A NEW TRANSMISSION STRATEGY FOR SCALABLE MULTIMEDIA DATA ON OFDM SYSTEMS

Transcription

1 15th European Sgnal Processng Conference (EUSIPCO 27), Poznan, Poland, September 3-7, 27, copyrght by EURASIP A NEW TRANSMISSION STRATEGY FOR SCALABLE MULTIMEDIA DATA ON OFDM SYSTEMS Heykel Houas, Cléo Baras and Inbar Fjalkow ETIS ENSEA/UCP/CNRS UMR-851, 9514 Cergy-Pontose, France heykel.houas@ensea.fr, cleo.baras@ensea.fr, nbar.fjalkow@ensea.fr ABSTRACT We consder the real tme transmsson of scalable multmeda data wth a constraned end-to-end Qualty of Servce. In the Raylegh channel case, the resultng optmzaton problem s solved by the Flexble Transmsson (FT) algorthm. Ths resources allocaton strategy allows to transmt the sole layers wth guaranteed QoS at the recever. Here, we propose to extend the FT method to the OFDM case. The resultng algorthm s ncorporated nto a multmeda applcaton usng MPEG-4 speech frames, n order to compare the performance of our approach wth those of other unequal error protecton schemes. Smulaton results show the effcency of our soluton and ts adaptvty to frequency channel selectvty. 1. INTRODUCTION The growng demand on multmeda content delvery has renforced the necessty of effcent transmsson systems dedcated to multmeda data. These systems have to comply wth the dversfcaton of the transmsson networks, and more partculary the emergence of new wreless and xdsl communcaton technologes, and the development of new multmeda scalable codng tools. Ther desgn s related to the major constrant of end-to-end Qualty-of-Servce (QoS). Ths QoS reflects the dstorton ntroduced on the transmtted data by the source codng and the transmsson operatons. It s defned by the applcaton (streamng,...) wth respect to the source codng algorthm and the recepton devce (moble phone, PC...). Thus, our research study context 1 s focused on the development of an effcent multmeda transmsson scheme that ensures the QoS for any source codng tools and applcaton scenaros. Regardng State-of-The-Art strateges, ncludng jont sourcechannel codng [2], the desgn of such a system requres the adaptaton of each aspect of the transmsson scheme (from source codng algorthm to resources allocaton polcy) wth respect to the channel condton and the recepton devce. Frst, the source codng algorthm choce can be smplfed when the source encodng technque s scalable. Indeed, the btstream frames are splt nto layers, correspondng to the compressed verson of the same multmeda data wth dfferent rates and dfferent source codng qualtes. Therefore, the source dstorton can be drectly related to the number of transmtted layers, wthout changng the source codng algorthm. Furthermore, ths scalable structure establshes a herarchcal organzaton of the emtted data nto senstvty classes to channel transmsson errors 2. Ths error senstvty, known to be unequal [3], requres to adapt (even optmze) the transmsson resources wth respect to the emtted data senstvty, channel condtons and system constrants, yeldng Unequal Error Protecton (UEP) schemes. State-of-The-Art transmsson strateges have two major drawbacks n our applcaton scenaro. Frst, most strateges [4] do not take advantage of the degree of freedom offered by the scalable encodng process on the source rate adaptaton. Second, most resources allocaton procedures [3, 5] are based on the mnmzaton of an emprcal dstorton metrc over possble resources. Ths metrc s 1 Ths study s part of the FP6/IST project M-Ppe [1] and s co-funded by the European Commsson. 2 We assume that the source decoder s not robust (as t s the case for most exstng standards). often tabulated wth respect to a consdered scalable encoder for varous transmsson condtons. Nevertheless, t s a complex task, whch can become costfull when the number of codng tools supported by the transmsson scheme ncreases. Moreover, the consdered optmzaton crteron s not relevant n our research study context, where the end-to-end QoS s specfed by the applcaton and belongs to the system constrants. In [6], we propose a resources allocaton polcy, called Flexble Transmsson (FT), sutable for ths applcaton context n the Raylegh channel case. It s based on: (1) an orgnal optmzaton crteron, that maxmzes the source rate (.e. the number of transmtted layers) under QoS and load constrants and (2) a smple expresson of QoS as Bt Error Rates (BER) requred by each layer [4]. In ths paper, we propose an extenson of ths algorthm to the Orthogonal Frequency Dvson Multplexng (OFDM) context, takng nto account the scalable property of the transmtted data and the specfctes of the OFDM channel. The adapted transmsson parameters are chosen wth respect to the current channel state nformaton and the system constrants, provded by a cross-layer strategy. The paper s organzed as follows. Secton 2 descrbes the transmsson system. Secton 3 sums up the FT resources allocaton polcy desgned for Raylegh channels and detals ts expanson to the OFDM case, yeldng the FT-OFDM procedure. Secton 4 draws up the algorthm effcency through an applcaton example (MPEG-4 audo streamng for telephony over ADSL). Smulaton results are presented and compared to standard protecton schemes. Secton 5 concludes by summng up the major contrbutons of the proposed method. 2. REFERENCE TRANSMISSION SYSTEM DESCRIPTION The reference transmsson system s presented n fgure 1 and s detaled n the followng subsectons. 2.1 Source codng parameters Thanks to the scalable encodng process, multmeda data are represented by a btstream structured n frames or Transmsson Unts (TU). These frames are splt nto I max layers: one base layer and I max 1 enhancement layers, denoted by {L } 1,Imax ] wth decreasng mportance degrees. The length of the -th layer (n bts) s equal to a proporton p of the total frame length, denoted by N, so that: I max p = 1. The mportance degree affected to each layer characterzes [2]: the weght of each layer on the dstorton D s ntroduced by the source encodng process on the reconstructed multmeda data: the more enhancement layers are used durng the source decodng, the smaller the source dstorton s. the unequal senstvty of each layer to channel transmsson errors and therefore ther weght on the dstorton D c ntroduced by the channel on the decoded multmeda data. As proposed by [4], we suppose that the error senstvty of each layer can be featured by a bt error probablty value, denoted by B. Ths value s defned accordng to some perceptual qualty crteron so that, when the BER affectng the transmsson of the -th 27 EURASIP 2194

2 15th European Sgnal Processng Conference (EUSIPCO 27), Poznan, Poland, September 3-7, 27, copyrght by EURASIP Resources allocaton algorthm Feedback nformaton { g(n) 2 Es }n [1,N c] Channel estmaton Emtted multmeda data Source encoder m(i FT ) = {m } [1,IFT ] I FT R(IFT ) = {R } [1,IFT ] L L L L 1 L 1 L 1 L 1 Channel OFDM Channel encoder CHANNEL decodng L Imax EMITTER IFT IFT RECEIVER IFT Source decoder Receved multmeda data Fgure 1: Transmsson scheme ploted by the FT algorthm. layer s lower than B, the source decodng of ths layer has no (or few) nfluence on the channel dstorton D c. The set of {B } [1,Imax ] s then drectly related wth the end-to-end QoS of transmsson systems: ensurng the QoS conssts n ensurng the BER affectng each layer, denoted by BER, to be lower than B. 2.2 Transmsson model Avalable resources at the emtter At the emtter, source data are frst encoded wth a channel codng, chosen among rate-compatble channel encoder [7]. In ths paper, we use Rate Compatble Punctured Convolutonal (RCPC) codes, but t could easly be extended to any other rate-compatble encodng process (such as turbo codes). Thanks to the puncturng process, usng a mother convolutonal codng rate R 1 = N 1 1, dfferent codng rates R l are avalable: R l = P+l P, wth 1 l (P 1)N 1, where P s the puncturng perod. The coded data are then mapped wth one of the avalable sgnallng constellatons. These constellatons are BPSK or 2 2m -QAM wth constant symbol energy E s. The order m denotes the number of bts per symbol and wll refer to a specfc modulaton choce (m = 1 for BPSK, m = 2 for QPSK, etc...) OFDM channel Thanks to a long enough cyclc prefx, the frequency selectve channel s turned nto scalar channel gans {g n } n [1,Nc ] on the N c dfferent OFDM carrers: the Sgnal-to-Nose-Rato (SNR) for the n-th sub-channel s g n 2 E s. Assumng that the channel vares slowly, we defne T as the number of perods whle channel tme response s constant. The maxmum symbol load per coded TU, S max, s chosen equal to T N c and s defned at the physcal layer. It refers to the physcal payload. Moreover, we assume that our OFDM system s synchronzaton error free Recever At the recever, after Zero-Forcng equalzaton, a soft output maxmum a posteror demapper s used to compute soft decson values, appled as nput of a classcal Vterb decoder. A channel estmaton module s lnked wth the emtter through a feed-back channel { so to nform hm of the estmated channel condtons (.e. g n 2 E s ). The feedback channel and the channel estmaton at the recever are supposed perfect and the btstream header }n [1,N c ] transmsson s assumed to be error-free. 3. THE PROPOSED RESOURCES ALLOCATION PROCEDURE Regardng the system characterstcs, the resources allocaton procedure ams at choosng the adapted system parameters that ensures the QoS constrants. The system parameters are the number of transmtted layers, denoted by I FT, the modulaton order m and the codng rate R for each transmtted layer L wth [1,I FT ]. Furthermore, the QoS constrants are stated by the {B } [1,Imax ] bounds. The proposed strategy s based on the key dea that the enhancement layers should not be transmtted whle the prevous transmtted layers are not suffcently protected to respect the QoS bounds. Thus, the proposed allocaton polcy yelds to the adaptaton of the number of transmtted layers I FT at the channel codng stage. Wth scalable data, the I FT parameter s drectly related to the source rate R s (I FT ): R s (I FT ) = p N. (1) As a result, the allocaton strategy s related to an optmzaton problem: maxmzng the source rate (or I FT ) and fndng the assocated modulaton orders m(i FT ) = {m } [1,IFT ] and channel codng rates R(I FT ) = {R } [1,IFT ] over the possble transmsson parameters combnatons, denoted by P, wth respect to QoS bounds, data throughput constrants (gven by the symbol rate S max ) and channel transmsson condtons. In ths secton, we brefly remnd the FT resources allocaton procedure desgned to solve ths optmzaton problem n the Raylegh channel case and expound ts adaptaton to the OFDM context. I FT 3.1 FT algorthm: Raylegh channel case In the Raylegh channel case, the channel condtons are gven through an unque E s value. Snce the bt error rate of the -th layer s a functon of the SNR, the modulaton ( order ) m and the codng rate R, we wll denote t: BER Es,m,R. Moreover, we defne S (m,r ) = p N m R, the spectral effcency of L (that s the number of symbols transmtted wthn the coded L ). The optmzaton problem related to the allocaton polcy can now be set by the followng equaton, gven the SNR value E s : (I FT,m(I FT ),R(I FT )) = arg max (I,m(I),R(I)) P R s(i), (2a) [1,I],BER( E s,m,r ) B (2b) I S TU = S (m,r ) S max (2c) where S TU represents the transmsson load of the whole transmtted data (.e. the number of symbols used to transmt the coded layers L 1 to L IFT ). 27 EURASIP 2195

3 15th European Sgnal Processng Conference (EUSIPCO 27), Poznan, Poland, September 3-7, 27, copyrght by EURASIP The FT soluton, we brng to ths problem n [6], s an teratve algorthm. It processes from the frst layer L 1 to the last one by progressvely fllng the coded TU untl the maxmum symbol load S max s reached. Durng the process of the I-th layer L I, the algorthm selects the modulaton and codng rate par (m I,R I ) that mnmzes the spectral effcency S I of L I over the possble pars whch respect the BER constrant (2b). The matched number of transmtted classes s obtaned as soon as I S > S max : ndeed I f lex = I Adaptaton to the OFDM case We now propose to extend the FT algorthm to the OFDM context. The resources allocaton procedure has to defne, besdes the matched number of transmtted layers and the matched transmsson parameters, an adapted carrer allocaton polcy. The allocaton polcy has two purposes: Take the channel state nto account: ths state s now gven through multple SNR values on the dfferent sub-carrers, that s { g n 2 E s. Ths SNR varablty prevents us from d- }n [1,N c ] rectly applyng the FT algorthm, snce t has been desgned for an unque SNR nput value. Usng the FT algorthm wth the average SNR doesn t satsfy the QoS on the carrers wth lower SNR than the average and usng the algorthm wth the mnmal SNR value results n a severe overprotecton (yeldng a waste of resources). State the assgnment between OFDM sub-channels, layers and resources as presented n fgure 2: addtonal parameters have therefore to be consdered by the allocaton polcy, such as the set of OFDM sub-channels, denoted by N, used to transmt each layer L wth modulaton m and codng rate R Adapted SNR values choce to process FT The OFDM SNR varablty modfes the QoS constrants formulaton (2b). Indeed, the BER affectng the -th layer transmsson now depends on the BER values characterzng each OFDM subchannels { } assgned to L. We wll denote these BER values by BER () n. They can be expressed wth respect to the SNR { n N values g n 2 E s }n [1,N, the modulaton order m and the codng c ] rate R. For the -th layer, the QoS constrant gven by equaton (2b) can be smply ensured by consderng the worst BER confguraton: ths confguraton s acheved by the OFDM sub-channel assgned to L wth the lowest SNR or channel gan (.e. mn g n 2 ). The QoS n N constrants wll then be stated by the followng overestmaton: [1,I],BER BER () n mn () B wth n mn () = arg mn g n 2 n N (3) Ths overestmaton, sometmes too severe, can be relaxed consderng that only a percentage α (for nstance 8%) of the layer L comply wth the QoS requrement. Thus, denotng N (α) the set of OFDM sub-channels used to transmt the α p N frst bts of L, the QoS constrants wll fnally be formulated consderng the worst BER confguraton on the N (α) OFDM sub-channels dstrbuton, that s: [1,I],BER BER () B n (α) mn () wth n (α) mn Sub-carrers allocaton polcy () = arg mn n N (α) g n 2 The OFDM SNR dversty can also be exploted to enforce the unequal protecton of the data. State-Of-The-Art resources allocaton algorthms [4] propose to sort the sub-channels n a decreasng SNR order: thus, the best sub-channels (n terms of SNR) (4) T SNR L 1 Tme... L L If lex N 1 N 2 N If lex Unallocated sub-channels Frequency Fgure 2: Sub-channels allocaton polcy. can be allocated to the most mportant layers, whle the least one are used to transmt the least mportant layers. We propose to follow the same allocaton polcy n the extended FT algorthm. Supposng the OFDM sub-channels sorted wth decreasng SNR value, the layer L 1 coded wth ts assocated protecton scheme (R 1 (I FT ),m 1 (I FT )) wll be transmtted [ over the ] best SNR subchannels, whose ndex belongs to N 1 = 1, p1 N m 1 R 1 T. On the other sde, L IFT coded wth (R IFT,m IFT ) wll be transmtted wth the least SNR ones, whose ndex only depend on the prevous layers mappng Problem formulaton and soluton The FT procedure n the OFDM context deals wth fndng the adapted number of transmtted layers I FT, the adapted modulaton order m(i FT ), codng rates R(I FT ) and sub-carrers dstrbuton N (I FT ) = {N } [1,IFT ] (startng from the best SNR sub-channels to the least ones) over the possble confguratons set P that satsfy QoS and system constrants. It can now be stated by the followng optmzaton problem: (I FT,m(I FT ),R(I FT ),N (I FT )) = arg max (I,m(I),R(I),N(I)) P R s(i)(5a) [1,I],BER () B, wth n (α) n (α) mn () mn I S TU = S (m,r ) S max () = arg mn n N (α) g(n) 2 (5b) (5c) The soluton to ths problem, called FT-OFDM, follows the same prncples as the teratve FT algorthm desgn for the Raylegh channel case. It processes from the frst layer L 1 to the last one by progressvely fllng the coded TU untl the maxmum symbol load S max s reached. Durng the process of the I-th layer L I, the algorthm computes for each possble confguraton (m,r ) the requred sub-channels dstrbuton N I and the assocated subset N (α) I (wth respect to the sub-channels dstrbutons of the prevous layers). Then, the targeted SNR value (mn (α) n N g(n) 2 E s / ) over the consdered N I sub-channels s computed to evaluate the respect of the BER constrant defned by equaton (4). The (m I,R I,N I ) satsfyng the QoS constrant and mnmzng the spectral effcency S I of L I over the possble trplets s fnally selected. The matched number of transmtted classes s obtaned as soon as I S > S max, wth I FT = I EURASIP 2196

4 15th European Sgnal Processng Conference (EUSIPCO 27), Poznan, Poland, September 3-7, 27, copyrght by EURASIP 4. EXPERIMENTAL RESULTS 4.1 Applcaton to MPEG-4 speech scalable data An applcaton of the proposed algorthm to the transmsson of scalable speech data s at stake to evaluate the algorthm performance. Among the several source codng tools of the MPEG-4 standard [8], we focus on the CELP encoder wth the MultPulse Exctaton and the Bt-Rate Scalablty (BRS) tools. Ths coder can compress a speech sgnal sampled at 8 khz wth a scalable btstream characterzed by a 12 kbps btrate and a 4-layers structure. The base layer (wth 12 bts per frame) s generated wth a core CELP encoder operatng at 6 kbps usng a speech producton model based on an exctaton sgnal passed through an auto-regressve flter. The BRS tool provdes the 3 remanng layers and adds a 2 kbps per layer nformaton, that refnes the exctaton sgnal descrpton (wth 4 bts per frame and per layer). The whole btstream s fnally constructed as the successon of frames, each contanng N = 24 bts. 4.2 Test plan Audo scalable data parameters Accordng to the chosen scalable coder structure, the layer proportons are the followng: p 1 = 1 2, p 2 = p 3 = p 4 = 1 6. We assume that the QoS to ensure for our applcaton can be descrbed by the followng BER upper bound of each class: B 1 = 3.1 3, B 2 = 4,6.1 3, B 3 = 8.1 3, B 4 = Standard lnk parameters RCPC codes are generated from a mother convolutonal 1 wth rate 3 and enumerator polynoms G 1 = [133] 8,G 2 = [145] 8 and G 3 = [175] 8. The puncturng perod s chosen equal{ to 8. The lst of avalable code rates R s fnally: 8 R = 9, 8 1, 8 12, 8 14, 8 16, 8 18, 8 2, 8 22, 8 }. 24 The modulaton schemes choce are lmted to BPSK and 2 2m - QAM wth m {2,3,4} (QPSK to 64-QAM) OFDM parameters The OFDM parameters are the followng: the number of subchannels n an OFDM symbol s fxed to N c = 12, the OFDM coherence tme s T = 3, yeldng a maxmum symbol load per coded TU S max = T N c = Performance evaluaton The FT-OFDM algorthm effcency for the transmsson of speech scalable data s evaluated by the perceved qualty measurement of the decoded speech sgnal. Ths measure s computed usng the Perceptual Evaluaton of Speech Qualty (PESQ) software, descrbed by the IUT recommendaton P.862 [9]. It compares the uncoded orgnal speech sgnal to the receved speech sgnal and results n a Mean Opnon Score (MOS), n the range [;4], that reflects the dstorton ntroduced by the source codng and transmsson operatons on the decoded data. Ths measure s computed usng a 1s-duraton sgnal, sampled at 8 khz, and s averaged over 5 transmssons. The performance of the FT-OFDM algorthm wll be compared to three State-of-the-Art resources allocaton strateges, summed up n table 1: these strateges acheve data protecton aganst error by applyng dfferent codng rates and modulaton orders on each layer. Followng the ordered subcarrer selecton algorthm proposed by [4] n the OFDM channel context to mnmze the average error probablty affectng each layer transmsson, the modulaton order and the assocated OFDM sub-channel dstrbuton are chosen as follows: the layers are dstrbuted on the sub-channel regardng ther transmsson error senstvtes (the hgher senstve the layer s to errors transmsson, the better sub-channel SNR s used). Modulaton orders are chosen so that the whole coded frame data only fll the half OFDM sub-channels (OFDM sub-channels wth the lowest SNR are therefore dscarded). Scheme Layer L 1 Layer L 2 Layer L 3 Layer L 4 Strategy 1 R 1 = 8/24 R 2 = 8/24 R 3 = 8/24 R 4 = 8/24 16-QAM 16-QAM 16-QAM 16-QAM Strategy 2 R 1 = 8/18 R 2 = 8/12 R 3 = 8/12 R 4 = 8/12 QPSK 16-QAM 16-QAM 16-QAM Strategy 3 R 1 = 8/16 R 2 = 8/22 R 3 = 8/16 R 4 = 8/9 QPSK 16-QAM 16-QAM 16-QAM g(n) 2 E s / (db) Table 1: Consdered standard resources allocaton strateges (a) Carrers ndex PESQ s MOS (b) Strategy 1.5 Strategy 2 Strategy 3 FT OFDM mean SNR (db) Fgure 3: (a) Frequency response of the consdered ADSL channel, (b) Performances of the dfferent resource allocaton schemes wth respect to the ADSL channel mean SNR. 4.3 Obtaned results Performance n terms of perceved qualty Wth am at evaluatng the performance of the FT-OFDM algorthm on the perceved qualty of the decoded data, we use an ADSL channel model, measure through expermental smulatons on real ADSL channel. The frequency response of the consdered ADSL channel model s presented n fgure 3 (a). The mean MOS obtaned wth the PESQ algorthm after 5 transmssons of a speech sgnal are drawn n fgure 3 (b) for the three reference resources allocaton strateges and the FT-OFDM algorthm wth respect to the ADSL channel mean SNR value. These curves prove the effcency of our resource allocaton procedure: the FT-OFDM algorthm ponts out that the QoS can be guarantee untl a 7 db mean SNR. In ths operatng range, the perceved qualty of the decoded speech data obtaned wth FT-OFDM slowly vares n comparson to the reference schemes. Moreover, the ADSL channel mean SNR value, from whch the maxmum MOS value (obtaned when the speech data are only dstorted by the source codng process) s reached, s 7 db lower than those of the State-Of- The-Art strateges. Ths mprovement s the result of the FT-OFDM adaptvty on channel condtons contrary to the reference schemes. For mean SNR values lower than 7 db, the FT-OFDM algorthm ponts out that the QoS can not be guarantee and selects a resources confguraton by default. Snce the QoS constrants are no more satsfed, the FT-OFDM MOS values roughly decrease and reach smlar performance to those of the reference schemes (due to the sub-carrers dstrbuton) Resources allocaton results wth respect to channel dynamc Wth am at evaluatng the nfluence of OFDM channel (and more partculary ts frequency selectvty) on the FT algorthm, we desgn an arbtrary smulaton channel wth the followng { prncples: gven a mean SNR value denoted by E s, the SNRs g(n) 2 E s }n [1,f c ] each OFDM sub-channel are computed n order to follow a lnear 27 EURASIP 2197

5 15th European Sgnal Processng Conference (EUSIPCO 27), Poznan, Poland, September 3-7, 27, copyrght by EURASIP I FT S TU /S max (%) Dynamc (db) Mean SNR (db) Mean SNR (db) Dynamc (db) Fgure 4: Varaton of the number of transmtted layer I FT wth respect to the mean SNR and the dynamc values of the OFDM channel (wth α = 1%). Fgure 5: Allocaton rato of an OFDM symbol wth respect to the mean SNR and the dynamc values of the OFDM channel (wth α = 1%). decrease (n db) around the mean SNR value. In other words: g(n) 2 E s = E ( ) s (db) + n + 1 wth n [1,N c ], (db) 2N c where represents the dfference between the best sub-channel SNR and the least one. In ths case, reflects the frequency selectvty of the OFDM channel and wll be referred to as channel dynamc. Fgure 4 and 5 clarfy the FT-OFDM algorthm process wth respect to the OFDM channel. Fgure 4 presents the codng source rate (n terms of number of transmtted layer I FT ), computed by the FT-OFDM algorthm, wth respect to the mean SNR and the dynamc values. Fgure 5 draws the varaton of the allocaton rato of an OFDM symbol (.e. the rato between the number of symbols S TU used to transmt the coded data and the symbol load S max ) wth the OFDM channel (mean SNR and dynamc). Fgure 4 exhbts the strong adaptaton of the codng source rate for medum mean SNR value (between 3 and 1 db) mposed by the QoS requrements. Moreover, the algorthm shows a strong adaptvty of the source rate to the channel dynamc: n ths range, the source rate decreases when the channel dynamc ncreases. Indeed, the proporton of low SNR sub-carrers (compared to mean SNR value) ncreases wth the dynamc and the QoS constrants become more restrctng, snce they are related to the lowest SNR value of the allocated OFDM sub-channel. Wth hgh mean SNR values (greater than 1 db), source codng rate s no more adapted (I FT = 4), snce the channel condtons are good enough to transmt all layers. The FT algorthm adaptvty to the OFDM channel s now deferred to the sub-channel allocaton polcy. Indeed, fgure 5 shows that the allocaton rato of an OFDM symbol decreases wth the mean SNR value ncrease. For hgh mean SNR confguratons, layers transmsson s concentrated on OFDM sub-channels wth hghest SNR. These sub-channels, thanks ther hgh SNR, are able to support the layers transmsson wth hgh modulaton orders (generally 64-QAM) and low codng rates and stll respect the QoS constrants. Fnally these fgures exhbt the system operatng lmts under the end-to-end QoS constrants: when the mean value SNR s lower than -3 db, any QoS bounds can be ensured. Therefore, the FT algorthm decdes to transmt any scalable data (by ndcatng I FT = ). These lmts could be mproved by relaxng the constrants (QoS or symbol load). 5. CONCLUSIONS In ths paper, we proposed a resources allocaton procedure for scalable multmeda data transmsson over a frequency selectve channel usng an OFDM system. Our soluton allows not to transmt all layers n order to protect enough the transmtted ones, usng a channel-adapted carrers allocaton. The choce of the transmtted layers, yeldng the source codng rate adaptaton, s no more performed by the source encodng process but by the channel encoder. Smulaton results show the effcency of the proposed source codng rate adaptaton. It also exhbt good performance of the proposed algorthm n terms of QoS delvery and robustness to frequency channel selectvty. Thus, the proposed resources allocaton strategy seems to be sutable for desgnng wreless or xdsl communcaton systems that may : (1) be strongly constraned by channel varatons (outdoor, ndoor channels...), (2) requre flexblty and low complexty (especally for multcast applcatons). REFERENCES [1] S. Hakansson, M-Ppe, [2] G. Cheung and A. Zakhor, Bt allocaton for jont source/channel codng of scalable vdeo, IEEE Transactons on Image Processng, vol. 9, no. 3, pp , March 2. [3] C. Lamy-Bergot, N. Chautru, and C. Bergeron, Unequal error protecton for H btstreams over a wreless IP network, n ICASSP, vol. 5, Toulouse, France, May 26, pp [4] D. Dardar, M. G. Martn, M. Mazzott, and M. Chan, Layered vdeo transmsson on adaptve OFDM wreless systems, EURASIP Journal on Appled Sgnal Processng, vol. 4, no. 1, pp , 24. [5] H. Zheng and K. Ray Lu, Robust mage and vdeo transmsson over spectrally shaped channels usng mutlcarrer modulaton, IEEE Transactons on Multmeda, vol. 1, no. 1, pp , [6] H. Houas, C. Baras, and I. Fjalkow, Resources allocaton optmzaton for scalable multmeda data subject to qualty of servce contrants, n SPAWC, july 26. [7] J. Hagenauer, Rate-compatble punctured convolutonal codes (RCPC codes) and ther applcaton, IEEE Transactons on Communcatons, vol. 36, no. 4, pp , Aprl [8] F. Perera and T. Ebrahm, The MPEG-4 book. IMSC Press Multmeda Seres, Prentce Hall PTR, may 22. [9] IUT-T Recommendaton P. 862, Perceptual evaluaton of speech qualty (PESQ): An objectve method for end-to-end speech qualty assessment of narrow-band telephone networks and speech codecs, EURASIP 2198