Yet a Faster Motion Estimation Algorithm with Directional Search - - PDF document

▶
yet a faster motion estimation algorithm with directional
SMART_READER_LITE
LIVE PREVIEW

Yet a Faster Motion Estimation Algorithm with Directional Search - - PDF document

Dec 06, 2023 216 likes •351 views

15 th International Conference on Digital Signal Processing (DSP2007), July 2007, Cardiff, UK. Yet a Faster Motion Estimation Algorithm with Directional Search Strategies Speaker: Prof. W.C. Siu Ying Zhang+*, Wan-Chi Siu* and Tingzhi Shen+*

slide-1
SLIDE 1

1

W.C. Siu

Yet a Faster Motion Estimation Algorithm with Directional Search Strategies Speaker: Prof. W.C. Siu

Ying Zhang+*, Wan-Chi Siu* and Tingzhi Shen+* *The Hong Kong Polytechnic University, Hong Kong

+Beijing Institute of Technology, Beijing

15th International Conference on Digital Signal Processing (DSP’2007), July 2007, Cardiff, UK.

W.C. Siu

2

Introduction Outline

1

Proposed algorithm

2

Experimental results

3

Conclusion

4

slide-2
SLIDE 2

2

W.C. Siu

3

  • 1. Introduction of Motion Estimation
  • Motion estimation (ME) is a popular technique for hybrid video

coding, and has been adopted in video-coding standards to exploit the temporal redundancy existing between frames.

  • The most common approach of the ME is the block matching

algorithm (BMA).

  • It divides each frame into a number of non-overlapping

macroblocks (MBs), and to predict the motion activities of each

  • f these MBs by searching for the most similar block from a

reference frame which is the previous decoded frame usually. W.C. Siu

4

  • The motion of a MB is represented by a motion vector (MV)

which is the spatial displacement between the original position

  • f the MB and the position of the reference block.
  • Exhaustive full search (FS) method is the most basic method of

the BMA.

  • In the FS, a MB will find its best match in the reference frame

from all positions within a range called search window which is centered at the position of the current MB.

Introduction of Motion Estimation

Motion Estimation Error Signal Extraction

slide-3
SLIDE 3

3

W.C. Siu

5

Introduction

Motion Estimation and Fast Motion Estimation Algorithms Full Search Fast Full Search Fast search Pixel Decimation Search by Pattern: TSS, DS, BBGDS, .. Search by Scheme: MVFAST, PMVFAST, FAME, EPZS .. (with patterns) Proposed: by Scheme

ME

Jump to Terms

W.C. Siu

6

Proposed algorithm－Search patterns

Figure 1. The star based search pattern

The search pattern:

  • 1. the star (diamond) based search pattern
  • 2. the block based search pattern
  • 4
  • 3
  • 2
  • 1 0

1 2 3 4

  • 4
  • 3
  • 2
  • 1

1 2 3 4

slide-4
SLIDE 4

4

W.C. Siu

7

Proposed algorithm－Search patterns

  • 4
  • 3
  • 2
  • 1 0

1 2 3 4

  • 4
  • 3
  • 2
  • 1

1 2 3 4

Figure 2. The block based search pattern

The search pattern:

  • 1. the star based search pattern
  • 2. the block based search pattern

W.C. Siu

8

Proposed algorithm－Use of search pattern

Figure 3. Sample Search Procedure

A sample of search procedure with star based search pattern.

  • 4
  • 3
  • 2
  • 1

1 2 3 4

  • 4
  • 3
  • 2
  • 1

1 2 3 4 Step1 Step3 Step in the future Current best position Original position Step2

The Distortion value is decreasing , the sum of absolute differences (SAD)

slide-5
SLIDE 5

5

W.C. Siu

9

Proposed algorithm－Block classification

MV prediction MV Search

Final MV

Stage1 Stage2

Predicted MV = Predicted MV + additional MV

If additional MV of the block=0, Classify the block into the ordinary block group; Else Classify the block into the special block group.

In the previous frame: W.C. Siu

1 0

Proposed algorithm－Search pattern selection

If the current block is prejudged as a ordinary block, Choose the star based search pattern; Else Choose the block based search pattern. MV prediction MV Search

Final MV

Stage1 Stage2

Predicted MV Search pattern selection

In the current frame:

slide-6
SLIDE 6

6

W.C. Siu

1 1

Proposed algorithm－The directional information

SAD:1008

  • 4
  • 3
  • 2
  • 1

1 2 3 4

  • 4
  • 3
  • 2
  • 1

1 2 3 4 SAD:1745 SAD:1823 SAD: 982 SAD:2231 Current best position

Figure 4. The latest directional vectors

The directional information:

  • 1. the latest directional vectors
  • 2. the general directional vectors

Latest directional set right top

W.C. Siu

1 2

Proposed algorithm－The directional information

Figure 5. The general directional vectors

The directional information:

  • 1. the latest directional vectors
  • 2. the general directional vectors

General directional set right top

  • 4
  • 3
  • 2
  • 1

1 2 3 4

  • 4
  • 3
  • 2
  • 1

1 2 3 4 Step1 Step3 Step in the future Current best position Original position Step2

slide-7
SLIDE 7

7

W.C. Siu

1 3

Proposed algorithm－The directional information

Step 4,

Candidate points set Right, top points

The number of search points required can be reduced with the directional information.

Figure 6. Sample Search Procedure

  • 4
  • 3
  • 2
  • 1

1 2 3 4

  • 4
  • 3
  • 2
  • 1

1 2 3 4 Step2 Step3 Step1 Original position Step4

Latest directional set right top General directional set right top

Step 3, Policy: The Union of these two directional sets (‘Latest|General’)

W.C. Siu

1 4

Proposed algorithm－The flow chart

Search the surrounding area from the current best point with the selected search pattern. No No Yes Next block Early termination Classify blocks into Ordinary or Special blocks, & find the threshold for early termination. Note2 Note1 Calculate the statistical motion information. Yes Last block? Find SAD of the initial point.

Note2: MV=predicted MV+MV Note1: MV=predicted MV, for SAD < threshold

Obtain the predicted motion vectors and choose a new center for the search window.

slide-8
SLIDE 8

8

W.C. Siu

1 5

Some further Details:

Early termination approach:

We make use of the statistics of the previous frame, etc. to determine a threshold, for early termination. We assume that the numbers of stationary blocks between successive frames are similar.

To look for the centre of the search window:

motion vector = (0,0) with reference to the block location, (i,j), medium motion vector of (MV(i-1,j,t), MV(i,j-1,t) and MV(i+1,j-1,t)), MV(i,j,t-1), and MV(i+1,j+1) W.C. Siu

1 6

Experimental results

Table I: Performance comparison of different algorithms

PSNR : the average PSNR value of the luminance component of the decoded frames. SpUp : the speed-up on the average number of SAD checking points VS full search. SpUpt : the average speed up on real time measurement VS full search algorithm. Sequences Hall Container Flower PSNR SpUp SpUpt PSNR SpUp SpUpt PSNR SpUp SpUpt FS 37.21 1 1 35.64 1 1 23.87 1 1 TSS 37.22 31 27 35.58 31 28 23.36 31 25 DS 37.26 70 62 35.57 75 68 23.91 59 53 BBGDS 37.25 94 80 35.58 103 79 23.92 73 67 CDS 37.27 181 158 35.52 192 155 23.86 151 128 MVFAST 37.28 147 145 35.56 170 134 23.95 109 92 FAME 36.64 103 80 35.00 91 74 24.08 89 71 OurAlg. 37.34 445 273 35.75 786 370 23.96 157 118 bitrate=380K bit/sec.

slide-9
SLIDE 9

9

W.C. Siu

1 7

Experimental results

Speed up (real-time)

PSNR(dB) 31.8 80 100 120 140 160 32.2 32.3 32.4 BBGDS 31.9 20 40 60 32.0 32.1 32.5 Proposed MVFAST DS TSS CDS FAME

Further Results

W.C. Siu

1 8

Conclusion

The important points of this algorithm:

  • 1. The block classification;
  • 2. The directional information.

Simplicity, low memory, fast realisation speed and low error are our

  • bjectives.

Simulation results:

  • 1. Our approach is faster than all algorithms available to us in

the literature;

  • 2. It also has similar or even better performance on image

quality and bit-rates.

Jump to the End

slide-10
SLIDE 10

10

W.C. Siu

1 9

Future Work:

Sub-pixel motion estimation using directional information:

Elementary results obtained: a good trade-off between quality and realisation speed is obtained.

Motion Estimation with RDO setting:

  • 1. Simply using the Lagrangian Optimzation equation in H.264 (12.2)
  • 2. Similar results obtained as discussed

Further work on the simplification of the process to look for the centre of the search window, etc.:

motion vector = (0,0) with reference to the block location, (i,j), medium motion vector of (MV(i-1,j,t), MV(i,j-1,t) and MV(i+1,j-1,t)), MV(i,j,t-1), and MV(i+1,j+1) and the simplification

W.C. Siu

The End Thank you!

slide-11
SLIDE 11

11

W.C. Siu

2 1

Additional slices:

Performance comparison (with RD control)

  • -the average value of the results obtained with 18 sequences

0.1686 31.93 TSS 0.1679 31.91 CDS 0.1681 32.21 DS 0.1680 32.12 BBGDS 0.1661 32.08 FAME 0.1657 32.26 EPZS 0.1658 32.35 MVFAST 0.1653 32.41 Proposed 0.1653 32.58 FS Bits/pixel PSNR (dB)

Sequence

With RD

W.C. Siu

2 2

Additional slices:

Return

Performance comparison (with RD control)

  • -the average value of the results obtained with 18 sequences

0.1655 0.1660 0.1665 0.1670 0.1675 0.1680 0.1685 31.9 32.0 32.1 32.2 32.3 32.4 32.5 32.6 Bits/pixel PSNR (dB) Proposed MVFAST EPZS CDS BBGDS DS FAME TSS

slide-12
SLIDE 12

12

W.C. Siu

2 3

Experimental results

bits/sec 380000 (for rate control) 760000 (for rate control) PSNR Bits/pixel PSNR Bits/pixel FS 32.46982 0.1659 34.98632 0.3127 Proposed 32.40655 0.1656 34.88912 0.3126 MVFAST 32.33186 0.1659 34.85154 0.3128 EPZS 32.27764 0.1658 34.80252 0.3128 FAME 32.01032 0.1664 34.65904 0.3150 BBGDS 32.08266 0.1683 34.54872 0.3148 DS 32.15516 0.1679 34.63003 0.3143 CDS 31.88913 0.1688 34.39547 0.3156 TSS 32.06936 0.1681 34.57621 0.3149