Scene Change Scale: An Indicator of Multiplexing MPEG Streams for ...

Scene Change Scale: An Indicator of Multiplexing MPEG Streams for The Assessment of QoS Guarantees Over ATM Networks Abdulfattah Mashat and Mourad Kara School of Computer Studies University of Leeds, Leeds LS2 9JT, UK E-mail: fabdul,[email protected]

Abstract

A variety of networked video applications use the MPEG coding scheme to reduce the required bandwidth for their transmission over the network. Encoded video trac is a correlated and bursty trac with high value of peak to mean ratio (burstiness). For ecient trac management in a high speed network, it is important to know the basic characteristics of such encoded video trac. This information can be used to study the performance of video transmission through this network. In addition, it can be used to develop appropriate control schemes for handling multimedia trac. An important reason for the uctuations in the overall bit rate is the scene changes within the video stream. As a result, an MPEG stream may have several spikes (peaks) due to scene changes. These may cause cell loss when multiple streams are multiplexed at an ATM switch. Therefore, we need to analyse their magnitude in order to achieve ecient management for this type of trac. This could be achieved by managing and classifying the video streams according to the amount of movement within the same stream. This paper presents a technique to classify MPEG streams using the amount of activity within each stream. Firstly, we analyse the variation of bit rate of an MPEG stream. We then present a method and an algorithm to identify the scene changes within the MPEG stream. Furthermore, based on the classi cation process of MPEG streams, we introduce a 'Scene Change Scale' exhibiting (grading) the amount of activity within the MPEG stream. The scale is used to demonstrate the impact of scene changes on QoS requirements. Consequently, we correlate the scale with the results obtained from simulation experiments. We show that this scale can be used to assess the QoS guarantees in terms of delay bounds and cell loss ratio (CLR) for multiplexed MPEG video streams.

1 Introduction For ecient trac management in a high speed network, it is important to know the basic characteristics of multimedia trac. This information can be used either to study the network utilisation or to develop appropriate control schemes for handling multimedia trac [Venturin95]. In order to achieve that, a real trac source should be analysed based on the measurements of the existing multimedia applications. In general, multimedia trac may be classi ed as data, audio and video trac. Audio and video represent real-time trac while data represents nonreal time trac. Each trac class could be characterised according to the trac behaviour and Quality of Service (QoS) requirements. For instance, there are two main factors that can have an in uence on the behaviour of the video trac: QoS requirements and the encoding schemes [Stamoulis94]. Nowadays, video becomes an increasingly important component of multimedia communications because of increasing the user demand for video and rapid advances in coding algorithms. The focus of this paper is on a particular coding algorithm which has recently received a great attention, namely the Motion Picture Experts (MPEG) standard. We have focused on MPEG because it is widely available and it has been standardised by CCITT. In addition, MPEG is 101/1

an example of Variable Bit Rate (VBR) video trac. VBR MPEG trac is expected to be one of the major trac sources for high speed networks. Compressed video trac (MPEG) exhibit complex patterns which vary from one stream to another. Managing VBR video trac is a very dicult problem due to the statistical properties of video stream which are depending on the coding scheme and the content of the video sequence [Kara97]. In Asynchronous Transfer Mode (ATM) networks environment, there are two main QoS requirements for VBR video trac which are dier from other trac types: Cell Loss and Cell Delay. The Cell Loss Ratio (CLR) is the ratio of the number of lost cells to the total number of cells sent by a trac source within an interval of time. CLR is an ATM-speci c metric and has a great impact on the QoS of VBR video trac due to the compression technique which remove the redundancy from the video images. One reason for a cell loss is buer over ows at a switching node [Onvural95]. The second requirement, the Cell Delay (CD), can be observed in various ways: coding, packetisation, propagation, transmission, switching, queuing and reassembly delay. The Cell Delay requirement can be described as set of delay constrains (bounds). The delay constrains vary based on the video services. For instance, interactive services require a short delay while the delay bound is less important in the case of distribution services. The aim of our work is to present a simple method and an algorithm to identify the scene changes within an MPEG video stream. Based on the amount of scene changes, A Scene Change Scale will be presented to exhibit the amount of activity within the MPEG stream. Furthermore, the impact of scene changes on QoS requirements will be explored. Our prime measurer of interest is the CLR at an ATM multiplexer. This means that we are considering the amount and magnitude of scene changes within the MPEG sequence in our analysis. For our analysis, we have used real empirical data sets for various MPEG sequences to de ne the statistical characteristics. This paper is organised as follows: In the next section, we present brief statistical characteristics of various MPEG sequences to determine their statistical properties. We then present a method to identify the scene changes within an MPEG stream. In section 3, we undertake several simulation experiments to show the performance of a multiplexed MPEG stream at a statistical multiplexer. We conclude in Section 4.

2 Characteristics of MPEG-Coded This section characterises VBR MPEG streams in terms of their statistical behaviour. The statistical behaviour of various MPEG streams could be explored using three statistical measures, namely distribution, autocorrelation function and scene change. The rst two measures have been covered in our previous work [Kara97]. The distribution parameters are important to describe and understand the main features of MPEG trac while the correlations have a great impact on the queuing performance of a statistical multiplexer [Sriram86]. However, the scene changes is being the focus of our study because the scene changes are one of the most important reasons for the uctuations in the over all bit rate within the video stream [Lazar93]. This section is organised as follows: Firstly, we present a brief description of MPEG-I encoder. We then present a brief analysis of MPEG behaviour. Lastly, A method and an algorithm are introduced to identify the scene changes within an MPEG stream.

2.1 MPEG encoder

The MPEG coding algorithm was developed initially to store the compressed video on a digitalstorage media [Pancha93]. MPEG is a exible coding scheme which makes this type of coding widely spreadable and the most frequently used standard for video encoding [Bunzel94]. A variety of the video applications (including video conferencing) use the MPEG coding scheme for reducing the required bandwidth [Rose95a]. There are two main types of MPEG coding schemes for video: MPEG-I and MPEG-II. Two more types are currently under development: MPEG101/2

4, a standard for multimedia application, and MPEG-7, a content representation standard for information search [ISO/IEC97]. The basic scheme of MPEG coding is to predict motion from frame to frame in the temporal direction, and then to use DCTs (Discrete Cosine Transforms) to organise the redundancy in the spatial directions. Thus, MPEG coding is a combination of interframe and intraframe coding techniques. Considering the output of MPEG-I encoder, the reduction can be achieved by producing three types of frames I, P and B frames (See Figure 1):

 I Frame (Intra frame): 'I' frames are simply a frame coded as a still image, not using any

past history.  P Frame (Predictive frame): 'P' frames predicted from the most recently reconstructed 'I' or 'P' frame.  B Frame (Bidirectional predictive): 'B' frames are predicted from the closest two 'I' or 'P' frames, one in the past and one in the future. GOP I

B

B

P

B

B

P

B

B

P

B

B

I

B

Mpeg Frame

Figure 1: Encoded MPEG Video Stream As a result, MPEG-I can be distinguished from other encoding schemes by bi-directional temporal prediction [Conti96]. Each of these frames uses a dierent coding algorithm. An MPEG encoder repeats these frames periodically. Each frame contains a two dimensional array of picture elements called pels. The output of the encoded stream (the sequence of decoded frames) contains a deterministic periodic sequence of frames such as [IBBPBBPBBPBB] which is called Group Of Pictures (GOP). An MPEG video stream can be classi ed into three layers: scene layer (contain almost the same pictures), GOP layer (contain deterministic periodic sequence of frames) and frame layer (contain dierent types of frames). The duration of these layers varies from several seconds to tens of milliseconds. MPEG trac can be characterised using dierent levels: macroblock, slide, frame, GOP , or even the entire MPEG stream. We will use the frame and the GOP level for our statistical studies.

2.2 Statistical Analysis of MPEG pattern

It is dicult to characterise a video trac by using a short sequence of real data (few seconds). For our statistical analysis study, we use a long (about 30 min) sequence of real MPEG video which contains 40000 frames. Empirical data sets for MPEG video streams have been retrieved from the ftp site [Wurzburg95]. The video sequences have been encoded at the Institute of Computer science, University of Wurzburg. These sets represent frame size traces from MPEGI encoded video sequences with encoder input 384X288 pel (Berkeley MPEG-encoder ver. 1.3 has been used). The traced videos were captured in motion-JPEG format from VCR (VHS) with a captured rate between 19 to 25 fps. Table1 shows some simple statistical results (frame-based) that we have obtained for some video sequence classes including movies, sport events, TV shows and video conferencing. The statistical properties of these video sequences are dierent depending on the moving activities of the sequence. For example, the size of B frames in the sport sequences has a large amount of change (in some cases it has the same size as P frames). This indicates a large number of movements in the input encoded sequence. The amount of activities within the video stream aects the support and management of QoS quarantees. Therefore, the amount of activities has 101/3

to be considered at the modelling process. Generally, it is possible to classify an MPEG sequence into three classi cations according to the amount of movements during the video sequence: High, Moderate and Low activity classes. It has been shown that the statistical behaviour of the video sequences for the same class is almost the same [Kara97]. However, we need to consider another precise way of classi cation based on the amount of activities within the video stream. Video Sequence

Sequence Type

Mean () Cell/frame

CoV(  )

Stdev() Cell/frame

Peak Cell/frame

Peak/Mean

Dino StarWars Race News Talk video Conference

Jurassic Park Movie Star Wars Movie Formula 1 car race TV News TV Talk show Set-top Conference

35 25 80 40 38 16

1.13 1.38 0.69 1.27 1.14 1.93

39 34 55 51 43 30

312 325 527 495 279 121

9.14 13.4 6.58 12.36 7.34 7.66

Table 1 Where : CoV is the Coecient of Variation  And Stdev is standard deviation .

The size of GOP (In our trace, the summation of frame sizes for every 12 consecutive frames) is in uenced by the video activity within the same video sequence. For instance, If there are lots of movements within a scene, the GOP size will be high and vice versa. Graphs 1 (a-c) show the time series graphs for some video sequences to compare the frame sizes. In addition, a high activity causes large sizes of P and B frames. For instance, the frame size of type I is very large in the 'Race' sequence (Graph 1(b)). In contrast, B and P frames in the 'video conferencing' sequence are small compared to I frames. This is because the amount of movements in the 'video sequence' is not large. In other words, there is a small amount of scene changes. Time series graph shows frame sizes of ’Settop’ sequence

Time series graph shows frame sizes of ’Race’ sequence

45000

Time series graph shows frame sizes of ’Dino’ sequence

140000

80000 ’dino_.IPB’

40000

70000

120000 35000

60000 100000

25000

20000

Frame Size (Bits)

Frame Size (Bits)

Frame Size (Bits)

30000 80000

60000

50000

40000

30000

15000 40000 20000

10000 20000

10000

5000

0

0 0

50

100

150

200

250 Frame

300

350

400

450

500

Graph 1 (a)

0 0

50

100

150

200

250 Frame

300

350

Graph 1 (b)

2.3 Scene Change Identi er

400

450

500

0

50

100

150

200

250 Frame

300

350

400

450

500

Graph 1 (c)

This section presents a method and describes an algorithm for identifying the amount of scene changes within an MPEG stream. In the visual sense, a scene can be de ned as a portion of the movie without sudden changes in view [Krunz96]. One of the most important reasons for the

uctuations in the over all bit rate are the scene changes within the video stream [Lazar93]. In an empirical data set for a traced MPEG trac, a signi cant change in the sizes of two consecutive GOP is an indication of a scene change. We have used GOP sizes for our analysis because a GOP contains most of the picture's detail (for our data traced, Every GOP is composed of one 'I' frame which contains most of the picture information and three 'P' frames. In addition, there are 8 frames of type 'B' which contain information of any changes with the previous GOPs).

101/4

Time Series of a traced GOP MPEG traffic

2nd Difference of a traced GOP MPEG traffic

700000

500000 ’DINO_GOP.dat’ 400000

600000 300000 500000

200000 100000

BIts

400000

0 300000

-100000 -200000

200000

-300000 100000 -400000 0

-500000 0

500

1000

1500

2000

2500

3000

3500

0

500

1000

1500

GOP

2000

2500

3000

3500

GOP

Graph 2

Graph 3

The method can be used for any MPEG sequence to identify the scene changes within the sequence. We will use 'Dino' sequence to demonstrate our method. The time series plot, see Graph 2, shows several spikes (peaks) due to the scene changes. These may cause cell losses when multiple streams are multiplexing at an ATM switch [Hyman96]. Therefore, we need to analyse their magnitude. This could be achieved by managing and classifying the video streams according to the amount of movement within the same stream. As described above, a scene change occurs when a GOP size is abnormally larger than its neighbours. Based on this fact, we can quantify the scene change in the following way: Let us assume that fXi g is the size of a GOP: fXi : i = 1; 2; :::; N g . At a scene change, the second dierence, Diff2 , will be large in magnitude and negative in sign [Hyman96].

Diff2 = ((Xi+1 ? Xi ) ? (Xi ? Xi?1 )) Graph 3 shows the plot of the second dierence for 'Dino' stream. Every large spike could be an indicator for a scene change. In order to quantify only the signi cant scene changes, we divide the second dierence by the average of the past few seconds (t). The period of the last few seconds, t, might vary. In some studies, the average length of a scene might range from 3 to 7 seconds [Krunz96]. We have tested various values for t, all of them gave similar results for fYi = 2; 3; :::; N ? 1g: Y = (Xi+1 ? Xi) ? (Xi ? Xi?1 ) i

1 t

Pi

j =i?t Xj

A signi cant scene change can be identi ed with every negative large spike when we plot the division result Yi from the above equation. We chose a Threshold (T ) as a critical value, where 0 < j T j < Maxy . The number of spikes below the Threshold indicates the amount of large movements within the same MPEG stream (See Graph 4). Lower values of the Threshold , j T j, capture more scene changes. In order to capture only the large scene changes, it is more obvious when T is below the mean value of fY g. The following algorithm depicts the method which identify the scene changes within an MPEG traced stream: READ Xi from a traced GOP le of size N Threshold = T FOR i = 1 TO N Diff2 = (Xi+1 ? Xi) ? (Xi ? Xi?1 ) CALCULATE  = Meanii?t Xi 2 Y = Diff  IF Yi < T THEN Scene Change is identified END LOOP 101/5

In order to justify our criteria, (See Graph 5) we plot fXi g time series and the second dierence Diff2 . It is clear that there is a good match between the two series. Division Result of a traced GOP MPEG traffic 2

Diff2 GOP Time Series

Threshold 1.5

1

0.5

0

-0.5

-1

-1.5

-2 0

500

1000

1500

2000

2500

3000

3500

0

500

1000

1500

GOP

2000

2500

3000

3500

GOP

Graph 4: Scene Change Identifying

Graph 5: The time series of fXi gand the associated Diff2 plots

This method gives an indication on the uctuations in the over all bit rate of an MPEG video stream. This indication could be varied from one stream and another due to the amount of activities within the stream. In order to classify MPEG streams according to the amount of activity and movement within the stream. We need to map the output of the scene changes method to a scale which gives a more precise indication on the amount and the strength of the bit rate variation. We have tested and scanned 21 traced MPEG-I streams using our method to de ne and classify the amount of activity for each stream. For a given Threshold value, these streams are scaled to be presented in a 'Scene Change Scale' which ranged from 0 to 1 [0:1]. If the amount of activity within the stream is limited, then the stream will be allocated near to 0 and visa versa. Graphs 6(a,b) show these streams on the Scene Change Scale with dierent values for the Threshold. It is important to notice that with a lower Threshold value, the strength value of the stream on the Scale will be increased. Scene Changes Scale with Threshold = -0.1

Scene Changes Scale with Threshold=-.05 ’Asterix’ ’Dino’ ’Lambs’ ’Mtv1’ ’News2’ ’Sbowl’ ’Soccer’ ’Talk1’ ’Term’ ’bond’ ’Atp’ ’Fuss’ ’MrBean’ ’News1’ ’Race’ ’Simpsons’ ’Star2’ ’Talk2’ ’Video’ ’Mtv2’

0

0.2

0.4 0.6 Scene Changes Scale

0.8

’Asterix.5’ ’Atp.5’ ’Bond.5’ ’Fuss.5’ ’Lambs.5’ ’Movie2.5’ ’MrBean.5’ ’Mtv.5’ ’Mtv1.5’ ’News1.5’ ’News2.5’ ’Race.5’ ’Sbowl.5’ ’Simpsons.5’ ’Soccer.5’ ’Star2.5’ ’Talk1.5’ ’Talk2.5’ ’Term.5’ ’Video.5’

1

0

0.05

0.1

0.15 Scene Changes Scale

0.2

0.25

0.3

Graph 6 (a,b)

3 Experimental Evaluation In the ATM trac management context, It is more common to test the QoS performance at a statistical multiplexer and to allocate the ecient buer size and bandwidth resources [Krunz96]. This section describes simulation experiments and presents the simulation results. The results describe the impact of the scene changes activities on QoS requirements relating the result to the Scene Change Scale.

101/6

3.1 Simulation Model

In an ATM network, cells have to be merged from dierent sources and routed to dierent destinations via switch paths. Then the cells will share the transmission links for part of their journey. The process of multiplexing and switching cells involves temporary storage of cells in a nite sized buer and the arrival cells form a queue in order to be served [Pitts96]. Therefore, the main tasks for the switch and the multiplexer are to provide a temporary storage for the arrival cells and then route them to the correct outgoing port while maintaining their QoS requirements. There are many ways to arrange a switch to provide a temporary storage [Perros96]. Depending on where buers could be placed in the switch: Input buering, Output buering and Cross-point matrix. We will study one of these ways which is called 'Output buering' where multiple sources have been multiplexed into a buer with one outgoing port (link).

CellCellCellCellCell

Service Rate C

Buffer size B

Figure 2: Multiplexing of multiple MPEG streams Each outgoing port and its buer could be represented as a queuing process. This type of process is known as an ATM Multiplexer [Perros96]. Any queuing process could be described as arrival customers, service time by the server, number of service channels and the buer capacity. The arrival customers can be speci ed as an input to the buer with an average number of arrivals per unit of time or they could be described by an average time between the arrivals. The arrival average could be either constant (deterministic) or variable (stochastic). The service time can also be described by a service rate. We will study a case when the arrival customers are VBR MPEG streams and there is only one service channel to serve the arrivals and the buer capacity is the waiting space (See Figure 2). The waiting space could be nite or in nite space, but in a real system the capacity must be nite. If the system capacity has been exceeded, then any incoming arrivals will be lost. Furthermore, larger buer sizes will increase the waiting time for arrivals to be served. Therefore, a trade-o between the delay and cell loss requirements should be achieved.

3.2 Performance Results

We have simulated the transmission of various video connections on an ATM multiplexer with a single link and a buer whose size B determined by the delay constraints (D) on data transmissions out of the multiplexer: B = D  L, where L is the Link speed. The cells arrive to the multiplexer from a number of real video MPEG connections (based on the empirical data sets). Each connection generates a frame consisting of a variable number of cells. For our sequence, the connection rate is 24 frames/sec. The First-Come-First-Serve (FCFS) discipline policy is used at the multiplexer. Our primary measure of interest is the CLR. When there is a space in the buer, then any incoming cells will be queued, otherwise the cells will be dropped. Due to the space limitation, only the results of the three chosen traces from the real MPEG sequences are presented, namely: 'Movie', 'Dino' and 'Talk' sequence. According to the Scene Change Scale, these traces represent three various classes of VBR MPEG streams: high, mod101/7

erate and low amount of activities. It can be seen from Graph 7, for all buer sizes studied, the losses resulting from the use of these sequences are compared. It is clear that the 'Movie' sequence produces the highest CLR while the 'Talk' produced the lowest. For small buers, the CLR is quite high. As buer increases, the CLR values start to decrease slightly. The delay bound (D) gives a similar performance because the buer size is a function of the delay constraints. Furthermore, it is important to notice that the 'Movie' sequence owns the highest value on the Scene Change Scale. We have tested the same simulation using various MPEG sequences. We have found that there is a strong positive correlation between the CLR obtained for the sequence performance and its associated position on the Scene Change Scale. Another observation can be seen from Graph 7, 'Dino' sequence produces CLR higher than the 'Talk' sequence, even though the mean value of the 'Talk' sequence is higher. For the sake of brevity, the CLR results for all sequence are not shown in this manuscript. However, we have compared the losses resulting at various loads. For instance, Graph 8 shows the performance of the 'Movie' sequence at loads: U=20%, U=40% and U=70%. It is clear that a higher loads resulting more value of CLR. 0.19

0.09 Dino Movie Talk

0.18

U = 70% U = %40 U = 20%

0.08

0.17 0.07 0.16 0.06

0.05

0.14

CLR

CLR

0.15

0.13

0.04

0.12 0.03 0.11 0.02 0.1 0.01

0.09 0.08

0 0

100

200

300

400

500 600 Buffer (Cells)

700

800

900

1000

Graph 7

0

100

200

300

400

500 600 Buffer (Cells)

700

800

900

1000

Graph 8

4 Conclusions An encoded video trac is a correlated and bursty trac with a high value of burstiness (peak to mean ration). In this paper, we have presented a brief description of the statistical analysis of various classes of MPEG streams. In order to explore the uctuations in the overall bit rate for an MPEG stream, we have introduced a simple method and algorithm to identify the scene changes within an MPEG stream. As a result, we have mapped the amount of scene changes into a Scene Change Scale which can be used to exhibit the amount of activity within the MPEG stream. We have also explored the impact of the scene changes on the QoS requirements. The primary measure of interest was the CLR. We have related the CLR results obtained from several simulations with the Scene Change Scale. We have found that there is a strong positive correlation between the CLR result and the Scale. In other words, Scene Change Scale can be used as an assessment of QoS guarantees over ATM networks. There are several directions to further this work. We would explore further statistical analysis of the magnitude for scene changes. Consequently, we would nd a suitable statistical model for the scene changes within an MPEG stream. Another thread we are exploring is a scene change based model for MPEG trac.

Acknowledgements: We would like to thank O. Rose for providing the data sets for the MPEG video sequences. We wish to thank M. Rahin, R. Wade and S. Hassan for useful discussions during the preparation of this manuscript. 101/8

References

[Bunzel94 ]M. Bunzel and S. Morris , 'Multimedia applications development', McGraw Hill

Inc, 1994. [Conti96 ]M Conti, E Gregori and A Larsson, 'Study of the Impact of MPEG-I Correlations on Video Sources Statistical Multiplexing', IEEE Journal on Selected Areas in Communications, Vol 14, No. 7, September 1996. [Danzig92 ]P. Danzig, S. Jamin, R. Cacerest, D. Mitzel and D. Estrin, 'An Empirical Workload Model for Driving Wide-Area TCP/IP Network Simulations', Journal of interworking, 1992. [Habib92 ]I. Habib and T. Saadawi, 'Multimedia Trac Characteristics in Broadband Networks', IEEE Communications Magazine, July 1992. [Heyman92 ]D. Heyman, A. Tabatabai and T. V. Lakshman, 'Statistical Analysis and Simulation Study of Video Teleconference Trac in ATM Networks', IEEE Transactions on Circuits and Systems for Video Technology, Vol. 2, No. 1, March 1992. [Hyman96 ]D. Heyman and T. Lakshman, 'Source Models for VBR Broadcast-Video Trac', IEEE/ACM Transaction s on Networking, Vol. 4, No. 1, Bebruary 1996. [ISO/IEC97 ]ISO/IEC working group, http://drogo.cselt.stet.it/mpeg/ . [Kara97 ]M. Kara and A. Mashat, 'Statistical Analysis and Modelling of MPEG Sources for Workload Characterisation of Distribution Multimedia Application', 5th IFIP workshop on Performance Modelling and Evaluation of ATM Networks, July 1997. [Krunz96 ]M. Krunz and S. Tripathi, 'Scene-Based Characterization of VBR MPEG-Compressed Video Trac", Technical Report TR-3573, Institute for Advance Computer Studies, Department of Computer Science, University of Maryland, 1996. [Lazar93 ]A. Lazar, G. Paci ci and D. Pendarakis, 'Modeling Video Sources for Real-time Scheduling, Technical Report 324-93-03, Columbia University, Department of Electrical Engineering and Center for Telecommunications Research, April 1993. [Onvural95 ]R. Onvural, 'Asynchronous Transfer Mode Networks:Performance Issues', Artech House INC, 1995. [Pancha93 ]P. Pancha and M. El Zarki, 'Bandwidth Allocation Schemes for Variable Bit Rate MPEG Sources in ATM Networks', IEEE Transactions on Circuits and Systems For Video Technology, Vol 3, No. 3, June 1993. [Pitts96 ]J. Pitts and J Schormans, 'Introduction to ATM Cesign and Performance', Wiley and Sons Ltd, 1996. [Rose95a ]O. Rose, 'Statistical properties of MPEG video trac and their impact on trac modelling in ATM systems', Report No. 101, Institute of Computer Science, University of Wurzburg, February 1995. [Rose95b ]O. Rose, 'Simple and Ecient Models for Variable Bit Rate MPEG Video Trac', Report No. 120, Institute of Computer Science, University of Wurzburg, July 1995. [Rose96 ]O. Rose, 'Estimation of the Hurst Parameter Long-Rang Dependent Time Series', Report No. 137, Institute of Computer Science, University of Wurzburg, February 1996. [Stamoulis94 ]G Stamoulis, M Anagnostou, ' Trac source models for ATM networks: a survey', Computer Communications Vol 17, number 6, June 1994. 101/9

[Sriram86 ]K Sriram and W. Whitt, 'Characterization superposition arrival processes in packet

multiplexers for voice and data', IEEE Journal on Selected Areas in Communications, Vol 4(6), september 1986. [Venturin95 ]R. Venturin, ' Trac Source Modelling of Multimedia Application for ATM', Informacija Telekomnni Kacije Automati, Vol 14, Iss 1-3, 1995. [Wurzburg95 ]ftp://ftp-info3.informatik.uni-wuerzburg.de/pub/MPEG/

101/10