AN ENHANCED SMOOTHING SCHEME FOR

MPEG VIDEO STREAMS TRANSMISSION

Joelson Tadeu Vendramin, Keiko Verônica Ono Fonseca

Centro Federal de Educação Tecnológica do Paraná – CEFET-PR,

Programa de Pós Graduação em Engenharia Elétrica e Informática Industrial – CPGEI,

Av. Sete de Setembro, 3165, Curitiba-PR, Brasil

Keywords: video smoothing, MPEG video transmission, video trace files.

Abstract: Compressed video transmission over UDP/IP-based networks leads some challenging studies. One of them

refers to the problem of minimizing the burstiness of video compressed traffic as it leads to poor network

bandwidth utilization. Video smoothing has been proposed as one solution to this problem. This paper

presents a smoothing algorithm to reduce such burstiness. The algorithm can bring good results in terms of

peak bandwidth reduction while keeps a simple implementation if compared to other smoothing schemes.

1 INTRODUCTION

Today, with the development of Internet and the

introduction of new network architectures, it is

possible to transmit multimedia real-time traffic that

have stringent Quality of Service (QoS)

requirements (Bakiras, 2002). That made services

such as Video-on-Demand (VoD), videophone, and

TV-quality videoconferences feasible. In order to

optimize storage and/or network resource capacity,

video compression has been largely applied. The

basic idea of video compression is to explore its

spatial and temporal redundancies. One widely used

family of open compression standards suitable for

transmitting video frames over UDP/IP-based

computer networks is that defined by the Motion

Picture Experts Group (MPEG).

Video compression, however, can present a

bursty nature and its traffic can lead to poor

bandwidth utilization. Therefore, effective

transmission schemes are required to transmit

streamed video across the network. In a client-server

scenario these schemes are known as smoothing

algorithms, as they provide a new and less bursty

transmission plan at the server side.

In that perspective, this paper presents a

smoothing algorithm developed to achieve higher

video quality under a given network capacity

condition, while making an efficient bandwidth

utilization. The algorithm uses the properties of a

compressed video in the MPEG format. This paper

is organized as follows. Section 2 describes the

MPEG video standard; section 3, presents concepts

of smoothing a video stream and section 4 the

algorithm implementation itself is described. Section

5 brings the experiments developed and the

simulation scenario. Finally, section 6 addresses the

main results and section 7 concludes the paper and

presents some future directions.

2 MPEG VIDEO STANDARD

Basically MPEG codes the video in a layered

structure where the most important element is the

frame (essentially the screen area). There are three

kind of frames in MPEG: Intracoded (I), predictive

(P) and bidirectional (B). Stand-alone static pictures,

coded with JPEG (Joint Pictures Expert Group)

algorithms, compose I-frames. P-frames are

composed with the block to block differences with

the last I or last P-frames. Finally B-frames are

predicted from nearest past I or P- frame and future I

or P-frame. These three frame types are arranged in

a pattern called Group of Pictures (GoP), which is

unique and repeats itself along the duration of the

video stream (see figure 1).

The MPEG video stream over an IP-based

computer network will be seen as a quasi-periodic

traffic, representing the GoP structures, with high

network utilization during the transmission of I

379

Tadeu Vendramin J. and Veronica Ono Fonseca K. (2004).

AN ENHANCED SMOOTHING SCHEME FOR MPEG VIDEO STREAMS TRANSMISSION.

In Proceedings of the First International Conference on E-Business and Telecommunication Networks, pages 379-384

DOI: 10.5220/0001393103790384

 SciTePress

frames and low-to-moderate network utilization

during the transmission of P and B-frames. Such

traffic generates Variable Bit Rates (VBR) that can

degrade the network bandwidth utilization if not

properly controlled. Hence, this justifies the

importance of study smoothing mechanisms (Zhang,

1997).

3 SMOOTHING OF THE MPEG

STREAM

The video transmission rate variability is a major

challenge in achieving optimal bandwidth

utilization. Considering the transmission of stored

video from a server to a client across the network, a

smoothing procedure tries to make the traffic less

bursty, thus avoiding the waste of network

resources. Also, as a general real-time design

principle, less bursty (or smoother) traffic is easier to

manage (Salehi, 1998), for example, when

multiplexing several streams.

Basically, smoothing techniques can be divided

into three categories: Smoothing by temporal

multiplexing, smoothing by stream aggregation and

smoothing by work-ahead (Salehi, 1998).

Smoothing by temporal multiplexing is achieved

when a buffer is introduced somewhere in the path

between the video server and the client. This can

smooth the stream’s peak rate but introduces extra

delay (Wrege, 1996). The stream aggregation

smoothing is statistical based. By the Central Limit

Theorem, when several independent streams share a

given resource their aggregate bandwidth

requirements converge to a Normal distribution,

provided that they have service times with finite

variance (Salehi, 1998). Finally, the work-ahead

smoothing is achieved when the server sends the

video data ahead of schedule. The constraints are

that (i) the data is available to be sent, and (ii) the

client has sufficient buffer space to receive it

(Reibman, 1992).

The work-ahead smoothing algorithms are often

referred as off-line smoothing in the sense that all the

work is done prior any transmission.

Another way to classify these algorithms is based

on some optimization criteria. For example the

critical bandwidth allocation (CBA) tries to

minimize the number of bandwidth increases (Feng,

1995; Sechrest, 1995), while the minimum changes

bandwidth allocation (MCBA) tries to minimize the

number of rate decreases (Jahanian, 1995). Another

optimization criterion (Salehi, 1998) is to achieve

the greatest possible reduction in rate variability.

A work-ahead smoothing algorithm, called

Enhanced Piecewise Constant Rate Transmission

and Transport (e-PCRTT) is particularly important,

since it was used as the “starting point” of our

algorithm. The e-PCRTT algorithm divides the

video stream into fixed size time intervals and then

computes, for each of them, a transmission rate that

will not overflow nor underflow the client’s buffer,

thus creating a transmission plan. As it knows, a

priori, all frame lengths and the client buffer

capacity, it can significantly reduce the traffic

burstiness (Hadar, 2001).

4 ALGORITHM PROPOSAL

We propose an extension (modification) of the e-

PCRTT algorithm. Instead of working with fixed

size time intervals, our algorithm tries to extend as

much as possible the calculated transmission rate,

always trying to fit the transmission plan adjacent to

the middle of the buffer’s limits, as will be seen

later.

(Feng, 1997) analyses several smoothing

algorithms and concludes that the PCRTT is

somewhat unstable under some conditions mainly

because of the partitioning of frames at fixed

intervals.

By removing this fixed size dependence and

maintaining the e-PCRTT’s simplicity as compared

to other smoothing schemes (like CBA or MCBA),

our algorithm can bring good results in the peak

bandwidth reduction.

The algorithm was developed using C language.

As input, it takes a text file (representing a original

video trace data

) and the client’s buffer size. Its

output is a new text file modified to reflect the new

transmission plan.

Figure 2 shows how the algorithm works. The L

curve represents the accumulated bytes sent to the

client during the video stream and the U curve is

equal to L curve plus the client’s buffer size. The

transmission plan is the traced line that must stay in

the middle path between L and U, or in other words,

the transmission plan must be feasible.

A video trace data is, basically, a text file that

carries video information such as: frame size, type, a

sequence number and its time.

I frames

B frames

P frames

GoP

Figure 1: An MPEG GoP pattern and the three

frame types.

ICETE 2004 - WIRELESS COMMUNICATION SYSTEMS AND NETWORKS

380

As can be observed, the transmission plan is

composed by successive constant bit rate runs.

These bit rates are incremented (points a and c) or

decremented (points b and d), thus preventing buffer

underflow or overflow, respectively.

Figure 3 brings the algorithm’s pseudocode.

Observe that it can be considered an hybrid version

of e-PCRTT and MCBA.

5 EXPERIMENTS AND

SIMULATION MODEL

To evaluate the algorithm two sets of experiments

were conducted. At the first experiment set, we

evaluate the mean bit rate and the peak bit rate,

using different trace files and different buffer sizes.

The following video trace files, obtained from

(Fitzek, 2001) were used: Jurassic Park (a film);

Silence of the Lambs (another film); ARD Talk (an

speech) and Susi und Strolch (a cartoon).

These files are coded with MPEG-4 format but

use the entire scene as a unique video object, so they

can be considered, without leak of accuracy, as

MPEG-2 streams. We adopted the same trace files

for sake of result comparison.

The duration of all trace files were truncated to

1800 seconds (half an hour), than each of them was

used as input to our algorithm with five different

buffer sizes: 256 k, 512 k, 1024 k, 2048 k and 3072

kbytes.

Finally, for each trace file (the original and

smoothed) the mean and the peak bit rates were

extracted. The peak-to-mean ratio, as will be

described, is assumed to be a performance metric of

the smoothing algorithm.

At the second experiment set, we evaluate the

performance of the smoothed stream in a client-

server scenario where a video source competes for

bandwidth with a concurrent traffic, i.e, any

network

application that shares a common link with the

video stream. A typical example is a File Transfer

Protocol (FTP) session.

A six-node NS-2 (Network Simulator) (Fall,

1999) scenario was built, implementing the client-

server video model and a client-server competing

FTP traffic. This model is depicted in figure 4.

All the network elements are identified as nodes,

and at each node, an agent can be attached. Here s

and s

represent the video server and the competing

traffic source, respectively. Their destinations nodes

are d

and d

. Node pairs define the network links.

Each of them has its own bandwidth, delay and

packet discard discipline parameters. The bottleneck

link is r

- r

; if it has not enough bandwidth to deal

with the aggregate traffic, packet losses will occur

and will be registered at an NS’ trace output file.

Node s

represents a FTP traffic source (Fall, 1999).

The following points were considered during the

experiment 2:

• The trace file used in all simulations was

Jurassic Park (and its 1024 k, 2048 k and

3072 kbytes smoothed versions);

• The smoothed trace files for buffers of 256 k

and 512 kbytes were not used in the

simulations;

• The simulation has a duration of 1800 s;

• The FTP source act three times during the

simulation: At 100, 500 and 1500 s, having

the following durations: 100, 500 and 100

seconds, respectively.

The next sections bring the performance metrics

used in the simulation as well its results.

6 RESULTS

Figure 5 brings the peak-to-mean ratio as function of

the client’s buffer size for the four trace files. In this

graph, a zero buffer means no smoothing (that is, the

original trace file).

The lower this ratio is, less bursty will be the

traffic. As it can be seen, better results are from the

ARD Talk trace: With a client’s buffer of 3072

kbytes the peak-to-mean ratio is almost 1. The Susi

und Strolch trace also brings good results, but for

buffer sizes upper to 1024 kbytes. Jurassic Park has

almost a linear behavior as the buffer increases and,

finally, the worst performance was observed in

Silence of the Lambs.

Transmission plan

curve

L curve

Bytes

Figure 2: A smoothing transmission plan.

AN ENHANCED SMOOTHING SCHEME FOR MPEG VIDEO STREAMS TRANSMISSION

381

For the second set of experiments, we are

basically interested in the following performance

metrics: (1) The use of bandwidth in the link

between the routers r

– r

(see figure 4); (2) The

client buffer instant occupancy; (3) The end to end

delay; (4) If there are lost packets, the effect of this

in the video frames.

As proposed in (Ito, 2002), we adopted two

performance metrics: the Frame Error Ratio (FER),

defined as the fraction of frames in error in each

video stream, and the network Useless Output (UO),

defined as the fraction of useless or unusable video

data from all video streams in a multiplexed

scenario. Frames in error are considered frames with

a lost fragment and all its propagated frames (P

and/or B), while the unusable video data is the data

received by users as part of an unrecoverable frame.

The use of the bandwidth in the link between the

routers for the Jurassic Park transmission as a

function of time is depicted in figure 6. In figure 6(a)

we have the original trace file and in figures 6(b), (c)

and (d), we have smoothed files for buffers of 1024

k, 2048 k and 3072 k, respectively. In all plots the

solid line refers to the traffic due to video only and

the dotted line is the total traffic (video plus FTP).

Observe that when FTP is present it takes all the

available bandwidth (2 Mbps) and, when not

present, the two lines are coincident.

// Part 1: Reads the input (file and buffer) and computes L/U curves

Read (Original_trace_file, Buffer_size)

For each frame in (Original_trace_file) do

Computes L[i] (sum of past frame sizes until frame i)

Computes U[i] = L[i] + Buffer_size

End

// Part 2: Computes the transmission rates inside each smoothing interval

“start_point” of transmission plan begins in L[1]

While exist frames in (Original_trace_file) do

Increments frames by 1

NewIncrement = True

While (NewIncrement = True) do

Computes “next_point” in transm.plan as (L[i]+U[i])/2

Computes “transm.rate” based on “next_point” and “start_point”

If (“transm.rate” upper U[i] inside interval) then // overflow

NewIncrement = False

If (“transm.rate” lower L[i] inside interval) then // underflow

NewIncrement = False

End

Defines the smoothing interval as (“start_point”, “next_point”)

Associates the “transm.rate” with the smoothing interval

Define a new “start_point” based on “next_point”

// Loop to start a new interval…

End

// Part 3: Adjusting the frame_time and frame_size

// for each video frame arrival time

For each video frame arrival time in (Original_trace_file)

Defines new (frame_time, frame_size) based on transm.plan

Write (Smoothed_trace_file)

End

Figure 3: Smoothing algorithm pseudocode

Figure 4: The simulation scenario for experiment 2.

- video server

- video client

- concurrent traffic source

- concurrent traffic destination

- routing nodes

100 Mbps

1 ms delay

2 Mbps

10 ms delay

0 500 1000 1500 2000 2500 3000 3500

1,0

1,5

2,0

2,5

3,0

3,5

4,0

4,5

5,0

5,5

Razão pico-média

buffer [kbytes]

Jurassic

ARDTalk

SusiUndSt

Silence

Buffer [kbytes]

Peak-to-mean ratio

Figure 5: The peak-to-mean ratio.

ure 4: The simulation scenario for ex

eriment 2

ICETE 2004 - WIRELESS COMMUNICATION SYSTEMS AND NETWORKS

382

The client buffer instant occupancy must show

that the buffer never underflows nor overflows. As

the use of the buffer depends fundamentally how the

codec (present at video client) consumes the video

frames, we made the following assumptions: (1) The

buffer will increase as the video packets arrive from

network; (2) The client will take from the buffer an

entire GoP or, equivalently, 12 frames (see figure 1)

every 480 ms; (3) The video playback at the client

will start after two complete GoP are present in the

buffer, this leads an initial playback delay of about

960 ms.

Considering those assumptions, we observed

that, in some instants, the nominal buffer limit was

reached. This can be easily overridden by increasing

in 10% the real buffer limit, or, by specifying a

lower value to the smoothing algorithm.

We can observe that in all cases, the end to end

delay increment is lower (about 6 %), but the delay

jitter has a higher increment (about 16 %). The

delay’s increment is because smoothing introduces

extra packet fragmentation and the delay jitter’s

increment is a consequence of the intensive use of

the client buffer.

The effect of the lost packets of video frames

was observed during the original trace file

transmission and in the trace smoothed for a 1024

kbytes of buffer. In all other cases, no losses were

present. This is mainly because the smoothing has

reduced the burstiness of the video traffic.

As explained before, FER is the fraction of

frames in error in the video stream. For the original

trace file transmission the calculated FER is

1,04 % and for the 1024 kbytes trace file, this metric

is 0,03 %. If these values were estimated only by the

packet loss ratio they would be 0,53 % and 0,01 %,

respectively.

The useless output (UO) represents the fraction

of unusable video frames that were delivered to the

client. For the original video traffic, 3,83 % of all

video data is unusable, while for the 1024 kbytes

trace file, we have 0,34% as useless output.

200

400

600

800

1000

1200

1400

1600 1800

200

400

600

800

1000

1200

1400

1600

1800

2000

2200

[kb

ps]

tempo [s]

video

total

0 200 400 600 800 1000

1200

1400

1600

1800

200

400

600

800

1000

1200

1400

1600

1800

2000

2200

[kb

ps]

tempo [s]

video

total

(a)

(b)

200

400

600

800

1000

1200

1400

1600 1800

200

400

600

800

1000

1200

1400

1600

1800

2000

2200

[kb

ps]

tempo [s]

total

video

0 200 400 600 800 1000

1200

1400

1600

1800

200

400

600

800

1000

1200

1400

1600

1800

2000

2200

[kb

ps]

tempo [s]

total

video

Time [s] Time [s]

Used bandwidth [kbps]

Figure 6: The used bandwitdth between routers for the client’s buffer size of: (a) no buffer; (b) 1024 kbytes;

kbytes; and (d) 3072 kbytes

AN ENHANCED SMOOTHING SCHEME FOR MPEG VIDEO STREAMS TRANSMISSION

383

7 CONCLUSION

The first set of experiments is related to the

smoothing algorithm itself. By analyzing the

smoothed video traces we can conclude that the

efficiency of the algorithm depends on the type of

the video. The

Susi und Strolch trace file have the

greatest reduction (about 65 %, based on the peak-

to-mean ratio metric), the worst performance was

observed in the Silence of the Lambs trace file (only

28 % of reduction). We think this happened because

in the

Susi und Strolch the instants of bursts are,

predominantly, punctual, while in the Silence of the

Lambs the burst is spread in time.

The smoothing algorithm proposed in this paper

is based on two other smoothing algorithms: e-

PCRTT and MCBA. We have showed that it can

bring good results in terms of peak bandwidth

reduction, keeping the simplicity of the e-PCRTT

computation associated with the ability of extend the

transmission rate runs as much as possible, which is

a characteristic present in MCBA.

The second set of experiments explores the effect

of transmitting such a smoothed video over a simple

network configuration where video competes with a

concurrent traffic (FTP). One characteristic of the

FTP traffic is that it occupies all the available

bandwidth when it is present. In this case, the

smoothing contributed to optimize the use of the

bandwidth by reducing the bursts.

The study of the client buffer occupancy is

interesting as it ensures that the buffer will never

overflow nor underflow, as this is the assumption of

the smoothing algorithm.

The end to end delay in all smoothed trace files

has almost the same value, which in turn, is higher

than in the original trace file transmission. The

delay’s variation is more sensitive in the smoothed

traces. This was expected by the principle of the

smoothing: The new transmission plan imposes

more or less delay once it changes the transmission

rate (increasing or decreasing it).

Finally, the study of packet loss is very important

in a video transmission since a single lost packet can

degrade a whole GoP. We measured this effect

through the FER and UO metrics, but since we are

not dealing with packet discard discipline, the study

is merely informative.

Some kind of intelligent packet discard discipline

at the routers leading in a better network utilization

can also be addressed.

REFERENCES

Bakiras, S., Li, V.O.K., 2002. Maximizing the Number of

Users in an Interactive Video-on-Demand System. In

IEEE Transactions on Broadcasting, Vol. 48, No. 4,

281-292.

Ito, M.R., Bay, Y., 2002. A Packet Discard Scheme for

Loss Control in IP Networks with MPEG Video

Traffic. In

8th IEEE International Conference on

Communication Systems 2002

Zhang, Z., Kurose, J., Salehi, J.D., Towsley, D., 1997.

Smoothing, Statistical Multiplexing and Call

Admission Control for Stored Video. In

IEEE Journal

on Selected Areas in Communications

, Vol. 15, 1148-

1166.

Salehi, J.D., Zhang, Z. Kurose, J., Towsley, D., 1998.

Supporting Stored Video: Reducing Rate Variability

and End-to-End Resource Requirements through

Optimal Smoothing. In

IEEE/ACM Trans.

Networking, Vol. 6, 397-410.

Wrege, D., Knightly, E., Liebeherr, J., Zhang H., 1996.

Deterministic delay bounds for vbr video in packet-

switching networks: Fundamental limits and practical

tradeoffs. In

IEEE/ACM Trans. Networking, Vol. 4,

352-362.

Reibman, A.R., Berger, A.W., 1992. Constraints on

variable bit-rate video for ATM networks. In

IEEE

Transactions on Circuits and Systems for Video

Technology

, Vol. 2, 361-372.

Feng, W., Sechrest, S., 1995. Smoothing and buffering for

delivery of prerecorded compressed video. In

Proc. of

the IS&T/SPIE Symp. on Multimedia Comp. and

Networking, 234-242.

Sechrest, S., Feng, W., 1995. Critical bandwidth allocation

for the delivery of compressed video. In

Comput.

Communication

, Vol. 18, 709-717.

Jahanian, F., Feng, W., Sechrest, S., 1995. Optimal

buffering for the delivery of compressed prerecorded

video. In

Proc. of IASTED/ISMM International

Conference on Networks

Hadar, O., Cohen, R., 2001. PCRTT Enhancement for

Off-Line Video Smoothing. In

The Journal of Real

Time Imaging

, Vol. 7, No. 3, 301-314.

Fall, K., Varadhan, K. (eds.), 1999. NS – Notes and

Documentation, tech. rep., The Vint Project.

Fitzek, F.H.P., Reisslein, M., 2001. MPEG-4 and H.263

Video Traces for Network Performance Evaluation. In

IEEE Network, Vol. 15, No. 16, 40-54.

Feng, W., Rexford, J., 1997. A Comparison of Bandwidth

Smoothing Techniques for the Transmission of

Prerecorded Compressed Video. In

Proc. IEEE

Infocom

, 58-66.

ICETE 2004 - WIRELESS COMMUNICATION SYSTEMS AND NETWORKS

384