1

Statistical Analysis of Sign Language VideoConference Traffic in Multipoint IP Sessions

S. Kouremenos, D. Kouremenos, S. Domoxoudis and A. Drigas

National Center for Scientific ResearchN.C.S.R. ‘Demokritos’Athens, Greece

Gallaudet University - Washington D.C.April 2004

2

VideoConference Traffic Modeling

A problem of major importance.Valuable insights about the resulting network load.A theoretical assessment of the network performance.Extensively studied in literature - Full Theoretical Models have been proposedComplex Procedure

Variation of videoconference session parameters Number of participantsTarget Video bit rateTarget Frame rate

Variation of videoconference content (H&S, Movies, Sign Language)Different Versions and Implementations of Video Codecs(H.261, H.263, H.263+, H.264)

3

Why Sign Language VideoConference Traffic Modeling is a separate research thread?

Increasing availability of affordable communication channels (ISDN, ADSL ) for the end-users (signers).Readily available videoconferencing software (MS NetMeeting).Established video coding standards (such as H.261 and H263).Exchange of bandwidth demanding qualitative video information - minimum video bit rate of 384 kbs and frames per second are at least 15 reported although 30 is ideal.Videoconferencing traffic modeling research has been tested ONLY on head&shoulders or movies.

4

Sign Language VideoConference Requirements Official application profile document of ITU

T1607250-99

20

12

8

SQCIF(112 × 96)

QCIF(176 × 144)

CIF(352 × 288)

Spatial resolution

Temporal resolution, fps

Good usability

Usable with some restrictions

Very limited usability

No practical usability

QCIF – 15fps

5

VideoConference Topologies

CLIENT TO CLIENT

(One-point Communication )

More Flexible

Less QoS capabilities

CLIENT TO MCU

(Multipoint Communication)

Better Synchronization, Control and QoS

Demand of large bandwidth for Continuous Presence

6

Experiments Description

Multipoint - Continuous Presence Video Conference Sessions between Native Greek Signers CISCO MCU 3510 in High Quality Mode (CIF)

Target Video Bit Rate = 320KBits/secTarget Frame Rate = 15fps

MS NetMeeting (to ensure the direct usefulness and applicability of our results)

QCIF H.261 and H.263 encoded video – Best Quality

1h Duration

7

Experiments Quantities at the Frame level

4742842892474282252508784910834000126850175860246830309660Variance

17271724172739661729117242941354343713877Average Frame Size (Bytes)

15151571599867Frame Rate(fps)

208.36207.69208.36217.22206.59867.63217.64216.79215.68215.95Video Bit Rate (Kbits/sec)

18001800Duration (sec)

1515Target Frame

Rate(fps)

3201280

Target Video Bit Rate (For

the MCU)(KBits/sec)

320320

Target Video Bit Rate (For Terminals) (KBits/sec)

Switched Presence - H263 Continuous Presence - H261 Scenario

MCUVC4VC3VC2VC1MCUVC4VC3VC2 VC1Terminal

21Exp

8

Sign Language VideoConference Traffic Analysis (1)

The frame sizes sequence is a Stationary Stohastic process with an AutoCorrelation Function Exponentially decaying and a Gamma-like Distribution.

0 100 200 300 400 5000

0.2

0.4

0.6

0.8

1

LAG

Auto

core

latio

n

H.261 Encoded Frame Sizes Sequence

0 100 200 300 400 5000

0.2

0.4

0.6

0.8

1

LAG

Auto

core

latio

n

H.263 Encoded Frame Sizes Sequence

Non MonotonicMonotonic

9

Sign Language VideoConference Traffic Analysis (2)

0 1000 2000 3000 4000 5000 60000

0.2

0.4

0.6

0.8

1

1.2

1.4x 10

-3

Frame Size (Bytes)

Den

sity

H.261 Encoded Frame Sizes Sequence

HIGH MOTIONLOW MOTION

0 1000 2000 3000 4000 50000

0.5

1

1.5

2

2.5x 10

-3

Frame Size (Bytes)

Den

sity

H.263 Encoded Frame Size Sequence

HIGH MOTIONLOW MOTION

The frame sizes sequence Distribution exhibits a Symmetrical Gamma-like Distribution similar in all cases

Large FrameSizes

Small Frame Sizes

10

Traffic Modeling (AutoCorrelation Function)

0 100 200 300 400 5000

0.2

0.4

0.6

0.8

1

lag

Auto

corr

elat

ion

H.261

SampleCompound Exponential Fit

0 100 200 300 400 5000

0.2

0.4

0.6

0.8

1

lag

Auto

corr

elat

ion

H.263

SampleCompound Ex ponential Fit

ρκ = wλ1κ + (1-w)λ2

κ, with |λ2| < |λ1| < 1

Compound Exponential Fit

Short Term Correlation (λ2)

Long Term Correlation (λ1)

What matters is the λ1 parameter

11

Traffic Modeling (Probability Density Function)

0 1000 2000 3000 4000 50000

0.2

0.4

0.6

0.8

1

1.2

1.4x 10

-3

Frame Size

Den

sity

H.261

SampleMOM

0 500 1000 1500 2000 2500 30000

0.5

1

1.5

2

2.5x 10

-3

Frame Size

Den

sity

H.263

SampleMOM

11 1( )( )

p xxf x ep

µ

µ µ

−−⎛ ⎞

= ⎜ ⎟Γ ⎝ ⎠, 0, 0p xµ > ≥ 1

0( ) p up u e du

∞ − −Γ = ∫2mp

v= v

mµ =

Gamma Density Function

MOM Method: and

12

Full Theoretical Models in LiteratureMarkov Chain Models

DAR(1) Model (D.P. Heyman)

ρ is the autocorrelation coefficient at lag-1

Q is a rank-one stochastic matrix with all rows equal to the probabilities resulting from the negative binomial density corresponding to the Gamma fit for the frame size distribution

C-DAR(1) Model (S. Xu, Z. Huang, and Y. Yao )

Ι (1 )darP Qρ ρ= + −

( )cdar darP f P I= −ln

( 1)f Tρ

ρ=

−The continuous version of DAR(1), where T is the frame rate of the videoconference traffic

13

Our generalization of C-DAR(1) model for Sign Language

Contribution of Results for simple and accurate modeling when using Sign Language

ρ=λ1 (close to 0.998) – Conservative Choice

Q is constructed via the MOM Method with p and µ parameters

68<p<72 and 43<µ<50 for H.261

62<p<70 and 24<µ<28 for H.263

14

Queuing Analysis via the C-DAR model and the fluid-flow methodAssume our model has M states, and the V is the rate vector V= (V1 , V2 , …, VM), where Vi is the video bit rate in state i. Then the traffic can be expressed as (Q, V ), where Q is the Transition Rate Matrix derived from the C-DAR model.The queue occupancy is a continuous random variable x, 0<x< K, where K is the queue buffer capacity. Define the steady-state probability distribution function Fi(x) as the joint probability that the buffer occupancy is less than or equal to x, when in the i state of the source model.If:

Then we have:

And the frame sizes overflow probability is:

where

1 2( ) [ ( ), ( ),..., ( )]TMF x F x F x F x=

ur1 2( , ,..., )MD diag d d d=

where di = Vi – C

( ) ( )TD F x Q F x⋅ = ⋅ur ur

( ) /i ii

Poverflow V C P V= −∑

i i

i S

V Vπ∈

= ⋅∑0

( ) ( ) ,i i

iPi Pi q t K

F K iπ−

+

∈Ω= = =

− ∈Ω

| ii V C+Ω = >

| ii V C−Ω = <

15

Analysis Results vs Trace Driven Simulation

0 5 10 15 20-2

-1.5

-1

-0.5

0

Buf f er Thres hold (Kby tes )

log1

0(Pr

[ove

rflow

])

Buf f er Ov erf low Es timation - C=220KBits /sec - V C2 - H.261

C-DA R

Simulation

0 5 10 15 20-2

-1.5

-1

-0.5

0

Buf f er Threshold (Kby tes)

log1

0(Pr

[ove

rflow

])Buf f er Overf low Estimation - C=220KBits /sec - V C1 - H.261

C-DA R

Simulation

0 5 10 15 20-2

-1.5

-1

-0.5

0

Buf f er Threshold (Kby tes )

log1

0(Pr

[ove

rflow

])

Buf f er Overf low Es timation - C=208KBits /sec - V C1 - H.263

C-DA R

Simulation

0 10 20 30 40 50-2

-1.5

-1

-0.5

0

Buf f er Thres hold (Kby tes )

log1

0(Pr

[ove

rflow

])

Buf f er Ov erf low Es timation - C=220KBits /s ec - V C2 - H.263

C-DA R

Simulation

16

Further Work

Experiments with different clients (CuSeeMe, VCON)

Modeling Analysis of the new codec H.264

Analysis of the traffic from the MCU in continuous presence mode

17

Thank you