CONSTANT BITRATE CONTROL FOR A DISTRIBUTED VIDEO

CODING SYSTEM

Mariusz Jakubowski

, João Ascenso

and Grzegorz Pastuszak

Institute of Radioelectronics, Warsaw University of Technology, 15/19 Nowowiejska Str., Warsaw, Poland

Instituto Superior de Engenharia de Lisboa – Instituto de Telecomunicaçőes

R. Conselheiro Emídio Navarro, 1, Lisbon, Portugal

Keywords: Wyner-Ziv coding, distributed video coding, rate control.

Abstract: In some distributed video coding (DVC) systems, the total bitrate depends mainly on the key frames (Intra

coded) quality and on the side information accuracy. In this paper, a rate control (RC) mechanism is

proposed to achieve and maintain a certain target bitrate for the overall Intra and WZ bitstream, mainly by

adjusting online the Intra frames quality through the quantization parameter (QP). In order to obtain a

similar decoded quality of Intra and WZ frames, the relevant parameters: QP for the key frames and the

quantization index (Q

Index

) for WZ frames are controlled jointly. The major novelty of this work is a

statistical model that expresses the relationship between Q

Index

and WZ frames bitrate. The proposed rate

control solution is integrated into the VISNET2 WZ codec and the experimental results demonstrate the

efficiency of the proposed algorithm to reach and maintain the target bitrate.

1 INTRODUCTION

Around 2002, a new video coding paradigm known

as distributed video coding (DVC) has emerged,

inspired by two Information Theory results from the

70’s: the Slepian-Wolf theorem (Slepian and Wolf,

1973) and the Wyner-Ziv theorem (Wyner and Ziv,

1976). The main advantage of DVC lies in emergent

application scenarios such as wireless video

surveillance, low-power video sensor networks and

mobile camera phones. In such applications, there

are strong requirements in terms of low encoding

complexity or a more balanced complexity

distribution between the encoder and decoder.

Improved error resilience is also a desired feature

since most of the considered channels are quite

noisy (e.g. wireless channels). In such scenarios,

DVC fits well, since it explores the video statistics,

partially or totally, at the decoder not at the encoder

side, as in traditional video coding solutions, e.g. in

MPEG-x and H.26x standards. In DVC, one of the

most interesting cases is the source coding of a

source X, while a source Y, known as side

information, is available at the decoder only. Wyner

and Ziv showed that for lossy coding under certain

conditions (Wyner and Ziv, 1976), there is no loss of

coding efficiency if the dependency between X and

Y is explored at the decoder with reference to the

case where joint encoding is performed (i.e. X and Y

are available at the encoder). This interesting result

opens the possibility to design a system where two

statistically dependent signals are compressed in a

distributed way (separate encoding, joint decoding)

while still achieving the coding efficiency of

conventional predictive coding schemes (joint

encoding and decoding). However, practical DVC

codecs did not yet achieve this target performance,

especially when low complexity encoding is a major

requirement.

One of the most interesting and used DVC

architectures is based on turbo codes and a feeback

channel (FC) to perform rate control at the decoder.

The feedback channel has a key role, since the

decoder, knowing the available side information, can

test for successful decoding (i.e. if most of the errors

were corrected) and ask for the necessary bitrate to

achieve a certain target quality (established by the

encoder). Actually, in this solution there is no bitrate

control. A certain quality is established by the

encoder and the decoder just spends the necessary

rate to achieve it.

However, when the video transmission occurs in

constant bandwidth or bandwidth limited channels, it

is necessary to have a fixed target encoding bitrate

131

Jakubowski M., Ascenso J. and Pastuszak G. (2008).

CONSTANT BITRATE CONTROL FOR A DISTRIBUTED VIDEO CODING SYSTEM.

In Proceedings of the International Conference on Signal Processing and Multimedia Applications, pages 131-138

DOI: 10.5220/0001937501310138

 SciTePress

for the whole transmission. In this case, the encoder

must allocate the bitrate among each coding unit

(e.g. frame) and control the encoder parameters, i.e.

adjust the quantization parameter, in order to spend

the allocated bits efficiently.

In this context, this paper presents an encoder

rate control technique which achieves a constant

bitrate while minimizing changes in the quality of

the decoded sequence.

2 VISNET2 WZ VIDEO CODEC

The overall Wyner-Ziv (WZ) coding architecture for

the VISNET2 video codec is illustrated in Figure 1.

This codec follows the architecture proposed in

(Brites et. al., 2006), except for the encoder rate

control module which is proposed in this paper.

The coding process starts by the division of the

video frames into key frames and Wyner-Ziv (WZ)

frames. Then, one or two key frames are encoded

using the H.264/AVC Intra mode (Wiegand et. al.,

2003) in order to guarantee that each GOP is

delimited by key frames. The quality and thus the

rate of each key frame is defined mainly by the

quantization parameter (QP).

The frames in between are WZ frames, which are

simply coded with a H.264/AVC 4×4 block-based

discrete cosine transform (DCT) followed by the

aggregation of DCT coefficients in 16 frequency

bands b

. Each band is uniformly quantized and

bitplanes are created and sent to the turbo encoder.

The encoder establishes the final decoded quality by

defining for each band b

the respective number of

bitplanes M

for which WZ bits are generated, i.e.

the amount of bitplanes that will have a small error

probability after turbo decoding. There are 8 4×4

quantization matrices (Brites et. al., 2006), which

define different M

values for each DCT band b

allowing to achieve different rate-distortion (RD)

performances. The quantization matrices used by

both encoder and decoder are defined by the Q

Index

parameter.

At the decoder, for each WZ frame, the side

information Y

, an estimate of the X

frame, is

created by motion compensated interpolation (MCI)

based on two references, one temporally in the past

and another in the future (for GOP = 2 the references

correspond to the key frames). Then, the DCT

transform is applied to the side information and,

with a Laplacian correlation model, soft-input

information is obtained for the turbo decoder. The

iterative turbo decoder uses the received parity bits

and the soft-input side information and attempts to

generate the decoded (with small error probability

) quantized symbol stream. If the decoding is not

successful (P

> 10

-3

) the decoder requests via the

feedback channel for more parity bits, until

successful decoding (P

< 10

-3

) is achieved. Each

bitplane of each band is turbo decoded starting from

the most significant biplane and the DC coefficient

band. A zig-zag scan order is followed for the DCT

bands. After turbo decoding all bitplanes of all DCT

bands for which WZ bits were sent, the quantized

symbol stream is obtained. Next, in the

reconstruction module, the side information is used

together with the decoded quantized symbol stream,

to obtain the decoded X

frame after the IDCT

transform.

Finally, the key frames and WZ coded frames are

mixed again to generate the decoded video sequence

with a quality defined by the QP (for key frames)

and Q

Index

(for WZ frames) encoding quantization

parameters. The bitrate is spent according to the side

information quality, i.e. the accuracy of the MCI

estimation.

A novel encoder rate control module is proposed

in this paper (see Figure 2) which needs as an input

the bits spent on the WZ and key frames and

allocates the available bitrate among the WZ and

Intra key frames by changing the QP (for key

frames) and Q

Index

(for WZ frames) according to the

rate control algorithm proposed in the next Section.

3 PROPOSED RC ALGORITHM

In Figure 2, the flowchart of the proposed rate

control algorithm is presented. For each GOP of the

sequence, the WZ encoder is run with some initial

value of Q

Index

and generates the parity bits. Next,

key frames are encoded with a new QP value which

is selected based on the bitrate of the previous GOP

and the predicted bitrate in the current and next

GOP. The WZ decoder, invoked in the next step,

uses these key frames and the parity bits produced

by the WZ encoder. In the last step, a new value of

Index

parameter is selected according to the QP

value in order to obtain similar WZ and Intra frames

quality. The procedure is repeated up to the last

GOP.

In the next subsections, a detailed description of QP

and Q

Index

selection procedures is given.

SIGMAP 2008 - International Conference on Signal Processing and Multimedia Applications

132

Figure 1: VISNET2 WZ Video Codec Architecture.

Select new

QIndex

Run

WZEncoder

Change QP

Run

IntraEncoder

Run

WZDecoder

Last GOP

YES

STOP

START

Figure 2: Rate control algorithm flowchart.

3.1 QP Selection Procedure

The QP selection procedure is a key element of the

algorithm. With the QP value is possible to control

the bitrate and quality of the Intra frames and,

indirectly, of the WZ frames, since the side

information (critical in the WZ codec RD

performance) is created based on the Intra frames.

The selection algorithm takes into account the bitrate

of the previous GOP and a predicted increase or

decrease of the WZ frames bitrate caused by a

possible change in Q

Index

. If the previous bitrate is

greater or smaller than the target one, QP should be

modified. It is known that in H.264/AVC a change

of QP by 1 corresponds to a change in bitrate of

approximately 12% (change of QP by 6 means that

bitrate is halved or doubled) (Wiegand et al., 2003).

According to this rule, the relationship between the

bitrate of Intra frames and the QP parameter can be

expressed as:

QPQP

−

(1)

where R

, R

are the previous and the predicted

Intra frames bitrate, respectively. QP

and QP

are

the previous and predicted QP, respectively. First,

the target bitrate of Intra frames in the current GOP,

, is estimated by

IPR

R −=

(2)

where R

is the target bitrate, IP is the period of Intra

frames in an encoded sequence, FR is the frame rate,

and R

WZ1

is the predicted bitrate of WZ frames in the

current GOP. To obtain a good estimation of R

WZ,

couple of experiments were performed to study the

influence of the Q

Index

and QP

parameters on the

overall WZ rate. In Figure 3, the average bitrate of

WZ frames according to the Q

Index

for the Foreman

QCIF sequence with fixed QP equal to 24 is shown.

In Figure 4, the average bitrate of WZ frames for

the same sequence according to the QP parameter,

maintaining fixed Q

Index

equal to 4, is presented. It

can be seen that from QP = 0 up to about 30, the

bitrate of WZ frames is almost constant and starts to

increase faster approximately from QP = 36. An

CONSTANT BITRATE CONTROL FOR A DISTRIBUTED VIDEO CODING SYSTEM

133

important conclusion follows from this study: if

Index

does not change, the bitrate of WZ frames in

the current GOP will remain similar to the previous

one. If Q

Index

changes, the bitrate of WZ frames in

the current GOP will change , but in a more difficult

way to predict. As seen in Figure 3, when Q

Index

incremented from 7 to 8 the bitrate is almost

doubled, but from 4 to 5 only an increase of 10% is

observed.

Several experiments were performed on the

Coastguard, Foreman, Hall Monitor, and Soccer

QCIF sequences in order to obtain a model which

describes the dependence of the bitrate of WZ

frames on Q

Index

when it changes from one value

) to another (Q

). This model can be presented

in the form of transition table (TT) which models the

WZ frames bitrate for different Q

and Q

pairs.

Each element of the table is described by the

following equation:

][

][][

]][[

IWZ

IWZIWZ

QRQR

QQTT

−

(3)

where R

] and R

] are the bitrates of WZ

frames at two different quantization indexes. For

each coded sequence it is possible to evaluate (3)

and obtain a different table. Table 1 and Table 2

show the model obtained for the Soccer and

Coastguard sequences. The strategy proposed here to

cope with this inter sequence variation is to use the

transition table with average values from the four

previously defined sequences, and update it with

actual values during the coding process. These initial

values are shown in Table 3. Now, the R

WZ1

term in

(2) can be expressed as

]][[*

10001 IIWZWZWZ

QQTTRRR +=

(4)

where R

WZ0

is the previous bitrate of WZ frames, Q

and Q

are the previous and current Q

Index

values

which are used to index the transition table. When

is estimated, QP

can be estimated by

substituting R

by R

in (1)

201

log*6

QPQP −=

(5)

However, in order to ensure a fast approach to

the target bitrate set at the encoder and to maintain

relatively smooth changes in quality, some

additional constraints are established for the QP

variation. First of all, QP can change freely between

0 and 51 only in the second GOP. The first GOP is

coded with some initial QP value and the outcome is

unknown. It can be far above or far below the target

bitrate, so it is necessary to get close to the target as

fast as possible. However, a rapid change of QP can

cause a temporary peak of bitrate within one GOP

and in reaction, QP in the next GOP would have to

change strongly again in the opposite direction. The

result would be rapid changes of PSNR between

GOPs which causes a flickering artefact, i.e. a

negative subjective impact in video quality. To

avoid such an instability, it is proposed to limit the

QP variation between consecutive GOPs: QP can be

increased by five and decreased by three at

maximum. It was found that the ability to decrease

the bitrate is more critical than increasing.

Additionally, the algorithm tries to predict the

consequences of decreasing the QP in the next GOP.

Foreman QCIF, 15 Hz

115

165

215

12345678

QIndex

- WZ bit rate [kb/s]

Figure 3: Bitrate characterization in WZ frames for

different Q

Index

values.

Foreman QCIF, 15 Hz

0 5 10 15 20 25 30 35 40 45

WZ bit rate [kb/s]

Figure 4: Bitrate characterization in WZ frames for

different QP values.

For example, if QP variation causes a change of

Index

in the next GOP which will lead to a large

increase of the WZ frames bitrate, the algorithm will

prevent such a situation and reduce the amount of

QP decrease. This approach allows a relatively

SIGMAP 2008 - International Conference on Signal Processing and Multimedia Applications

134

stable quality (PSNR) and the bitrate close to the

target after a few first GOPs.

Table 1: Bitrate transition table of WZ frames for the

Soccer sequence.

QI1

QI0

1 2 3 4 5 6 7 8

1 0 0.25 0.48 1.77 1.69 2.60 3.72 6.79

2 -0.20 0 0.18 0.99 1.15 1.87 2.77 5.22

3 -0.32 -0.15 0 0.69 0.82 1.43 2.19 4.26

4 -0.60 -0.50 -0.41 0 0.08 0.44 0.89 2.12

5 -0.63 -0.53 -0.45 -0.07 0 0.34 0.75 1.90

6 -0.72 -0.65 -0.59 -0.30 -0.25 0 0.31 1.17

7 -0.79 -0.73 -0.68 -0.47 -0.43 -0.24 0 0.65

8 -0.87 -0.84 -0.81 -0.68 -0.65 -0.54 -0.39 0

Table 2: Bitrate transition table of WZ frames for the

Coastguard sequence.

QI1

QI0

1 2 3 4 5 6 7 8

1 0 0.47 0.75 1.87 2.16 3.83 5.86 13.68

2 -0.32 0 0.19 0.94 1.15 2.28 3.65 8.96

3 -0.43 -0.16 0 0.63 0.80 1.76 2.91 7.37

4 -0.65 -0.49 -0.39 0 0.10 0.69 1.39 4.12

5 -0.68 -0.53 -0.45 -0.09 0 0.53 1.17 3.64

6 -0.79 -0.69 -0.64 -0.41 -0.35 0 0.42 2.04

7 -0.85 -0.78 -0.74 -0.58 -0.54 -0.29 0 1.14

8 -0.93 -0.90 -0.88 -0.80 -0.78 -0.67 -0.53 0

Table 3: Averaged transition table.

QI1

QI0

1 2 3 4 5 6 7 8

1 0 0.44 0.72 1.82 2.11 3.54 5.19 10.75

2 -0.30 0 0.19 0.96 1.16 2.14 3.28 7.09

3 -0.41 -0.16 0 0.64 0.81 1.64 2.59 5.79

4 -0.64 -0.49 -0.39 0 0.10 0.60 1.18 3.14

5 -0.68 -0.54 -0.45 -0.09 0 0.45 0.98 2.74

6 -0.78 -0.68 -0.62 -0.37 -0.31 0 0.36 1.56

7 -0.83 -0.76 -0.72 -0.54 -0.49 -0.26 0 0.88

8 -0.91 -0.87 -0.85 -0.75 -0.72 -0.60 -0.46 0

3.2 Q

Index

Selection Procedure

The Q

Index

parameter is selected according to the QP

value in order to maintain a similar quality of WZ

and Intra frames, reducing the flickering effect

which is quite important from the subjective point of

view. In (Areia et al, 2008), Q

Index

for four sequences

(Coastguard, Foreman, Hall Monitor, and Soccer)

are matched with QP parameters which give similar

Intra frames quality. These values are collected in

Table 4. Because an appropriate model which relates

these two parameters is difficult to find, it is

proposed to use the QP average values for these four

sequences (Table 4) and select the Q

Index

which is

matched with the QP equal or smaller than the

current QP value. For example, if current QP equals

38, the closest value in Table 4 is 39, which means

that Q

Index

3 is selected.

Table 4: Points of equivalent quality for the key frames

and WZ frames. C. – Coastguard, F. – Foreman, H. M. –

Hall Monitor, S. – Soccer.

Index

C. F. H. M. S. Avg.

1 39 42 37 45 41

2 38 40 36 44 40

3 38 39 35 42 39

4 35 36 33 38 36

5 34 35 32 38 35

6 33 33 31 35 33

7 31 31 29 31 31

8 27 26 25 26 26

4 EXPERIMENTAL RESULTS

The proposed algorithm was integrated into the

VISNET2 DVC codec, as shown in Section 2. For

the experiments, six QCIF format sequences were

used. Four of them had already been used in

previous experiments: Coastguard, Foreman, Hall

Monitor, and Soccer. Two additional sequences,

Paris and Stefan, were taken to verify if the used

parameters are not sequence dependent. The test

conditions are shown in Table 5.

Table 5: Test conditions.

Sequences Coastguard, Foreman,

Hall Monitor, Paris, Soccer,

Stefan

Intra Period 2

Domain Transform

Initial Q

Index

/QP 3/39

Key Frames Codec H.264

Frame Rate 15

Target Bitrate [kb/s] 250

In Figure 5, the resulting total bitrate for each

GOP and the PSNR for each frame is shown for all

test sequences. The initial Q

Index

/QP value is set at

3/39 for all sequences. For most of them, it gives the

initial bitrate far below the target one. It makes

CONSTANT BITRATE CONTROL FOR A DISTRIBUTED VIDEO CODING SYSTEM

135

possible to demonstrate the capability to reach fast

the target bitrate and maintain it fixed as desired. It

can be seen that in most cases the bitrate approaches

the target one after a few first GOPs. Except for Hall

Monitor, after the five first GOPs the difference is

not greater than 20%. A slow increase of the bitrate

in case of the Hall Monitor sequence is a price for a

relatively stable PSNR and a result of a lack of an

accurate model to express the relationship of the QP

and Q

Index

parameters. The algorithm does not

decrease the QP parameter further because it would

change the Q

Index

value, in such a way that a rapid

increase of the WZ frames bitrate would lead to an

overall bitrate overflow. The bitrate overflow for the

Coastguard sequence is visibly correlated with a

rapid decrease of PSNR at the frame number 38.

This frame is very blurred and causes an unexpected

increase of the WZ bitrate without the change of

Index

, together with a large decrease of PSNR. The

mechanism built in the algorithm efficiently

compensates the excess of bitrate by increasing QP

immediately. For the remaining sequences, bitrate is

very close to the target one but never exceeds it.

The second subject of interest was the

differences in the quality of Intra and WZ frames

within one GOP. In general, they remain within a

range of 1 - 2 dB after coding of a few first GOPs.

However, in more dynamic regions of a sequence

these differences can achieve 3-5 dB (Coastguard,

Foreman, Soccer) or in extreme cases even more

than 10 dB (Coastguard, Soccer).

5 CONCLUSIONS

The proposed method for DVC rate control confirms

its efficiency in terms of achieving and maintaining

the required bitrate. Thanks to the limitations

imposed on the QP variation, differences in the

quality between Intra and WZ frames fall, in

general, within a range of 1-2 dB. However, there is

still a lot of room for improvement. In future work, a

more accurate mechanism for the Q

Index

selection

and for the bitrate prediction of WZ frames should

be developed.

ACKNOWLEDGEMENTS

The work presented was developed within activities

of VISNET II, the European Network of Excellence,

(http://www.visnet-noe.org), founded under the

European Commission IST 6FP programme.

REFERENCES

Slepian, J., Wolf., J., 1973, “Noiseless Coding of

Correlated Information Sources”, IEEE Trans. on

Information Theory, vol. 19, nº 4, pp. 471 - 480.

Wyner, A., Ziv, J., 1976. “The Rate-Distortion Function

for Source Coding with Side Information at the

Decoder”, IEEE Trans. on Information Theory, vol.

22, nº 1, pp. 1 - 10.

Brites, C., Ascenso, J., Pereira, F., 2006. Improving

Transform Domain Wyner-Ziv Video Coding

Performance. In ICASSP’2006, IEEE Int. Conf. on

Acoustics, Speech and Signal Processing, May 14-19,

Toulouse, France.

Areia, J. D., Pereira, F., Fernando, W.A.C., 2008. Impact

of the Key Frames Quality on the Overall Wyner-Ziv

Video Coding Performance. In ELMAR-2008, 50th

International Symposium ELMAR-2008, September

10-13. Zadar, Croatia (submitted).

Wiegand, T, Sullivan, G. J., Bjontegaard, G., and Luthra,

A., 2003. Overview of the H.264/AVC Video Coding

Standard. In IEEE Trans. Circuits Syst. Video

Technol., vol. 13, no. 7, pp. 560-576.

SIGMAP 2008 - International Conference on Signal Processing and Multimedia Applications

136

(a)

(c)

(e)

(g)

(b)

(d)

(f)

(h)

Coastguard QCIF, 15 Hz

100

150

200

250

1 11213141516171

GOP number

Bit rate [kb/s]

Fore man QCIF, 15 Hz

110

130

150

170

190

210

230

250

1 11213141516171

GOP number

Bit rate [kb/s]

Hall QCIF, 15 Hz

100

120

140

160

180

200

220

240

1 112131415161 7181

GOP number

Bit rate [kb/s]

Paris QCIF, 15 Hz

150

170

190

210

230

250

1 11213141516171

GOP number

Bit rate [kb/s]

Coastguard QCIF, 15 Hz

1 21 41 61 81 101 121 141

Frame number

PSNR [dB

]

Fore man QCIF, 15 Hz

1 21 41 61 81 101 121 141

Frame number

PSNR [dB

]

Hall QCIF, 15 Hz

29.5

31.5

33.5

35.5

37.5

39.5

1 21416181101121141161

Frame number

PSNR [dB

]

Paris QCIF, 15 Hz

26.5

27.5

28.5

29.5

30.5

31.5

32.5

1 21 41 61 81 101 121 141

Frame number

PSNR [dB

]

CONSTANT BITRATE CONTROL FOR A DISTRIBUTED VIDEO CODING SYSTEM

137

(i)

(j)

(k)

(l)

Figure 5: Bitrate and PSNR changes: (a), (b) Coastguard, (c), (d) Foreman, (e), (f) Hall Monitor, (g), (h) Paris, (i), (j)

Soccer, and (k), (l) Stefan.

Socce r QCIF, 15 Hz

115

135

155

175

195

215

235

255

1 11213141516171

GOP number

Bit rate [kb/s]

Socce r QCIF, 15 Hz

13.5

18.5

23.5

28.5

33.5

38.5

1 21416181101121141

Frame number

PSNR [dB

]

Stefan QCIF, 15 Hz

150

170

190

210

230

250

1 6 11 16 21 26 31 36 41 46

GOP number

Bit rate [kb/s]

Stefan QCIF, 15 Hz

1 112131415161718191

Frame number

PSNR [dB

]

SIGMAP 2008 - International Conference on Signal Processing and Multimedia Applications

138