Delay Constrained Fuzzy Rate Control for Video Streaming over DVB-H

Transcription

1 Delay Constraned Fuzzy Rate Control for Vdeo Streamng over DVB-H Mehd Rezae, Moncef Gabbouj, 3 med Bouazz,, Tampere Unversty of Technology, 3 Noka Research Center, Tampere, Fnland mehd.rezae@tut.f, moncef.gabbouj@tut.f, med.bouazz@noka.com Abstract A novel fuzzy vdeo rate control algorthm (RCA) wth delay and qualty constrants s proposed for vdeo streamng over DVB-H (Dgtal Vdeo Broadcastng for Handheld termnals) applcaton. n ths applcaton a varable bt rate (VBR) bt stream s used to mprove the qualty and compresson performance of encoded vdeo at the expense of bufferng delay whle there s a holdup as channel changng delay. DVB-H uses a tme-slced transmsson scheme to reduce the power consumpton for the recever. The tme-slce scheme amplfes the channel changng delay. Two man factors n channel changng delay are the tme untl a meda decoder s refreshed by a random access pont such as an DR (nstantaneous Decoder Refresh) pcture n H.64/AVC and the delay to compensate the varaton n bt rate. The average decoder refresh delay can decrease by frequent DR pctures n the bt stream whle t drops the qualty and compresson effcency remarkably. On the other hand, a VBR vdeo provdes hgher qualty at the expense of hgher bufferng delay. Hence force, there s a multlateral contradcton between the qualty and delay n ths applcaton. The proposed RCA has been optmzed to provde hgh qualty VBR vdeo wth mnmum channel changng delay. Smulaton results on H.64/AVC vdeo codng standard show that encoded bt streams by proposed RCA strctly obey the delay and qualty constrants.. ntroducton Dgtal Vdeo Broadcastng for Handheld termnals (DVB-H) s an ETS standard specfcaton for brngng broadcast servces to handheld recevers []. DVB-H s manly based on the DVB-T specfcaton for dgtal terrestral televson, addng to t a number of features desgned to take nto account the lmted battery lfe of small handheld devces and the partcular envronments n whch such recevers must operate. To reduce the power consumpton n handheld termnals, the servce data s tme-slced and then t s sent nto the channel as bursts at a sgnfcantly hgher bt rate compared to the bt rate of audo-vsual servce. Tme-slcng enables a recever to stay actve only a fracton of the tme (Burst), whle recevng bursts of a requested servce []. Channel changng delay n DVB-H refers to the tme between the channel swtchng and the start of the meda renderng. The channel changng delay for newly-joned recpents conssts of several parts ncludng: delay untl the start of the desred tme-slce or burst, recepton duraton of a complete burst, delay to compensate the sze varaton of bursts, delay to compensate the varaton of bt rate, delay to compensate the synchronzaton between the assocated streams of the streamng sesson and delay untl a meda decoder s refreshed by a random access pont to produce correct output samples. One of the crtcal factors n channel changng delay s the tme untl a meda decoder s refreshed to produce correct output frames, whch can be mnmzed f burst s started wth a random access pont such as an DR pcture. n DVB-H, the content encodng and the encapsulaton to bursts are mplemented ndependently and t s hard to algn the locaton of DR pctures to the burst boundares. The average decoder refresh delay can decrease by frequent DR pctures n the bt stream. Another man factor n channel changng delay s the requred ntal bufferng delay to compensate the varaton n bt rate. Whle n low delay vdeo communcatons bt streams wth constant bt rates are requred, n streamng applcaton more varaton n bt rate s acceptable. Normally, a varable bt rate RCA operates closer to the constant qualty optmum, whch results n a better average qualty. We revewed several varable RCAs and we proposed a RCA for moble applcatons n our prevous work [3]. Furthermore, we proposed a fuzzy RCA n [4] optmzed for hgh delay streamng applcaton. n ths paper a new RCA optmzed for streamng over DVB-H applcaton s proposed. roceedngs of the 6 nternatonal Conference on ntellgent nformaton Hdng and Multmeda Sgnal rocessng (H-MS'6) /6 $. 6

2 n vdeo streamng over DVB-H applcaton we need specal bt streams. Frst, we want to utlze the advantage of VBR vdeo to mprove the qualty and compresson performance. Second, frequent DR pctures are desred n the bt stream to decrease the decoder refresh delay. Whle n a smlar qualty an DR pcture consumes a bt budget from 5 to tmes more than an nter-predcton () pcture, usng frequent DR pctures drops the compresson effcency and qualty remarkably. f we encode the DR pctures wth a qualty smlar to pctures to use the advantage of VBR vdeo, t needs a hgh ntal bufferng delay to compensate the varaton n bt rate. On the other hand, f the DR pctures are encoded wth a qualty much lower than pctures, we do not use the advantage of varable bt rate vdeo. Hence force, there s a strong multlateral contradcton between qualty and delay n ths applcaton. We need to fnd an approprate operatng pont for the vdeo RCA n VBR between constant qualty and constant bt rate to optmze the average qualty and delay. Consderng the qualty and delay constrants, a fuzzy RCA for vdeo streamng over DVB-H applcaton s proposed n ths paper. The proposed RCA has been optmzed such that t utlzes the advantage of VBR and also t mnmzes the channel changng delay. Smulaton results on H.64/AVC vdeo codng standard show that the proposed RCA can provde VBR vdeo wth a hgh level average qualty and wth small delay. Secton of ths paper presents detals of proposed RCA. Smulaton results are provded n Secton 3. The paper s closed wth conclusons n secton 4.. roposed Rate Control Algorthm The proposed RCA has been optmzed to provde VBR bt streams ncludng frequent DR pctures wth a relatve constant vsual qualty and low average bufferng delay. Moreover, t has a small degree of complexty from the processng power pont of vew. The proposed RCA controls the bt rate by adjustng the quantzaton parameter (Q) on pcture base. t utlzes a relatve small sze vrtual buffer, a fuzzy controller and several other tools to calculate the Q for dfferent types of pctures. Whle frequent DR pctures exst n the bt stream and the sze of vrtual buffer s relatve small, the bt allocaton to DR pctures has a remarkable mpact on the encodng results so the Q of DR pctures should be computed carefully. Many theoretc and expermental results have been used n bt allocaton to dfferent vdeo pctures. The key concept s to mnmze the fluctuaton n Q and thereafter n qualty whle a target average (on all random access ponts) bufferng delay should be acheved. The proposed RCA s dvded n two man parts. The frst part utlzes a fuzzy controller to compute the Q for pctures and the second part uses more other tools to calculate the Q of DR pctures. The DR pctures at the scene cuts are handled dfferently from the normal predefned frequent DR pctures. Detals of proposed RCA are presented below... Q of nter-redcton ctures The Q for a pcture s defned by a fuzzy rate control system. Fgure. depcts the block dagram of proposed fuzzy rate control system. The fuzzy controller and vrtual buffer are the basc parts of the proposed rate controller. Furthermore, a low pass flter (LF) smoothes the feedback sgnal from the rate to the fuzzy controller. Moreover, a gan control block (G) tunes the gan of feedback loop adaptvely accordng to the target delay as: G = GC TR / BS, () where TR and BS denote the target bt rate and the buffer sze respectvely. G C s a constant coeffcent typcally about.5. The Q for the current pcture ( Q ) s sum of the output of fuzzy controller ( Q ) and F the Q used for encodng prevous frame ( Q ): Q ( ) = Q( ) + QF ( ). () From the controllng pont of vew the man part of computed Q for frames s a delayed verson of prevous used Q and the control or varaton of Q s provded by the fuzzy controller. The vrtual buffer used by controller smulates the bufferng process of the decoder n the recevng sde of streamng and t nearly compromses wth hypothetcal reference decoder (HRD) models used n dfferent vdeo codng standards. The occupancy of buffer s updated after encodng each vdeo frame as: BF( + ) = BF ( ) B ( ) + ( TR/ FR), (3) Uncompressed Vdeo Vdeo Encoder Q Q F Q Compressed Vdeo Delay Q( ) G Fuzzy Controller Rate LF B Vrtual Buffer To Decoder Buffer Fullness Fgure. Block dagram of the fuzzy vdeo RCA roceedngs of the 6 nternatonal Conference on ntellgent nformaton Hdng and Multmeda Sgnal rocessng (H-MS'6) /6 $. 6

3 Where BF denotes the buffer fullness and B shows the number of bts consumed by the prevous encoded or DR frame. FR ndcates the frame rate. The sze of vrtual buffer s one the most mportant parameters and t s defned accordng to the target average bufferng delay on all random access ponts whch s set by user. The fuzzy controller has been optmzed such that the average buffer fullness durng operatng tme s about 6% to 65% of buffer sze. Therefore, the sze of vrtual buffer s computed as: BS = TR TD/.65, (4) where BS denotes the sze of vrtual buffer, TD stands for the target average delay. More detals about the fuzzy controller are presented n the sequel.... Fuzzy Controller The fuzzy controller has basc modfcatons relatve to proposed fuzzy RCA n [4]. The fuzzy controller has two nput sgnals as: x = BF / BS, (5) B FR RX x = +, (6) TR DR where B denotes the consumed bt budget by the prevous encoded pcture. DR s the nterval of frequent DR pctures n the bt stream. RX ndcates the relatve codng complexty of DR pctures to pctures and t computed as: RX = B / B, (7) where B and B denotes the average consumed bt budgets by the all encoded DR and pctures respectvely. f the prevous encoded pcture s an DR pcture, the value of B n (6) s set by the value of B / RX. Moreover, a smple low pass flter (LF) s used to smooth the varaton n B before nput to fuzzy controller. The mpulse response of used LF s: H ( z) = m /( m + z ), (8) where m s a constant value and good results are obtaned wth m=.. All the fuzzy rules are summarzed n the table. The content of table specfes the output of the controller. The letters H, L, M, V, X, U and S correspond to lngustc specfcatons of Hgh, Low, Medum, Very, Extremely, Ultra and Super. As an example from the table t can be expressed as: f ( x s VL and x s M) Then (Output s VH ). The nput sgnals are specfed by ther fuzzy membershp functons (MSF). Numbers of 7 and 9 MSFs for the two nputs x and x were consdered. The lngustc fuzzy rules and MSFs were desgned based on some theoretc and expermental results and TABLE. SUMMARZATON OF THE F-THEN FUZZY RULES VH SH UH VXH XH VVH VH H MH M H UH VXH XH VVH VH H MH M ML MH VXH XH VVH VH H MH M ML L x M XH VVH VH H MH M ML L VL ML VVH VH H MH M ML L VL VVL L VH H MH M ML L VL VVL XL VL H MH M ML L VL VVL XL VXL XL VVL VL L ML M MH H VH x provded experences n [3] and [4]. Furthermore, an optmzaton process was performed for fne tunng of the fuzzy MSFs. The dstrbutons of MSFs are shown n the Fgure. The desred central values for the output of fuzzy system correspond to VXL, XL, VVL, VL, L, ML, M, MH, H, VH, VVH, XH, VHX, UH and SH n the table are -6, -5, -4, -3, -, -,,,, 3, 4, 5, 6, 7 and 8 respectvely. We used a well-known fuzzy system wth two nputs usng roduct nference Engne, sngleton fuzzfer, and centre average defuzzfer whch s: f N N = = ( x, x ) = N N y = = µ ( x ). µ A A µ ( x ). µ A A ( x ( x ) ), (9) where f ( x, x ) denotes the output and N { A, A,..., A } are fuzzy sets wth =, { µ ( x A )} N and { µ ( x )} membershp functons defned for A N nputs x and x. The centre of output fuzzy set ( B ),denoted by y, s chosen as the output desred value. More nformaton about the fuzzy system () are presented n [5]... Q of DR ctures The Q of DR pcture s computed based on the frame complexty, target bt rate, target average delay, buffer fullness and scene cut nformaton. There are two types of DR pctures, the normal frequent DR pctures whch are placed n locatons wth a constant frequency and the DR pctures whch are placed at scene cuts. The Q of all DR pctures s computed as: Q = QR + AC ( S + B + D), () where Q denotes the Q of DR pcture. A C s a constant value n the range of (_) that s used for fne tunng of the rate control accordng to vdeo content propertes. A bgger A provdes more C aggressve control. Q s a reference value for the R Q whch s computed dfferently for the two types of roceedngs of the 6 nternatonal Conference on ntellgent nformaton Hdng and Multmeda Sgnal rocessng (H-MS'6) /6 $. 6

4 .5 Membershp Functons of Lngustc Varables XL VVL VL L ML M MH H VH budget n the meddle range even there s no correlaton between scenes nput: x.5.5 VL L ML M MH H VH nput: x Fgure. Membershp functon of the lngustc varables DR pctures. The man control of Q s mposed by three controllng sgnals ncludng S, B and D. The sgnals S and B adapt the Q accordng to the codng complexty of vdeo pcture and buffer fullness respectvely. Moreover, D adapts the Q accordng to the target average delay. Whle the reference Q defnes a reference value for the Q, the controllng sgnals make some varaton around the reference value. More detals about the controllng sgnals and the reference Q are presented n the sequel.... Reference Q The reference Q for two types of DR pctures are handled dfferently. Whle the reference Q s the man part of fnal Q, t has an mportant rule from the qualty and delay ponts of vew. For the frequent DR pctures, the dea s to have a qualty as much as close to neghborng pctures. mplementng a low pass flter smlar to (8) on the Q of encoded frames gves a local average value whch s used as reference Q for the frequent DR pctures. The DR pctures at scene cuts may have correlaton wth prevous encoded pctures or not. Therefore, any estmaton ndependently of prevous encoded pctures or only based on encoded pctures may lose the bt budget or the qualty. From ths pont of vew fndng a ft Q for the DR pcture at scene cut s qute challengng. We proposed a method for calculaton of Q for DR pctures at scene cuts n [4]. The Q was computed based on target bt rate and the codng complexty usng a rate-dstorton model and a codng complexty measure. A smple alternatve soluton s proposed as: Q R = ( Q + Q m )/, () where Q ndcates the average Q on all encoded pctures and Q m s a constant Q n the mddle range e.g. (8-3) for H.64/AVC. The average Q keeps the qualty of pcture close to qualty of prevous pctures when there s a correlaton between the contents of consequent scenes. The Q guarantees a bt m... Codng Complexty Adaptaton The complexty adaptaton sgnal or S controls the Q of DR pcture accordng to the codng complexty of pcture. Usng a smple frst-order rate-dstorton model, t can be derved as: S = AS Q ( X / X ), () where Q denotes the average of Q on all encoded DR pctures. A namely complexty adaptaton factor S s a constant value typcally about.7. X denotes the codng complexty of the DR pcture and X s the average of X on all encoded DR pctures. Dfferent crtera such as varance or used measure n [4] can be used for the codng complexty...3. Buffer Fullness Adaptaton The buffer fullness adaptaton sgnal or B controls the Q of DR pcture accordng to the sze and occupancy of buffer. The buffer fullness adaptaton sgnal s calculated as: 3 B = 4 38r + 4r 4r, B 8, (3) where r = BF / BS. Several ponts are consdered n desgn of functon above. Frst, a more aggressve control s provded when the buffer fullness s closer to crtcal condtons.e. underflow or over flow and a looser control s provded when the buffer s far from crtcal condtons to prevent unnecessary varatons n qualty. Second, whle an DR frame can consume a bt budget several tmes more than an nter frame, a buffer close to overflow condton can go to normal state by an DR frame wth a normal Q so a fullness close to over flow s less crtcal than to under flow. Fnally, t s asymmetrc concernng to buffer fullness and buffer sze to be approprate to fuzzy controller...4. Delay Adaptaton Accordng to (4) the sze of vrtual buffer s set based on target bt rate and target average delay. The dea s whle a smaller buffer sze s used for smaller target delay, the allocated bts to DR pctures should be smaller properly to prevent buffer underflow. The delay adaptaton sgnal bases the Q of DR pcture wth a constant value accordng to the target delay. Usng a frst-order rate-dstorton model the delay adaptaton sgnal can be derved as: D= A ( / D Q TD TD ), (4) where Q denotes the average of Q on all encoded frames. A s a constant value typcally about.55. D TD ndcates a constant value as a reference pont for roceedngs of the 6 nternatonal Conference on ntellgent nformaton Hdng and Multmeda Sgnal rocessng (H-MS'6) /6 $. 6

5 the delay n whch no adaptaton s needed. E.g. TD =.75 when BS =. 5 TR. 3. Smulaton Results To evaluate the proposed RCA, a set of smulatons were run on dfferent vdeo contents usng H.64/AVC codng standard. We repeated a number of the known vdeo sequences ncludng Foreman, Carphone, News, ars, Salesman, New York, Football, Hall and Glasgow to make sutable long (5 mnutes) vdeo sequences. Moreover, four other vdeo sequences ncludng: sport, news, musc vdeo and move contents captured from TV were used for smulaton. The vdeo sequences were encoded for dfferent target bufferng delays ncludng.,.3,.4,.5,.6 and.7 second. For each bufferng delay, smulatons were repeated for dfferent values for the frequency of DR pctures from.5 Hz to Hz. From the delay pont of vew smulaton results show that the encoded bt streams strctly obey the delay constrant. To evaluate the proposed RCA from the qualty pont of vew we compared the results of proposed RCA wth the results of constant Q case as constant qualty case. To acheve the exact average target bt rate n the constant Q case, the vdeo sequence s dvded nto sectons and the two sectons are encoded by two consequent Qs. The target bt rate s acheved by controllng the number of vdeo frames n each secton. A part of smulaton results are presented n table. The results are provded n two dfferent values for the frequency of DR pctures ( and.5 Hz) for the bt rate of 3 kb/s and the frame rate of 5 f/s and QVGA pcture format. Whle theoretcally the constant Q case almost corresponds to the maxmum average qualty case, smulaton results show that the average qualty of encoded vdeo by the proposed RCA has a SNR about. db hgher than the constant qualty case whle the controlled bt streams need small bufferng delays (about.4 s n average). Moreover, we measure the mean and varance of Q used for each case. Small values for the varance of Q (about 3.4 n average) means small varaton n qualty and t means a hgh level vsual qualty for compressed vdeo. Ths measure for a constant bt rate vdeo can be very hgh and for the descrbed constant Q case t has an average value about.5. Moreover, smlar smulatons on the mentoned known vdeo sequences for the frame rate of 3 f/s and QCF pcture format, were provded. db enhancements n SNR relatve to constant Q case wth.6 second bufferng delay and the value of. for varance of Q. TABLE 3. COMARSON THE RESULTS OF ROOSED RCA WTH THE CONSTANT QS CASE N H.64/AVC ENCODER WTH TARGET RATE OF 3KB/S, 5 F/S, QVGA CTURE FORMAT. Vdeo Content Sport News Move 4. Conclutons A novel fuzzy vdeo RCA wth delay and qualty constrants s proposed for vdeo streamng over DVB- H applcaton. The proposed RCA were optmzed such that t can utlze the advantage of varable bt rate vdeo to mprove the qualty and compresson performance and also t decreases the channel changng delay by optmal bt allocaton to frequent DR pctures n the bt stream and wth optmal varaton n the bt rate. Smulaton results on H.64/AVC vdeo codng standard show that t can provde varable bt rate vdeo wth a hgh level average qualty and wth a relatve small delay. 5. References DR Constant Q roposed RCA c. Mean Mean SNR Mean SNR Var. Fre. Delay Q (db) Q (db) Q (s) Musc Vdeo Average [] ETS, Dgtal Vdeo Broadcastng (DVB): Transmsson systems for handheld termnals, ETS standard, EN 3 34 V.., 4 [] G. Fara, J. A. Henrksson, E. Stare,. Talmola, DVB- H: Dgtal Broadcast Servces to Handheld Devces, roceedng of The EEE, V. 94, No., Jan. 6 [3] M. Rezae, S. Wenger and M. Gabbouj, Vdeo Rate Control for Streamng and Local Recordng Optmzed for Moble Devces, EEE nt. Sym. on ersonal ndoor and Moble Rado Communcatons (MRC 5), Berln, Sep. 5 [4] M. Rezae, M. M. Hannuksela, M. Gabbouj, Low- Complexty Fuzzy Vdeo Rate Controller for Streamng, EEE nt. Con. on Acoustc, Speech and Sgnal rocessng (CASS), Toulouse, France, May 6. [5] L. X. Wang, Adaptve Fuzzy System and Control: Desgn and Stablty Analyss, Englewood Clffs, NJ: rentce-hall, 994. roceedngs of the 6 nternatonal Conference on ntellgent nformaton Hdng and Multmeda Sgnal rocessng (H-MS'6) /6 $. 6