`Development of a MPEG Data Stream
`Characterization for Use with ATM
`Networks
`
`Olen L. Stokes
`IBM Corporation
`P.O. Box 12195, Research Triangle Park, N.C., 27709
`(919) 467-5200, olstokes@vnet.ibm.com
`
`Arne A. Nilsson
`North Carolina State University
`Department of Electrical and Computer Engineering
`Raleigh, N.C., 27695, (919) 515-5130,/ax (919) 515-5523
`nilsson@eos.ncsu.edu
`
`Abstract
`The various parameters which affect MPEG data streams are discussed. An AR model is
`developed to characterize both constant bit rate and variable bit rate videos. Three videos with
`different bit rate modes, audio content, frame type mixture, and group of pictures sizes are
`analyzed. The accuracy of the AR model in predicting the size of the next frame based on
`previous frames is examined. Possible enhancements to the AR model are also introduced. The
`role of the AR model in the development of ATM networks is discussed.
`
`ATM networks, MPEG, AR model
`
`Keywords
`
`1 INTRODUCTION
`
`"We are moving into an age of Information Networking, with users anticipating increasing
`freedom to communicate with people and retrieve information anytime, anywhere, and in
`multiple media" (Albanese, 1991). These expectations translate into expanding markets and
`profits when desirable new services can be provided economically. The key to these services and
`their revenue growth is video. In fact, "digital video is likely to become the dominant B-ISDN
`
`R. Puigjaner (ed.), High Performance Networking
`© Springer Science+Business Media Dordrecht 1995
`
`Netflix 1033 - Page 1
`
`
`
`170
`
`Part Five Multimedia over ATM
`
`bandwidth driver, particularly for on-demand entertainment services" (Terry, 1992). Providing
`this information will require higher speed services with more flexibility than is possible with
`today's networks. Broadband ISDN (B-ISDN) networks with Asynchronous Transfer Mode
`(ATM) services have been proposed to provide the needed bandwidth and versatility.
`"The study of the statistical properties of packet video streams, to model video sources, is a
`required step in the process of designing B-ISDN networks to handle heterogeneous traffic"
`(Pancha, 1994). For network control, the ability to predict data rates could enable effective
`congestion control measures. For a constant bit rate (CBR) connection, the amount of bandwidth
`required and the degree of multiplexing possible are easily determined at connection setup. The
`task becomes to insure that this guaranteed traffic arrives with an acceptable delay and jitter.
`However, for a variable bit rate (VBR) video connection, the required bandwidth and resulting
`multiplexing must be adjusted over time. By predicting the upcoming requirements, the network
`can adjust its allocations to meet the new conditions. Likewise, an encoder could request a new
`allocation based on its predictions (Pancha, 1993).
`
`2 MPEGVIDEO
`
`In order to effectively utilize the available bandwidth and to provide the desired picture quality,
`compressed video encoding/decoding methods are being developed. One such standard which is
`receiving much attention has been developed by the Moving Pictures Expert Group (MPEG).
`The current standard is commonly known as MPEG-1 and is directed at providing image quality
`comparable to VHS video and sound quality similar to audio CDs. Although it is targeted for
`digital storage media, MPEG "is flexible enough to be used in a variety of video applications"
`(Pancha, 1993) including transmission through B-ISDN ATM networks. Also, an MPEG-2
`standard is being developed for higher resolution images and correspondingly higher data rates.
`The data stream from an MPEG video is inherently variable. This comes in part from the three
`different algorithms for encoding a video picture. An intra-coded picture, or !-frame, is an
`encoding of the picture based entirely on the information in that frame. A predictive-coded
`is based on motion compensated prediction between that frame and the
`picture, or P-frame,
`previous reference frame (I- or P-frame). A bidirectionally predictive-coded picture, orB-frame,
`is based on motion compensated prediction between that frame and the previous or next
`reference frame (I- or P-frame). The size of the resulting frame varies significantly between
`frame types. !-frames are the largest while B-frames are the smallest.
`Further, within each frame type, the size of the resulting encoded information varies. The size
`of an !-frame varies based on picture content. P- and B-frames vary depending on the motion
`present in the scene as well as picture content. Also, the quantizer-scale parameter (q) can be
`varied by the encoder to change the size and corresponding quality of each frame. A small q
`value produces a higher quality picture that requires a larger data rate. A large q value generates
`a smaller data rate at the expense of a lower quality picture. The value(s) chosen by the encoder
`are determined by the desire to either provide a constant picture quality or maintain a particular
`data bit rate.
`An MPEG data stream which contains audio as well as image data also includes system level
`control information. The control information defines the mixing of data packets which contain
`either audio or image data. This interleaving of audio and video packets is not necessarily
`
`Netflix 1033 - Page 2
`
`
`
`A MPEG data stream characterization for use with ATM networks
`
`171
`
`constant, nor are the audio packets always equally sized. Therefore, the audio data rate also
`varies over certain time scales.
`An MPEG encoder chooses between two basic service types: constant bit rate (CBR) and
`variable bit rate (VBR). Note that bit rate refers to the rate at which data is written to or taken
`from storage media (such as a CD-ROM) or transmitted over a network. For VBR video, the
`data rate is allowed to vary over time while the picture quality potentially remains constant. For
`CBR video, the picture quality is varied to insure that the resulting data stream can be written,
`read, or transmitted at a constant rate. To accomplish this, the data rate and picture quality are
`manipulated during encoding by adjusting q.
`This CBR data stream does not imply a constant frame size. Rather, it means that the output of
`an encoder buffer can be read at a constant bit rate. By using a sufficiently large buffer which is
`partially filled prior to removing the first video information, and by varying q so that the buffer
`never overruns or underruns, the encoder can produce a data stream which appears to be a CBR
`source to the remainder of the system. The encoder system shown in Figure 1 can operate in
`such a CBR manner or in a VBR manner. The CBR versus VBR decision is made at encoding
`time.
`
`Figure 1 MPEG encoding and decoding systems.
`
`The MPEG data shown in Figure 1 is part of a system level data stream as discussed above.
`The control information is not shown, nor are any padding data packets which might be required
`to obtain a CBR stream. The system and pad data rates are significantly smaller than the video
`and audio data rates. Note that many of the MPEG data streams available for analysis contain
`only image or audio information, but not both.
`The decoder system shown in Figure 1 can also operate under CBR or VBR conditions. The
`decode buffer is initially filled to a level specified by the data stream before any actual picture
`and/or audio information is removed from the buffer and decoded. Thereafter, the buffer is filled
`on one end by the arriving data and emptied periodically on the other by the decoders. As long
`as the buffer does not overrun or underrun, the video can be displayed properly. A generic
`MPEG decoder, the system target decoder (STD), is defined in the MPEG standard (ISO, 1993).
`The encoder must insure that the data stream it produces can be played properly by a directly
`connected STD. Therefore, the encoder must define the decoder minimum buffer size, initial
`
`Netflix 1033 - Page 3
`
`
`
`172
`
`Part Five Multimedia over ATM
`
`delay, and audio and video data rates. Note that any delay jitter introduced by transmission over
`a network is not necessarily anticipated by the encoder.
`The initial delay to allow the decode buffer to partially fill normally causes no concern when
`playback is from a digital storage medium. However, when the playback is part of an interactive
`or real-time display, this delay may create problems. For example, an initial delay in the decode
`buffer of 250 msec will significantly increase the probability that the round-trip delay will have
`a noticeable affect on a conversation. In these cases, a VBR mode of operation, where a frame's
`data is transmitted as soon as it is encoded, may be required. The encoder can either maintain a
`constant picture quality or continue to use the CBR algorithm without buffering. This latter
`mode would allow the data at the remote site to be both displayed as well as buffered for storage
`and future CBR playback.
`The frame type (I, P, or B) selected for a particular frame is based on a repeating sequence of
`frames called a Group Of Pictures (GOP). Two typical groups are shown in Figure 2. The
`sequence used is chosen by the encoder. The MPEG standard does not completely specify the
`encoding process, but rather defines a syntax from which decode by the STD is possible. This
`allows numerous encoding options which increases the difficulties in handling and controlling
`MPEG traffic.
`
`- - One Group Of Pictures - - - - - -
`
`OOQJ
`
`A. One 1-Frame each 30 frames
`
`B. One 1-Frame each 12 frames
`
`Figure 2 MPEG frame display sequence.
`
`•nllmage l . . [!fiia9e] ~!:
`\7-""'
`U
`~j:.mmmm~~~~~mmmmrn···
`
`CD-ROM
`
`ToATM
`
`~ ATMlink
`
`From
`
`To
`
`···mmmm~~~~~mmmmrn···
`\7 Reassembly
`Decoder.------, V
`•nl1mage IBJ!IIImage I ...
`Figure 3 MPEG-1 data stream to/from A TM
`cells.
`
`Once the data stream is encoded, it can be used locally or transmitted through a network. It can
`be displayed immediately or written to a digital storage medium for future display. Whenever
`the data stream is transmitted through an ATM network, the data must be divided into blocks
`which match the size of the ATM cell payload. This means 44- to 48-octet blocks depending on
`the ATM Adaptation Layer (AAL) used. At the receiving end, the data from the cell payloads is
`reassembled into the original MPEG stream. This process is shown in Figure 3. Should the data
`not be interleaved, different data types would travel on separate ATM virtual connections.
`
`Netflix 1033 - Page 4
`
`
`
`A MPEG data stream chflracterization for use with ATM networks
`
`173
`
`3 MPEG CHARACTERIZATION
`
`"In spite of all efforts, so far no general consensus has been reached for how to model data rates
`generated by typical video codecs ... " (Heeke, 1993). The derivation of an MPEG-1 mathematical
`characterization is challenging. As discussed above, MPEG-1 data streams are affected by
`numerous factors including picture content, object motion, audio content, and encoder options.
`The result is a random sequence of frame sizes which is difficult to analyze. Figure 4 shows
`three examples of MPEG-1 data stream frame sizes. The sizes are in terms of ATM cells with
`44-octet payloads.
`
`f rame Stret tor Star wan Vld•o
`
`Frame Sb:e• fo r ApoiD VIdeo
`
`Fram• Sllu tor Hockoy VIdeo
`
`Figure 4 Frame sizes for sample video sequences.
`
`These MPEG-1 data streams represent three different sets of encoder parameters. The Star
`Wars sequence is based on the image-only VBR data stream described in (Pancha, 1993). It
`contains only I- and P-frames and utilizes a constant q. The Apollo sequence is a system level
`CBR stream with interleaved audio and image (Aris). It includes all three frame types. The
`Hockey sequence is also a VBR image-only stream (North Valley Research). However, this
`encoding varies q by frame type and uses B-frames as well as 1- and P-frames. All of these
`videos contain camera shot changes as well as changes in movement and camera panning.
`Using these frame sizes as a sequence of random numbers, the average and standard deviation
`can be calculated as well as the autocorrelation functions (ACFs). These first two statistics are
`
`Netflix 1033 - Page 5
`
`
`
`174
`
`Part Five Multimedia over ATM
`
`shown in Table 1. The ACFs from the Hockey sequence, normalized with respect to the variance
`of the frame sizes, is shown in Figure 5.
`
`Table 1 Example video sequence statistics
`Video
`Mean
`Standard Deviation
`(ATM cells)
`(ATM cells)
`227.69
`106.09
`131.99
`86.79
`100.71
`69.49
`
`Star Wars
`Apollo
`Hockey
`
`Picture
`Sequence
`IPPPPPPP .. .
`IBPBPBP .. .
`IBBBBPB .. .
`
`GOP Size
`(Pictures)
`16
`30
`20
`
`The properties of the GOP size and frame mixture are present in the normalized ACFs. Each
`of the ACFs shows a periodic peak corresponding to a lag equal to the GOP size. At this point, a
`frame is always being correlated with another frame of the same frame type. Since the Star Wars
`video utilizes only P-frames between I-frames, there is a gradual decay in the correlation
`between the frames until the lag reaches the same point in the GOP cycle. However, in the other
`two videos, there is a pattern of P- and B-frames which produce additional peaks in the ACFs.
`These peaks also occur when similar frames are being correlated (P-frames to P-frames and B(cid:173)
`frames to B-frames).
`
`Auto-R191111i'lt coetlcitllt lor lillckoy Vi<leo
`
`lL
`
`~1
`
`E
`! 0.4
`E p.2
`
`fJ
`
`10
`
`16
`20
`LagN1111btr
`Figure 5 ACFs for Hockey video sequence.
`
`26
`
`30
`
`.0.2L.;. _
`
`_.____3. _
`
`_.__--L _ _.___.J._...I--__.L__j
`I
`10
`12
`14
`16
`18
`20
`Cootlici.,l Number
`Figure 6 ARCs for Hockey video sequence.
`
`To model the sequences, the ACFs are used to calculate autoregressive coefficients (ARCs) for
`each sequence (Shanmugan, 1988). The number of coefficients used is based on the GOP size,
`with sufficient coefficients included to allow past frames from one GOP to be utilized. The
`resulting Hockey ARCs are shown in Figure 6.
`From the ARCs, an autoregressive (AR) model for frame sizes is created based on:
`
`n
`
`1 n~J
`
`n
`
`x = [~(ARCx .)] + e
`.£....J
`i=l
`where: Xn = frame size sequence
`ARCi = autoregressive coefficient for lag i
`GOP = size of one group of pictures
`en = excitation sequence
`
`(1)
`
`The en data stream should ideally be an uncorrelated white Gaussian noise random sequence
`(Shanmugan, 1988).
`
`Netflix 1033 - Page 6
`
`
`
`A MPEG data stream characterization for use with ATM networks
`
`175
`
`To determine the characteristics of the excitation sequence, the following equation is used to
`extract the sequence:
`
`e = x - "'(ARcx .)
`
`n
`
`GOP
`[
`n ~ 1 n~
`i=l
`
`]
`
`(2)
`
`Note that the frame sizes from the frames of the first GOP are used to seed the calculation.
`Thus, the excitation sequence begins with the excitation value for the first frame of the second
`GOP. The resulting excitation sequence for Star Wars is shown Figure 7. At places where scene
`changes occur, the excitation sequence contains larger absolute value excitations.
`
`Excha1on FunctiOn 1tir Star Wan VIdeo
`
`Nonnalzed ACFs 1t1r Exeha!IOn Func1on cl Star wars \'Idee
`
`FmmeNumb•
`Figure 7 Excitation function for Star Wars
`video sequence.
`
`-1'----"-----'-----'----'---'---'-----'
`lO
`20
`2S
`1S
`D
`10
`lag Numbtr
`Figure 8 ACFs for Star Wars excitation
`function.
`
`The ACFs for the excitation sequence are then calculated. For white-noise, the sequence
`should be uncorrelated (Shanmugan, 1988). As is seen in Figure 8, the derived Star Wars
`excitation sequence is nearly uncorrelated. There are, however, slight peaks at the values
`corresponding to lags which are multiples of one GOP. These peaks will be discussed in more
`detail below.
`Further, the excitation distribution should be Gaussian with a mean of 0.0 (Shanmugan, 1988).
`From Table 2, all three sequences have very nearly a 0.0 mean. To allow comparison of the
`distributions, the excitation is normalized using the standard deviation of the excitation
`sequence. The resulting excitation histogram should produce a Gaussian curve with a standard
`deviation of 1.0. The distributions are very similar to the expected shape, as demonstrated in
`Figure 9. Note also that the standard deviations of the excitation sequences are considerably
`smaller than the standard deviations of the corresponding video sequences.
`
`Table 2 Excitation sequence statistics
`Video
`
`Star Wars
`Apollo
`Hockey
`
`Standard Deviation
`Mean
`(:iTM S.!fJ!!J_ __ ~~__fE.M.siJ!!L_.~-~~ ...
`55.16
`1.31
`0.37
`33.59
`14.53 · - - -
`0.88
`
`Netflix 1033 - Page 7
`
`
`
`176
`
`Part Five Multimedia over ATM
`
`Pn~llablly Df!lllllr fw Elcillllllft Funcllilft of SDrWart VIHo
`
`F11111e Size Plldlcllon lcr Hook., VIII•
`40o.-----.---r---..--....---,----,
`.,,,, PNdicltd
`
`·4
`
`5
`4
`2
`0
`~
`Ezdalion I E1cilllion Slllldlrd DIViallon
`Figure 9 Excitation probability density.
`
`10
`
`Figure 10 Prediction for Hockey video
`sequence.
`
`There are, however, more excitations which exceed the three sigma (±3o) values than would
`be expected for a pure Gaussian distribution. These excitations occur at, and soon after, video
`scene changes. These large values increase the excitation standard deviation. This alters the
`relationship of the Gaussian reference curves to the displayed distributions shown in Figures 9,
`12, and 14.
`For network control, these calculated ARCs can predict the next frame size through the use of
`following equation:
`
`GOP
`
`P. = L(ARC;x._J
`i=l
`where: Pn =predicted MPEG frame sizes
`
`(3)
`
`The prediction begins for the first frame in the second GOP to allow seeding with actual past
`values. The prediction error is then be defined as:
`
`GOP
`
`E.= x.- P. = x.- L(ARCixn-i)
`i=l
`where: En = prediction error
`
`(4)
`
`A comparison of Equations 2 and 4 reveals that the resulting error sequence En is identical to
`the excitation sequence en. Therefore, if the ARCs are known, the next frame size should be
`predictable to the accuracy demonstrated by the excitation sequence.
`Unfortunately, the calculations of the ACFs and ARCs above involve the use of future
`knowledge. That is, the ACFs and ARCs are calculated using the entire frame size sequence.
`Then, those results are applied to the sequence as post-processing. This is not possible in a
`network control or a real-time encoding environment. In these cases, only previously received
`frame sizes are available. However, it should be possible to provide the size of the GOP to the
`network at connection setup time. Note that the mixture of frames within the GOP is not
`required.
`To predict based only on prior frames, the ACFs are estimated based on the frame sizes that
`have already been received. The ARCs are then obtained from the estimated ACFs. The ACFs
`are calculated using the biased estimator:
`
`Netflix 1033 - Page 8
`
`
`
`A MPEG data stream characterization for use with ATM networks
`
`177
`
`1 N-fmf-1
`R(m) = N ~ (x.xn+m)
`where: R(m) = autocorrelation function (ACF) for a lag of m
`N = Number of frame sizes included in the calculation
`
`(5)
`
`If calculated in this manner, the storage of all previously received frame sizes would be
`required. To limit the storage requirements to the number of frames in one GOP, the following
`calculations are made beginning with the first frame of the second GOP:
`
`Sum.(m) = Sum ... Am)+(x.x,_m)
`R,.(m) = Sum.(m)
`n
`where: Sumo= 0. 0
`X-(GOP-1}-··XO =seeded values from first GOP frame sizes
`Rn(m) =estimate n of ACF for lag m
`
`(6)
`
`(7)
`
`The estimated ARCs are then calculated based on the estimated ACFs and used to predict the
`n+ 1 frame size using Equation 3. Note that numerous frames must be received before the ACFs
`and ARCs converge to values near those shown in Figures 5 and 6. Following that, if desired to
`reduce calculations, the ACFs and ARCs need only to be updated periodically. However, Sumn
`must be updated every frame.
`Since the ACFs calculated in Equation 7 are not the same as those used to extract the
`excitation sequence en, the prediction error En will differ from the excitation en. Figure 10
`shows the actual and predicted frame sizes for the Hockey sequence using the above
`methodology. Figure 11 shows the resulting prediction error for the Apollo sequence. The ACFs
`and ARCs are allowed to settle for 250 frames before "steady-state" prediction is assumed.
`Figure 12 shows the normalized distribution of the prediction errors for the Star Wars sequence.
`This distribution closely resembles the excitation distribution indicating that the next frame size
`can be predicted within the accuracy of the excitation sequence.
`
`RtaHilll P11d~1ion Emu lor Apolo ViiiD
`
`Sttad!-51111 Pre~1ion Emil Plllllabity Dtnoily of Star Wall Vileo
`
`1000
`
`1050
`
`IIOCI
`
`1150
`
`1200
`
`12ti0
`
`1300
`
`1350
`
`1400
`
`Figure 11 Prediction error for Apollo video
`sequence.
`
`-4
`
`-2
`
`I
`4
`2
`0
`Emu I Excitllan Standard Dowialan
`Figure 12 Prediction error probability density.
`
`10
`
`As noted above, the ACFs for the excitation sequences indicate a small correlation at lags
`equal to multiples of the size of the GOP. If this correlation could be used to estimate the above
`frame size prediction error, the uncertainty in the next frame could potentially be reduced. The
`
`Netflix 1033 - Page 9
`
`
`
`178
`
`Part Five Multimedia over ATM
`
`resulting prediction algorithm would be similar to an autoregressive moving average (ARMA)
`calculation. The following formula is used:
`
`CE. =E.- pE. = [x.- ~(ARC;x.J]- %(EARC1E __ 1)
`where: CEn = corrected prediction error
`En =non-corrected prediction error
`pEn = estimated non-corrected prediction error
`EARC = ARCs calculated for the non-corrected prediction error En
`
`(8)
`
`A method similar to that used above for calculating the estimated ACFs and ARCs is used to
`estimate the coefficients (EACF and EARC) of the prediction errors shown in Figures 10-12.
`Once again, the coefficients must be allowed to approach "steady-state" values before the results
`are stable. For this estimate, 200 prediction errors are utilized. The resulting correction
`coefficients EARC are substantially smaller than the prediction coefficients ARC.
`The resulting corrected prediction errors CEn are shown in Figure 13 for the Apollo sequence.
`The large periodic errors are noticeably reduced. The errors are again normalized by the
`excitation standard deviation and the Star Wars distribution is shown in Figure 14. The corrected
`error distribution shows marginal improvement over the non-corrected prediction errors in
`Figure 12. The corrected distributions have slightly smaller standard deviations and fewer errors
`greater than +4a. This only slight improvement is reasonable considering the small size of the
`correlation peaks of the excitation sequence ACFs. The small EARCs also reinforce this.
`
`Rea~ Tine Cona!td PrediGIIon Ernr for Apollo VIdeo
`
`Probal!lly Densh)' lor Conalld PrediGIIDn Enrot Sial Wart Vileo
`
`3GO
`
`1000
`
`1050
`
`11CICI
`
`1150
`
`1250
`1200
`F•moNIII!Ib•
`Figure 13 Corrected prediction error for
`Apollo video sequence.
`
`1300
`
`1350
`
`1400
`
`8
`I
`2
`0
`~
`~
`Co111at1d Enrl Exei•tion Standard Doviltion
`Figure 14 Corrected prediction error
`probability density.
`
`~
`
`4 MPEG IN ATM NETWORKS
`
`The understanding obtained in the development of the above characterization of MPEG-1 video
`data streams can aid in the development of B-ISDN ATM networks which must transport these
`videos. For network control of a VBR connection, the network must manage a new load during
`each frame period. The more accurately that this load can be predicted, the better that statistical
`multiplexing can be performed. Also, a network policing function could use this predicted load
`to set the permitted data rate for the next frame. Should the encoder also use the same algorithm,
`it would know the load expected by the network for the upcoming frame. If a larger than
`
`Netflix 1033 - Page 10
`
`
`
`A MPEG data stream characterization for use with ATM networks
`
`179
`
`expected frame is produced (corresponding to a large positive prediction error), a layered
`encoder similar to that discussed in (Pancha, 1993) and (Pancha, 1994) could mark an
`appropriate number of cells as low priority.
`If the network does not provide this level of control, the encoder can also take preventive
`action. If the network supports a fast-reservation protocol, the encoder can request a different
`bandwidth allocation based on its own prediction of the next frame size. This mode of operation
`is well described in (Pancha, 1993) and the AR prediction method presented above provides an
`alternate method of predicting the next frame size.
`The decision to incorporate the additional complexity of generating predicted error corrections
`must be made after ATM networks are better understood. While the improvement is slight, this
`improvement might be worthwhile in network control environments. If heavily loaded networks
`are found to become unstable with small increases in load, the enhanced prediction may prevent
`problems. Since large positive prediction error values represent large underestimates of the
`frame size, reduction of errors greater than +4a might significantly increase stability.
`Further analysis might reveal other, less complex methods which can address the repeatability
`(correlation) of prediction errors. A simple strategy might be to increase the predicted frame size
`by some percentage of the prediction error which occurred one GOP prior. This approach could
`even be modified to correct the prediction only when the prior error was greater than some
`minimum value. Further analysis is required.
`
`5 CONCLUSIONS
`
`The task of defining a perfect model of an MPEG-1 data stream remains a challenge. However,
`an AR model can be used to predict the next MPEG-1 frame size within the tolerances of an
`excitation sequence whose properties can be determined. The AR model is constructed based on
`the previous frame sizes and the number of frames in a GOP. This prediction can be slightly
`improved by correcting the esti)llllte through the use of another AR prediction based on past
`errors. The resulting AR models can be used by network control functions as well as MPEG-1
`encoders.
`
`6 REFERENCES
`
`Albanese, A., Bussey, H., Weinstein, S. and Wolff, R. (1991) "A Multi-Network Research
`Testbed for Multimedia Communications Services," IEEE International Conference on
`Communications, Volume 1, 3.3.1-6.
`
`Aris Multimedia Entertainment, Inc., video sequence statistics derived from apl/lnc2.mpg video
`on Worldview CD-ROM.
`
`Heeke, H. (1993) "A Traffic-Control Algorithm for ATM Networks," IEEE Transactions on
`Circuits and Systems for Video Technology, Volume 3, Number 3, 182-9.
`·
`
`ISO Committee 11172 (1993) "Coding of moving pictures and associated audio for digital
`storage media at up to 1.5 Mbit/s," ISOIIEC JTC 1/SC 29.
`
`Netflix 1033 - Page 11
`
`
`
`180
`
`Part Five Multimedia over ATM
`
`North Valley Research, video sequence statistics derived from hockeyl.sif.variable video.
`
`Pancha, P. and El Zarki, M. (1993) "Bandwidth-Allocation Schemes for Variable-Bit-Rate
`MPEG Sources in ATM Networks," IEEE Transactions on Circuits and Systems for Video
`Technology, Volume 3, Number 3, 190-8.
`
`Pancha, P. and El Zarki, M. (1994) "MPEG Coding For Variable Bit Rate Transmission," IEEE
`Communications, Volume 32, Number 5, 54-66.
`
`Shanmugan, K. S. and Breipohl, A.M. (1988) Random Signals- Detection. Estimation. and
`Data Analysis, Wiley, 250, 316, and 599.
`
`Terry, J. (1992) "Alternative Technologies and Delivery Systems for Broadband ISDN Access,"
`IEEE Communications Magazine, Volume 30, Number 8, 58-64,.
`
`7 BIOGRAPHIES
`
`Olen L. Stokes received the B.S., M.S., and PhD. degrees from North Carolina State University
`in 1974, 1981, and 1995, respectively. He joined the IBM Corporation in 1974 working on
`communications controllers. From 1978 to 1984, he developed both hardware and microcode for
`the 3687 supermarket scanner, which was the first industrial user of holography. From 1985 to
`1989 he was a manager in IBM's Systems Network Architecture group. He then began working
`in the development of high speed communications products. He is currently a Senior Engineer
`in the IBM Networking Hardware Division in Research Triangle Park, N.C.
`
`Arne A. Nilsson received the M.S. and Ph.D. degrees from Lund Institute of Technology, Lund,
`Sweden in 1968 and 1976, respectively. He is currently a Professor with the Department of
`Electrical and Computer Engineering at North Carolina State University, Raleigh, North
`Carolina. In 1988 he became Director of the Center for Advanced Computing and
`Communications (formerly the Center for Communications and Signal Processing) and is now
`the Technical Director of that Center. He has published more than 70 papers in technical journals
`and conferences and has consulted for many government agencies and corporations.
`
`Netflix 1033 - Page 12