

algorithms like Three step search (TSS), New three

step search (NTSS) (Renxiang Li and Liou, 1994),

Four step search (FSS)(Po and Ma, 1996), Dia-

mond search (DS) (Zhu and Ma, 1997),Hexagonal

search (Zhu et al., 2002), Unsymmetrical Hexago-

nal search (UMHEX)(Zhibo Chen, 2002) ,Simpliﬁed

UMHEX(UMHEXS) (Toivonen and Heikkil, 2006),

Adaptive Rood Pattern Search(Nie and Ma, 2002)

Predicted Motion Vector Field Adaptive Search Tech-

nique(PMVFAST) (Tourapis et al., 2002) and En-

hanced Predictive Zonal Search (EPZS)are devel-

oped (Xu and He, 2008). Among these algorithms

TSS,NTSS,FSS were part of MPEG 1,Mpeg 2 and

UMHEX, PMVFAST and EPZS are part of current

video standards(Sarwer and Wu, 2009). Thus it will

be more appropriate to compare the results with these

algorithms.All these mentioned algorithms are mainly

focused on different shapes of search patterns and

criteria to select these patterns(K. Venkatachalapathy

and Viswanath, 2004). Figure 1 show some of the

search patterns taken by UMHEXS algorithm at dif-

ferent stages in the algorithm.

Researchers are developing the algorithms for

content aware ME to make the search more adaptive

(Yi-Ching Liawa and Zuu-ChangHong, 2009). Char-

acteristics of video data can be in terms of object

acceleration, inertia, linearity (DongYoon Kim and

Park, 2013) of the object etc. This paper is focus-

ing on Homogeneity of the object in a macroblock

with respect to adjacent macroblocks of the video

frame. The algorithm developed by Humaria Nisar

is implemented for the proposed work as base work.

In proposed algorithm initial search points are gen-

erated randomly which are used in tracking the non-

homogeneous/irregular motion of video data. Thus

this is an incremental work in the base algorithm .

This paper has ﬁve sections and from which sec-

tion 2 details about basic algorithm of homogene-

ity.Proposed work is mentioned in section 3 along

with Simulation and test conditions is discussed in

section 4. Results and conclusions are elaborated in

Figure 1: Example of search patterns in UMHEXS algo-

rithm.

section 5 and the references are listed used for this

work.

2 HOMOGENEITY ALGORITHM

Every video sequence contains varying amount of ho-

mogeneity.i.e.various macroblocks of frames of these

sequences belong to one group or another. Thus

making it easy for us to predict the motion vec-

tor with the help of previously coded homogeneous

blocks(Humaira Nisar and Choi, 2012). This al-

gorithm classiﬁes every macroblock into three cate-

gories :

• Homogeneous

• Non homogeneous

• Stationary

and depending upon these categories the search pat-

tern is applied adaptively.

2.1 Motion Vector Prediction

Motion vector prediction is carried out using spatio

temporal neighbours.The spatial neighboring blocks

in current frame are left, right, corner left, corner right

macroblocks, whereas the temporal domain provides

the collocated macroblock which is the macroblock

with the same location in the reference frame. These

neighbors provide three kinds of initial motion vec-

tors,

VISAPP2015-InternationalConferenceonComputerVisionTheoryandApplications

160

1. Zero motion vector(ZMV=[0,0])

2. PMV1=median(motion vectors of spatial neigh-

bors)

3. PMV2=motion vector of collocated macroblock

2.2 Homogenity Analysis

In video sequences, there exists high correlation be-

tween the neighboring blocks in the spatial and tem-

poral domains. If the current and neighboring blocks

belong to same object then they have consistent mo-

tion activity and hence these can be classiﬁed as ho-

mogeneous blocks. If the motion vectors are not con-

sistent then these blocks are considered as non homo-

geneous blocks. If the blocks are homogeneous then

a simple median prediction is carried out. The homo-

geneity coefﬁcients (HC) play important role in the

analysis.

Calculate the average of motion vectors of neighbor-

ing blocks in the current frame.





∑

i=1

(1)





∑

i=1

(2)

Here N is number of neighboring blocks of cur-

rent block (In this case N = 5 i.e. A, B, C, D blocks

from same frame and co-located block from reference

frame).

The homogeneity coefﬁcients are



∑

i=1

− MV

|MV

(3)



∑

i=1

− MV

|MV

(4)

HC = HC

+ HC

(5)

However, for MV

=0 or MV

=0, there are two possible

cases.

1. The x or y components of motion vectors of

neighboring blocks lie in opposite directions to

each other.

2. The neighboring blocks are stationary i.e. all

neighboring blocks have coefﬁcients (0,0).

In such cases HC are calculated using the mean of

absolute values of x and y components of the MVs.

If MV

=0 and MV

=0, then that block is con-

sidered as a Stationary block.

Smaller the value of HC means the block is homoge-

neous block and motion is consistent. Large value of

HC denotes that the block is non-homogeneous.

2.3 Predicted Motion Vector

The magnitude of predicted motion vector(PMV) pro-

vides the basis for motion classiﬁcation and hence de-

ﬁnes the motion content of the video. If the magnitude

of the PMV is greater than 1/2 of the search range,

motion of block is considered fast else it is medium

or slow. The PMV is calculated using following equa-

tion,

PMV = median(MV

, MV

) (6)

2.4 GMP Calculations

The SAD of the global minimum is generally small as

compared to the neighboring values. Hence the error

descent rate is quite sharp. The GMP can be calcu-

lated as,

GMP =

SAD

neighbourhood

SAD

centre

(7)

where SAD

neighbourhood

is the SAD of the point next

to the center point and SAD

centre

is the SAD of the

center point. The smaller the value of GMP, closer

we are to the global minimum. If the value of GMP

is larger then the search is far away from global min-

imum and we have to incorporate larger search pat-

tern and multiple predictors to identify the direction

of global minimum.

2.5 Early Termination

The early termination threshold is different for homo-

geneous and non-homogeneous blocks.For homoge-

neous blocks,

= mean(SAD

;SADl

;SAD

) (8)

For non-homogeneous blocks,

= min(SAD

;SADl

;SAD

) (9)

3 PROPOSED ALGORITHM

The algorithm described above distinguishes between

slow and fast motion of video sequence. But the per-

formance of the algorithm can be improved in case of

fast and irregular motion video sequences. The ran-

domized prediction can be performed in order to cater

this need. If the motion is fast and irregular then there

is a higher possibility that the neighboring blocks may

not be able to provide accurate motion vector predic-

tors hampering the performance in terms of motion

estimation time and PSNR. Thus randomized motion

prediction can be performed in the speciﬁed search

range and the performance can be lifted. The perfor-

mance of any algorithm is measured in terms of,

RandomInitialSearchPointsPredictionforContentAwareMotionEstimationinH.264

161

• Total encoding

• ME time

• PSNR of the luma component

• Bit rate

All the above mentioned parameters are interdepen-

dent. However, if you try to reduce the encoding and

estimation time, the PSNR decreases resulting in in-

creased bitrate. On the other hand,if we try to increase

the PSNR and decrease the bit rate, complexity in-

creases causing increase in encoding and estimation

time. Thus we can not use any one approach blindly

for all the applications. The approach is chosen de-

pending upon the application need.

Parameter Value

GOP Structure IPPPPPP

Proﬁle IDC Baseline(66)

Quantization Parameter I=28,P=28

Search Range 32

Entropy Coding Method CAVLC

Block distortion measure SAD

3.1 Algorithm

As mentioned earlier, fast and irregular motion can

be compensated with the help of random motion pre-

diction.The steps followed in this approach are given

below,

1. Calculate PMV of the macroblock (refer section

2.3).

2. If macroblock is either stationary or slow set pre-

diction range 16X16.For fast macroblock set it to

32X32.

3. Divide assigned prediction range in 4 parts.

4. From each part , one coordinate is randomly gen-

erated as Alpha from upper left,Beta from upper

right, Gamma from lower right and Delta from

lower left.

5. Center of prediction range is called as Epsilon.

6. Calculate SAD at each of 5 points.

7. Compare this SAD and ﬁnd point with minimum

SAD.

8. If this minimum SAD is lower than SAD of

Spatio-temporal neighbour then assign this ran-

domly generated point as Search center.If SAD

is not minimum ﬁnd minimum SAD point among

neighbour.

9. Apply pre-deﬁned search pattern on selected point

for further search.

4 IMPLEMENTATION AND TEST

CONDITIONS

Working environment The proposed algorithm is im-

plemented in JM reference software version 18.4 of

H.264 standard. The implementation is carried out

in Linux environment and the processor used is Intel

core i7 processor.

Following test conditions were used by making

changes in conﬁguration ﬁle.

The test sequences used are taken from different data

bases. The links are listed below,

http://trace.eas.asu.edu/yuv/

ftp://ftp.tnt.uni-hannover.de

ftp://ftp.ldv.e-technik.tu-muenchen.de

Test sequences are of different resolutions as shown

below.

• Bus(QCIF)

• Clair(QCIF)

• Coastguard(QCIF)

• Container(QCIF)

• Carphone(QCIF)

• Mobile(QCIF)

• Bus(CIF)

• Mobile(CIF)

• Shields(SD)

• Ice (HD)

Since most of the videos have motion changes or

scene changes after 150th frame, all these video se-

quences are executed for 300 frames. The algorithm

is meant for generating random data points in four

quadrants, it is iterated for 10 times for the same test

conditions and the average values of total time, ME

time, PSNR of Y and Bit Rate are presented in tabu-

lar form.

The performance of proposed algorithms is compared

with three algorithms already present in the JM 18.4

reference software, viz.,

1. Fast full search (FFSearch)

2. UMHEX

3. UMHEXS

5 RESULTS AND CONCLUSION

Table 2 is representing the actual values of the test pa-

rameters where as table 3 represents normalized val-

ues of parameters.It is observed from the results that

for all tested video sequences total encoding time and

VISAPP2015-InternationalConferenceonComputerVisionTheoryandApplications

162

Table 2: Results for Different Videos.

Video Algorithm Name Total Time (Sec) ME Time (Sec) PSNR (dB) Bit Rate

bus qcif

FFSearch 107.11 71.552 34.455 212.24

Umhex 76.168 39.304 34.409 225.99

Umhexsmp 68.477 32.435 34.424 214.49

Homogenity 59 22.797 34.361 231.23

Proposed 58.1507 21.6406 34.3655 243.61

carphone qcif

FFSearch 1180.938 1064.747 36.789 150.71

Umhex 262.952 146.67 36.729 153.85

Umhexsmp 221.124 105.73 36.73 151.68

Homogenity 203.334 88.935 36.711 153.88

Proposed 203.157 87.4971 36.7138 154.677

claire qcif

FFSearch 1526.664 1424.182 39.744 31.03

Umhex 226.552 126.101 39.596 31.29

Umhexsmp 167.9 68.321 39.533 30.94

Homogenity 172.077 72.528 39.615 30.96

Proposed 169.5722 68.9847 39.533 30.94

coastguard qcif

FFSearch 969.892 848.188 34.211 170.13

Umhex 273.97 150.728 34.204 170.99

Umhexsmp 228.006 107.192 34.207 170.56

Homogenity 192.873 72.631 34.185 170.81

Proposed 191.5548 70.7902 34.1913 171.276

container qcif

FFSearch 923.489 830.41 35.992 36.45

Umhex 174.499 81.791 35.961 36.86

Umhexsmp 144.53 51.832 35.941 36.58

Homogenity 140.355 48.513 35.939 36.98

Proposed 140.4631 48.1663 35.9345 36.659

foreman qcif

FFSearch 928.134 830.147 36.06 125.09

Umhex 231.246 132.687 35.993 129.7

Umhexsmp 203.169 104.072 36.001 126.23

Homogenity 175.967 78.738 36.011 130.96

Proposed 173.8248 76.202 35.9895 132.266

mobile qcif

FFSearch 963.292 777.915 33.22 378.69

Umhex 331.908 143.395 33.201 379.99

Umhexsmp 331.011 143.66 33.201 379.99

Homogenity 266.264 71.902 33.207 380.8

Proposed 255.4348 69.8024 33.2091 380.978

bus cif

FFSearch 1995.345 1740.457 35.029 940.92

Umhex 604.852 345.656 34.988 1005.84

Umhexsmp 547.367 293.197 35.023 949.58

Homogenity 444.73 188.965 34.966 1024.88

Proposed 429.1194 174.3156 34.9576 1050.853

mobile cif

FFSearch 3735.533 2942.544 34.12 1478.56

Umhex 1231.955 570 34.085 1489.51

Umhexsmp 1123.831 467.731 34.101 1482.85

Homogenity 960.884 308.059 34.083 1502.78

Proposed 954.6059 300.3608 34.0842 1499.15

shields sd

FFSearch 25606.11 22937.532 35.828 923.46

Umhex 6221.245 3441.304 35.777 937.41

Umhexsmp 5452.145 2800.285 35.8 922.94

Homogenity 4750.687 2119.94 35.782 931.61

Proposed 4660.5152 2014.2926 35.7696 935.503

ice hd

FFSearch 66966.356 64304.688 41.595 2167.16

Umhex 7963.397 5263.388 41.508 2385.62

Umhexsmp 5977.357 3365 41.54 2189.28

Homogenity 5791.631 3199.114 41.534 2254.16

Proposed 5690.359 3112.7386 41.535 2293.154

RandomInitialSearchPointsPredictionforContentAwareMotionEstimationinH.264

163

Table 3: Normalized Results for Different Videos.

Video Algorithm Name Total Time Saving (%) MET Saving (%) PSNR Loss(dB) Bit Rate Increases(%)

bus qcif

FFSearch

Umhex 28.888059 45.06932021 0.133507474 -6.478514889

Umhexsmp 36.06852768 54.6693314 0.089972428 -1.060120618

Homogenity 44.91644104 68.13925537 0.27281962 -8.947418017

Proposed 45.70936421 69.75542263 0.259759106 -14.78043724

carphone qcif

FFSearch

Umhex 77.73363208 86.22489662 0.163092229 -2.083471568

Umhexsmp 81.27556231 90.0699415 0.160374025 -0.643620198

Homogenity 82.78199194 91.64731152 0.212019897 -2.103377347

Proposed 82.79698003 91.78235769 0.204408927 -2.632207551

claire qcif

FFSearch

Umhex 85.16032342 91.14572435 0.372383253 -0.837898808

Umhexsmp 89.0021642 95.20279009 0.530897746 0.290041895

Homogenity 88.7285611 94.90739245 0.324577295 0.225588141

Proposed 88.89263125 95.1561879 0.530897746 0.290041895

coastguard qcif

FFSearch

Umhex 71.75252502 82.2294114 0.020461255 -0.505495797

Umhexsmp 76.49160938 87.36223573 0.011692146 -0.252747899

Homogenity 80.11397145 91.436922 0.075998948 -0.399694351

Proposed 80.24988349 91.65394936 0.057583818 -0.673602539

container qcif

FFSearch

Umhex 81.10437699 90.15052805 0.086130251 -1.124828532

Umhexsmp 84.34956995 93.75826399 0.141698155 -0.356652949

Homogenity 84.80165979 94.15794607 0.147254946 -1.454046639

Proposed 84.78995418 94.19969654 0.159757724 -0.573388203

foreman qcif

FFSearch

Umhex 75.08484766 84.01644528 0.185801442 -3.68534655

Umhexsmp 78.10994964 87.46342515 0.163616195 -0.911343832

Homogenity 81.04077644 90.51517382 0.135884637 -4.692621313

Proposed 81.27158363 90.82066188 0.195507488 -5.736669598

mobile qcif

FFSearch

Umhex 65.54440398 81.56675215 0.057194461 -0.343288706

Umhexsmp 65.63752216 81.53268673 0.057194461 -0.343288706

Homogenity 72.35895243 90.75708786 0.039133052 -0.557183976

Proposed 73.48313907 91.02698881 0.032811559 -0.604188122

bus cif

FFSearch

Umhex 69.68684613 80.13992877 0.117045876 -6.899630149

Umhexsmp 72.56780156 83.15402219 0.017128665 -0.920375802

Homogenity 77.7116238 89.14279411 0.179850981 -8.923181567

Proposed 78.49397473 89.98449258 0.203831111 -11.68356502

mobile cif

FFSearch

Umhex 67.02063668 80.62900674 0.102579132 -0.740585434

Umhexsmp 69.91510984 84.10453675 0.055685815 -0.29014717

Homogenity 74.2771915 89.53086173 0.108440797 -1.638080294

Proposed 74.44525587 89.79247889 0.104923798 -1.39257115

shields sd

FFSearch

Umhex 75.70406048 84.99706071 0.142346768 -1.510623091

Umhexsmp 78.70764048 87.79169006 0.078151167 0.056309965

Homogenity 81.44705697 90.75776766 0.128391202 -0.882550408

Proposed 81.79920652 91.21835514 0.163001005 -1.304117125

ice hd

FFSearch

Umhex 88.10836146 91.81492335 0.209159755 -10.08047398

Umhexsmp 91.0740895 94.76710003 0.132227431 -1.020690674

Homogenity 91.35143175 95.02506878 0.146652242 -4.014470551

Proposed 91.50265993 95.15939087 0.144248107 -5.813783938

VISAPP2015-InternationalConferenceonComputerVisionTheoryandApplications

164

ME time is saved. ME Time saved is ranging from

0.4% to 1.62%. Most of the videos considered as

test sequences are having higher motion in their video

contents thus it can be said that it is saving more time

than the other algorithms. Also as the resolution of

video is increased no substantial change is observed

in Homogeneity algorithm and proposed algorithm in

terms of time but bit rate is increasing by .04 to 0.08

ACKNOWLEDGEMENTS

This work is supported by Center of Excellence -

Signal and Image processing, College of Engineer-

ing, Pune,India under Technical Education Quality

Improving Program (TEQIP) PHASE-II.

REFERENCES

DongYoon Kim, H. L. and Park, H. (2013). Iterative true

motion estimation for motion-compensated frame in-

terpolation. In IEEE Transactions on Circuits and Sys-

tems for Video Technology, Vol. 23, NO. 3. IEEE.

Humaira Nisar, A. S. M. and Choi, T.-S. (2012). Con-

tent adaptive fast motion estimation based on spatio-

temporal homogeneity analysis and motion classiﬁca-

tion. In Pattern Recognition Letters 33. IEEE.

K. Venkatachalapathy, R. K. and Viswanath, K. (2004).

A new adaptive search strategy for fast block based

motion estimation algorithms. In Journal of Visual

Communication and Image Representation. 15. Sci-

ence Direct.

Nie, Y. and Ma, K.-K. (2002). Adaptive rood pattern search

for fast block-matching motion estimation. In IEEE

Transactions on Image Processing, Vol. 11, NO. 12.

IEEE.

Po, L.-M. and Ma, W.-C. (1996). A novel four-step search

algorithm for fast block motion estimation. In IEEE

Transactions on Circuits and Systems for Video Tech-

nology, Vol. 6, NO. 3. IEEE.

Renxiang Li, B. Z. and Liou, M. L. (1994). A new three-

step search algorithm for block motion estimation. In

IEEE Transactions on Circuits and Systems for Video

Technology, Vol. 4, NO. 4. IEEE.

Sarwer, M. G. and Wu, Q. J. (2009). Adaptive search area

selection of variable block-size motion estimation of

h.264/avc video coding standard. In IEEE Transac-

tions on Circuits and Systems for Video Technology,

Vol. 18, N. IEEE.

Toivonen, T. and Heikkil, J. (2006). Improved

unsymmetric-cross multi-hexagon-grid search algo-

rithm for fast block motion estimation. In JInterna-

tional Conference on Image Processing (ICIP). IEEE.

Tourapis, A. M., Au, O. C., and Liou, M. L. (2002). Highly

efﬁcient predictive zonal algorithms for fast block-

matching motion estimation. In IEEE Transactions

on Circuits and Systems for Video Technology, VOL.

12, NO. 10. IEEE.

Xu, X. and He, Y. (2008). Improvements on fast motion es-

timation strategy for h.264/avc. In IEEE Transactions

on Circuits and Systems for Video Technology, Vol. 18,

NO. 3. IEEE.

Yi-Ching Liawa, J. b. and Zuu-ChangHong (2009). Fast

block matching using prediction and rejection criteria.

In Journal on Signal Processing. ELSEVIER.

Zhibo Chen, Peng Zhou, Y. H. (5-13 December, 2002). Fast

integer pel and fractional pel motion estimation for

jvt. In Joint Video Team (JVT) of ISO/IEC MPEG &

ITU-T VCEG (ISO/IEC JTC1/SC29/WG11 and ITU-

T SG16 Q.6) 6th Meeting: Awaji, Island, JP. ITU-T

— ISO/IEC Joint Video Coding Experts Group (JVT)

File:JVT-F017.doc,.

Zhu, C., Lin, X., and Chau, L.-P. (2002). Hexagon-based

search pattern for fast block motion estimation. In

IEEE Transactions on Circuits and Systems for Video

Technology, Vol. 12,NO.5. IEEE.

Zhu, S. and Ma, K.-K. (1997). A new diamond search algo-

rithm for fast block matching mot ion estimation. In

International Conference on Information, Communi-

cations and Signal Processing. IEEE.

RandomInitialSearchPointsPredictionforContentAwareMotionEstimationinH.264

165