Multi-user video streaming using unequal error protection network coding in wireless networks

Transcription

1 Vukobratović and Stanković EURASIP Journa on Wireess Communications and Networking 202, 202:28 RESEARCH Open Access Muti-user video streaming using unequa error protection network coding in wireess networks Dejan Vukobratović * and Vadimir Stanković 2 Abstract In this artice, we investigate a muti-user video streaming system appying unequa error protection (UEP) network coding (NC) for simutaneous rea-time exchange of scaabe video streams among mutipe users. We focus on a simpe wireess scenario where users exchange encoded data packets over a common centra network node (e.g., a base station or an access point) that aims to capture the fundamenta system behaviour. Our goa is to present anaytica toos that provide both the decoding probabiity anaysis and the expected deay guarantees for different importance ayers of scaabe video streams. Using the proposed toos, we offer a simpe framework for design and anaysis of UEP NC based muti-user video streaming systems and provide exampes of system design for video conferencing scenario in broadband wireess ceuar networks. Introduction Rea-time muti-user (or muti-party) video streaming refers to a scenario where mutipe users, interconnected by a common communication network, perform rea-time exchange of video streams [,2]. Each of the users continuousy creates its own video stream and is interested in the continuous and rea-time recovery of the streams generated by a subset or the set of a the other users. Appication exampes incude video conferencing, mutiview video systems, muti-party peer-to-peer (P2P) video exchange, emerging mutimedia-oriented socia networking (e.g., see-what-i-see appications), etc. However, designing robust and efficient muti-user video streaming systems over wireess networks faces a number of chaenges, most notaby, the strict deay imits enforced by rea-time requirements and time-variabe wireess channe conditions responsibe for frequent packet osses. Network coding (NC) is a nove information processing technique appied in network nodes in which, instead of simpe forwarding of received data packets, the data packets are combined and resuting network coded packets are transmitted instead. The idea was first introduced for the singe-source muticast probem, where it was shown that, *Correspondence: dejanv@uns.ac.rs Department of Power, Eectronics and Communication Engineering, University of Novi Sad, Trg D. Obradovića 6, NoviSad, Serbia Fu ist of author information is avaiabe at the end of the artice unike routing, it achieves the capacity of the muticast connection [3]. For the singe-source muticast probems represented by directed acycic graphs with unit-capacity error-free edges, the cass of inear network codes achieves the muticast connection capacity [4]. Furthermore, random inear codes over sufficienty arge finite fieds open the way for simpe and fuy distributed network code design [5]. The random inear coding (RLC) approach is adapted for practica impementation in ossy packet networks [6,7], and suggested in a number of wireess networking appications [8,9]. To increase throughput and improve error resiience, NC has been recenty suggested for appications in mutimedia streaming [0-4], and in particuar, for muti-user video conferencing [5-7]. In [5], which is cosest to our work, RLC is investigated for muti-party video conferencing in wireess broadband ceuar systems. This study demonstrates that RLC appied within the centra node possess a potentia to reduce the end-to-end deay, increase throughput and improve the transmission reiabiity and system fairness. In this artice, we expore anaytica toos for the design and anaysis of a rea-time muti-user video streaming system that appies scaabe video coding and unequa error protection (UEP) RLC. We focus on a simpe scenario where wireess users exchange video streams over a common centra node with the goa of capturing 202Vukobratović and Stanković; icensee Springer. This is an Open Access artice distributed under the terms of the Creative Commons Attribution License ( which permits unrestricted use, distribution, and reproduction in any medium, provided the origina work is propery cited.

2 Vukobratović and Stanković EURASIP Journa on Wireess Communications and Networking 202, 202:28 Page 2 of 3 the fundamenta system behaviour. This work buids upon our recent theoretica anaysis of UEP RLC schemes for erasure channes [8]. Addressing a specific UEP RLC appication, the rea-time muti-user video streaming, this artice extends the ayer decoding probabiity anaysis of UEP RLC addressed in previous work to incude additiona performance measures such as expected decoding deays of different video ayers and evoution of the expected received video quaity of exchanged scaabe video streams over time. Using the presented set of anaytica toos, we offer a simpe framework for the design and anaysis of UEP RLC based rea-time muti-user scaabe video streaming systems. The framework provides fexibe approach for reiabe exchange of ayered video streams over dynamicay changing wireess channes. The appication of the proposed framework and the benefits over the standard RLC approach appied in [5] are demonstrated through the distortion-optimized system design exampes. The artice is organized as foows. Section RLC: an overview and UEP extension provides a background on RLC and its UEP extensions, and provides a decoding performance anaysis of the UEP RLC. The proposed muti-user video streaming setup is introduced in Section Muti-user video streaming using UEP NC. The same section formuates the distortion-based optimization of the muti-user video streaming system based on UEP NC. Seected UEP NC code design exampes are discussed in Section System optimization and resuts. The artice is concuded in Section Concusions. RLC: an overview and UEP extension Background and motivation For wireess broadcasting, NC is usuay motivated by the two-user packet exchange exampe in Figure (see e.g., [9]). Instead of repicating and independenty transmitting each user packet, the centra node XORs the incominguserpacketsandbroadcastsasingecodedpacket. As a resut, the number of packet transmissions required for two-way packet exchange between users reduces from four to three. The two-user exampe can be extended to muti-user scenarios using opportunistic binary NC (XOR-ing) for wireess broadcast networks, proposed in [20]. However, thebroadcastingnodeneedstoknowthebuffercontent of its neighbours in order to construct the optima encoded packet. On the other hand, extension to mutiuser scenario is possibe by appying RLC over received data packets using non-binary finite fied coefficents [5]. After N u users upoad their data to the centra node, the centra node broadcasts random inear combinations of users packets in a rateess fashion, unti each user recovers the data packets of other users (Figure 2). A user needs to receive any N u encoded packets broadcasted by the centra node in order to recover other users with high probabiity, if the finite fied size is sufficienty arge. In contrast, without NC, an aternative is repeated broadcasting of user packets in a data carouse fashion, or managing the one-to-many automatic repeat request (ARQ) mechanism, which is known to suffer a numberofdrawbacks(e.g.,excessivedeayandthefeedback imposion probem). In this artice, we extend the basic idea of Figure 2 to the case where users exchange scaabe coded video streams. Users messages are organized into ayers of different importance, starting with the most important and continuing with progressivey ess important ayers. If, e.g., due to poor channe conditions or ow access data rates, a user is unabe to fuy recover other users, it benefits from recovering as many importance ayers of other users messages starting from the most important ayer onwards. For increased protection of more important ayers over error-prone wireess inks, the advantages of the UEP forward error correction (FEC) coding are demonstrated in a number of research studies [2-23]. In this artice, we focus on the UEP RLC to expore the benefits of both rateess UEP FEC coding and NC. Random inear coding Let x = {x, x 2,..., x K } be a source message that consists of K equa-ength source packets. RLC appied over the message x produces encoded packets ω as random inear combinations of source packets with coefficients randomy seected from a given finite fied GF(2 q ): ω = U 2 U 3 U 2 U 3 BS BS x 2 BS x3 BS x x 2 x +x 2 x +x 2 U x x 5 x 4 U 4 U α i x i U 4 U U 2 U Figure Simpe two-user NC exampe. U 2 U 5 Figure 2 Simpe muti-user NC exampe (N u = 5). U 5

3 Vukobratović and Stanković EURASIP Journa on Wireess Communications and Networking 202, 202:28 Page 3 of 3 Ki= α i x i,whereα i is a randomy seected eement of GF(2 q ),andω is of the same ength as source packets. Each encoded packet carries a header information containing goba encoding coefficients within the goba encoding vector g = {α, α 2,..., α K } [6]. For unicast transmission, to reax the overhead requirements, both the transmitter and the receiver may use the pair of synchronized random number generators (RNG) to produce thesamesequenceofgobaencodingcoefficients.inthis case, ony a short RNG seed needs to be conveyed within the packet overhead. The RLC encoding can be repeated at the transmitter in a rateess fashion, unti the receiver coects enough encoded packets to decode the source message using the Gaussian eimination (GE) decoding. GE decoding introduces compexity imitations on the message ength K. However, for rea-time interactive mutimedia appications, sma vaues of K are acceptabe for practica depoyment [24,25]. UEP random inear coding Let x ={x, x 2,..., x K } be a ayered source message containing K equa-ength source packets cassified into L importance ayers. The source message starts with the most important base ayer (BL) and continues with progressivey ess important enhancement ayers (EL). The subset of the source message containing the -th ayer is denoted as x and contains k source packets, where Li= k i = K. We denote the subset of the source message containing the first ayers as x : and the number of source packets in the first ayers is K = i= k i (Figure 3). The UEP RLC scheme caed expanding window (EW) RLC was investigated in [8]. The EW RLC introduces a set of EWs over the ayered source message foowing the importance structure of the message. More precisey, for the L-ayer message, the set of L EWs is defined where the -th EW, L, contains the source bock subset x : (Figure 3). EW RLC encoding The EW RLC introduces the probabiistic encoding process over the set of EWs. For each encoded packet, one of the EWs is first seected using the predefined window st window x x 2 } } nd 2 window... th L window k k 2 k L Figure 3 Expanding window RLC. } x K seection distribution Ɣ(ξ) = L i= Ɣ i ξ i,whereɣ i is the probabiity of seection of the i-th window, L i= Ɣ i =. Then, an encoded packet is created by appying RLC ony over the seected window. Arbitrariy many EW RLC encoded packets can be produced by independenty repeating the encoding process for each encoded packet. EW RLC decoding A receiver coects correcty received EW RLC coded packets and decodes the source message (or subset of its ayers) using the standard GE decoder, as if standard RLC is used. An important difference is that the parts of the ayered source message coud be decoded even if ess than K encoded packets are received. For more detais on the design of EW RLC, we refer the interested reader to [8]. Performance anaysis of EW RLC In the foowing, we review a simpe upper bound for the set of decoding probabiities P d, (N) that the content of the -th window, L, is recovered at the receiver after receiving any N EW RLC encoded packets. The upper bound is genera and hods for any packet-eve coding and decoding scheme that appies probabiistic encoding and expanding windowing approach. Moreover, the EW RLC in combination with GE decoding achieves this bound as the fied size increases [8]. Let y = {y, y 2,..., y N } be a sequence of N received EW RLC encoded packets. For the derivation of upper bounds, y is competey described by the corresponding vector n = {n, n 2,..., n L },wheren denotes the number of received packets obtained by EW RLC coding over the -th window. We denote by y (and y ( :,respectivey) the subset of y containing the set of n N = ) i= n i received packets obtained by EW RLC encoding over the -th (the first ) window(s). For a given n, we define a set of variabes R (n), L, using the foowing recursion: R (n) = min(n, K ), R (n) = min(r (n) + n, K ),2 L. () Thus any received y can be recursivey transformed into R ={R (n), R 2 (n),..., R L (n)}. The vaues of R (n) provide an upper bound on the rank of the N K matrices whose rows are goba encoding vectors of N received packets in y :.Inotherwords,R (n) is the maximum number of source packets in x : that can be recovered from y :.UsingR we can simpy determine the set of ayers of x thereceivercanrecoverafterreceiving y. Namey,x : can be fuy recovered if R (n) = K. In addition, x : can be aso recovered if any of the arger

4 Vukobratović and Stanković EURASIP Journa on Wireess Communications and Networking 202, 202:28 Page 4 of 3 windows is recovered, i.e., if R m (n) = K m for any < m L, because arger windows contain smaer ones. Formay, the upper bound on P d, (N) foows by conditioning on n a : P d, (N n) = = I ( ) R = K L i=+ i (R j < K j ) (R i = K i ), j= where I( ) represents an indicator function equa to if its argument, which is a ogica expression, is true, otherwise, I( ) = 0, and i () ( j ()) is ogica or (and) of a sequence of ogica expressions parametrized by i (j). Conditioning on n is removed using the prior distribution over n: P Ɣ(ξ),N (n) = (2) N! n! n 2!...n L! Ɣn Ɣn Ɣn L L. (3) Finay, upper bound on P d, (N) is obtained as: P d, (N) (n,n 2,...,n L ): Li= n i =N P Ɣ(ξ),N (n)p d, (N n). (4) For a given ayered source message, P d, (N) depends ony on the seected window seection distribution Ɣ(ξ). b Therefore, designing EW RLC codes with desired P d, (N) behaviour reduces to the design of appropriate Ɣ(ξ) [8]. Muti-user video streaming using UEP NC System mode In this artice, we propose the UEP NC as a core component of a rea-time muti-user video streaming system. For simpicity, we focus on a wireess ceuar network exampe where N u mobie users (U, U 2,..., U Nu )participatein a muti-party video conferencing session over a common base station (BS) within a singe ce (Figure 4). The presented resuts are appicabe in simiar scenarios, e.g., if instead of a singe BS users connect to different BSs mutuay interconnected by high-speed inks (e.g., fiber optic) or if instead of a ceuar network we observe a Wi-Fi access point. In our scenario, every user continuousy segments its own video stream into groups of frames (GOF), where each GOF contains N gof frames, and compresses every GOF using a scaabe video coder. For each user U i and each compressed GOF, the output of the video coder is a ayered source message x (i) that contains K (i) source packets, each of ength b bits, organized into L ayers, where the -th ayer contains k (i) packets. The vaues of b and L are the same across a users whereas for each user U i and each GOF, the vaues K (i) and {k (i), k(i) 2,..., k(i) L } are in genera different. Video streaming among users in a session may be observed as a GOF-by-GOF exchange process. A singe GOF exchange period repeats every T gof = N gof /N fps seconds, where N fps is the number of frames per second of the video stream, and both N gof and N fps are equa across a users. For simpicity, we assume that GOF periods are aigned among different users, i.e., the messages x (i) are synchronousy generated by a the users. For every GOF period, the goa of each user is to share its own and coect other users GOFs, or at east as many of their ayers starting from the beginning onwards, within a strict deay imits. During each GOF period, the data exchange process can be divided into two phases. In the first, upoad phase, users simutaneousy upoad their data to the BS, and in the second, broadcast phase, the BS broadcasts the received data to a the users. We assume that orthogona channes are aocated between each user and the BS, and for the broadcast transmission by the BS, aowing for simutaneous transmission on a channes. Each wireess ink is modeed as a synchronous time-sotted packet-erasure ink where the fixed size encoded packets of ength b bits are transmitted using a fixed transmission data rate R and erasure probabiity ɛ. We assume U 2 U 3 U BS U 4 Figure 4 Muti-user video streaming system mode.

5 Vukobratović and Stanković EURASIP Journa on Wireess Communications and Networking 202, 202:28 Page 5 of 3 that the pair (R,ɛ) is in genera different on different wireess inks, it remains fixed during the transmission period T gof of a singe GOF, and may change between different GOF transmissions. The time sot duration T p = b/r represents a time required for a singe encoded packet transmission. c In the upoad phase, to protect data from erasures, the user U i encodes the source message x (i) using the EW RLC scheme defined by a window seection distribution Ɣ (i) (ξ) and streams the encoded packets using the user rate R (i). The BS recovers the users messages using N u independent GE decoders dedicated to different users. Afterdecodingasmanyuserayersaspossibe,theBScreates its own source message x (BS) that contains a or a subset of users message ayers. For exampe, if a N u users are competey recovered, the message x (BS) is of ength K (BS) = N u i= K (i) packets and contains L ayers. The - th ayer x (BS) comprises the tota of k (BS) = N u i= k(i) packets from the -th ayer data of a N u user messages, x (BS) ={x (), x (2),..., x (N u) }, as iustrated in Figure 4 for the two-ayer scenario. In genera, the BS may create the -th ayer x (BS) from the -th ayers of a subset of users. In the broadcast phase, the BS appies the EW RLC scheme defined by a window seection distribution Ɣ (BS) (ξ) over x (BS) and broadcasts the stream of encoded packets using the broadcast rate R (BS). Each user recovers the BS message x (BS) using the GE decoder where, prior to decoding, the user cances out its own packets from the received encoded packets. We note that the two phases may overap in time, i.e., the BS may start broadcasting a subset of (aready recovered) ayers of x (BS) before the upoad phase of a users is competed. Singe-ink anaysis: decoding and deay performance To anayze the system in Figure 4, we focus on a singe-ink transmission of UEP RLC coded ayered message between any transmitter-receiver pair during a fixed time period T. Instead of ayer decoding probabiities after fixed number of received packets, P d, (N) (Section Performance anaysis of EW RLC ), we shift our interest to ayer decoding probabiities after fixed transmission time, P d, (T). However, unike P d, (N), deriving P d, (T) requires introduction of a packet-eve channe mode. During the period of duration T, the transmission process consists of N T = RT/b encoded packet transmissions (packet sots). We redefine the received sequence y = {y, y 2,..., y NT } to describe the outcome of a N T transmissions, where y i may represent either the received encoded packet or a ost (erased) packet. The received sequence y can be described by vector n = {n, n 2,..., n L, n e }, where, as before, n i is the number of received encoded packets obtained by encoding over the i-th EW, N = L i= n i N T isthenumberof(correcty) received encoded packets during the interva T, and n e = N T N isthenumberoferasedpackets. For a fixed T, the number of received encoded packets N is dependent on the underying channe packet-oss mode. For simpicity, we assume a packet erasure channe mode that erases encoded packets independenty with erasure probabiity ɛ.toobtainp d, (T), we use conditioning over n: P d, (T) = P Ɣ(ξ),ɛ (n)p d, (T n), (5) n where P Ɣ(ξ),ɛ (n) is sighty atered version of (3) that accounts for n e erased packet events: N T! P Ɣ(x),ɛ (n) = n! n 2!...n L! n e! (6) [ Ɣ ( ɛ)] n [ Ɣ 2 ( ɛ)] n 2...[ Ɣ L ( ɛ)] nl (ɛ) n e. P d, (T n) can be obtained as the ayer decoding probabiity P d, (N n) after N received encoded packets (Section Performance anaysis of EW RLC, Equations () (2)) since, by knowing n, we directy obtain N = L i= n i. The knowedge of P d, (T) impicity provides information on the decoding deay of the -th message ayer. For exampe, one can search for minima transmission period T () th probabiity P d, (T () such that the -th message ayer decoding th ) > P() th,wherep() th is the threshod decoding probabiity set in advance. For more expicit deay information, one can obtain the probabiity distribution p d, (N T ) that the -th message ayer is recovered after exacty the N T -th time sot (and cannot be recovered before): p d, (N T ) = (7) = N T ( P d, (T = i T p )) P d, (T = N T T p ), i= and T p = b/r. The expected deay E [ N T ] for recovery of the -th message ayer is: E [ N T ] = N T = N T p d, (N T ). (8) The expected deay of recovery of the compete source message E[ N T ] for the EW RLC scheme is equa to the recovery deay of the ast L-th ayer: E[ N T ] = E L [ N T ]. E [ N T ] is expressed in terms of the number of time sots, but it can be easiy converted into absoute time vaues as E [ T] = E[ N T ] T p. Exampe. Let x be a ayered source message containing K = 60 source packets of size b = 3, 200 bits (400 bytes) divided into L = 2 ayers: the BL containing k = 20 packets and the EL containing k 2 = 40 packets. The EW RLC

6 Vukobratović and Stanković EURASIP Journa on Wireess Communications and Networking 202, 202:28 Page 6 of 3 P d, (T), (T) P d,/2 =0 P d, =0.3 =0.3 P d, =0.4 =0.4 P d, =0.5 =0.5 P d, =0.6 =0.6 P d, = T [s] Figure 5 P d, (T) and (T) for EW RLC over the range of Ɣ vaues. scheme defined by Ɣ(ξ) = Ɣ ξ + ( Ɣ )ξ 2 is appied over x producing continuous stream of 400-bytes ong encoded packets. The wireess ink towards the receiver is modeed as a synchronous rate R = 2 Mbit/s ink with packet erasure probabiity ɛ = 0.. Figure 5 presents the evoution of ayer decoding probabiities P d, (T) and (T) over time T at the receiver for the range of Ɣ vaues. As a baseine scheme, we start from the midde soid curve for Ɣ = 0, representing standard RLC appied over the whoe message, which resuts in simutaneous recovery of both message ayers. By increasing Ɣ, we obtain the UEP effect of the EW RLC scheme, where soid curves for the P d, (T) graduay shift to the eft, i.e., towards earier recovery of the most important data. The extreme case of Ɣ = resuts in the eariest recovery of the most important part represented by the eftmost dashed curve. The increase in Ɣ comes at the price of deayed decoding of ess important ayer (T). Figure 6 presents the expected ayer decoding deays E [ T] and E 2 [ T] as a function of Ɣ.Notethatsignificant decrease of E [ T] with the increase in Ɣ initay comes with a reativey sma oss in E 2 [ T]. Forexampe, for Ɣ = 0.5, E [ T] drops from 05 to 62 ms ( 4%)whie E 2 [ T] rises from 05 to 26 ms (+20%). The singe-ink anaysis can be extended to the scenario where the transmitter changes the appied EW RLC code during the transmission (i.e., switches between different Ɣ(ξ)). As an exampe, we derive P d, (T) for T > T,given that the transmitter has appied the EW RLC defined by Ɣ a (ξ) during 0 t T, and the EW RLC defined by Ɣ b (ξ) for t > T. For the ink parameters R and ɛ, the transmitter sends N = RT /b encoded packets using Ɣ a (ξ) and N 2 = R(T T )/b encoded packets using Ɣ b (ξ). The received sequence y =[ y y 2 ] is a concatenation of two sequences y and y 2 of ength N and N 2, respectivey. It can be described by the vector n = n n 2, where the vectors n and n 2 represent the description of E [T] E 2 [T] 0.6 E [T], E 2 [T] [s] Γ Figure 6 E [ T] and E 2 [ T] for EW RLC over the range of Ɣ vaues.

7 Vukobratović and Stanković EURASIP Journa on Wireess Communications and Networking 202, 202:28 Page 7 of 3 P d, (T), (T) P d, = = (T) Γ (a) =0.5, Γ(b) =0. P 0. d, (T) GF(2 8 ) Γ =0.5 (T) GF(2 8 ) Γ = T [s] Figure 7 P d, (T) and (T) forewrlcwhichchangesɣ (a) = 0.5 to Ɣ (b) = 0. at the time instant T = y and y 2 respectivey (as defined earier), and is the component-wise sum of two equa-ength vectors. Since n and n 2 foow probabiity distributions P Ɣ (a) (x),ɛ (n ) and P Ɣ (b) (x),ɛ (n 2) given by Equation (6), the ayer decoding probabiities are obtained as in (5): P d, (T) = (9) = P Ɣ (a) (x),ɛ (n )P Ɣ (b) (x),ɛ (n 2) P d, (T n), n n 2 where P d, (T n) foows from P d, (N = L i= n i n). Exampe 2. We continue Exampe by investigating the evoution of P d, (T) and (T) over time T at the receiver if the transmitter appies Ɣ a (ξ) = 0.5ξ + 0.5ξ 2 for the first T = 25 ms, and then changes to Ɣ b (ξ) = 0.ξ + 0.9ξ 2 (the remaining parameters are the same as in the previous exampe). Figure 7 compares the case where Ɣ a (ξ) changes to Ɣ b (ξ) with the case where Ɣ a (ξ) is used throughout the transmission. Figure iustrates thattheɣ(ξ) change has no effect on P d, (T) as it comes too ate (P d, (T = 0.25) = ), whereas the improvement of (T) for T > T is notabe due to the increase in the second window seection probabiity from 0.5 to 0.9, which points out to possibe adaptive (e.g., feedback-driven) updates of Ɣ(ξ) during transmission. In addition, Figure 8 demonstrates that the upper bound expressions for P d, (T) used in this artice match very we the exact cacuation of P d, (T) for a finite fied size GF(2 8 ) (markers) [8]. Finay, anaytica resuts presented above can be easiy extended to Gibert Eiot erasure channe mode with two states: the good and the bad state. This foows from the fact that the probabiity distribution of the number n e of erasures over N T consecutive reaizations of the channe (i.e., over time interva T) is known (e.g., see [26]). The remaining N T n e non-erased channe reaizations deiver encoded symbos from different EWs according to the mutinomia distribution aw (3). Figure 8 Time-diagram of different transmission phases in the system mode.

8 Vukobratović and Stanković EURASIP Journa on Wireess Communications and Networking 202, 202:28 Page 8 of 3 Distortion-optima system design In this section, we appy the singe-ink anaysis to anayze the muti-user video conferencing setup introduced in Section System mode. Our goa is to formuate the system design probem that eads to the EW RLC code design providing distortion-optima system performance. As a distortion measure, we use peak signa to noise ratio (PSNR) as a standard video quaity metric foowing directy from the mean square error (MSE) distortion measure. We use the terms video quaity and distortion interchangeaby whie refering to video quaity (PSNR) measure. We focus on a singe message (GOF) exchange cyce among the system users. After the initia deay of t = T gof needed for each user to acquire and compress N gof frames of video (assuming negigibe compression deay), the users start the upoad phase. During the upoad phase, the BS waits to receive enough encoded packets to recover the users messages or as many of their ayers with sufficienty high probabiity. The upoad phase duration T u is upper bounded by the GOF period duration T gof as, after this period expires, users are suppied with a new set of compressed messages, which marks the beginning of the upoad phase of the next message exchange cyce. From the set of recovered ayers, the BS creates its own message x (BS) which is broadcasted back to the users over the broadcast downink channe during the broadcast phase of duration T d. To simpify the anaysis, we assume the upoad and the broadcast phase do not overap, i.e., after the upoad phase of duration T u,theusersstoptransmitting and start istening the BS for the foowing period of duration T d. This anaysis provides guaranteed (owerbound) performance for the overapping phases case, as we discuss ater. We are interested in the system design that maximizes the tota average received video quaity at the user terminas after a given target system deay T = T gof + T u + T d. Note that, as T gof is constant and T isgiveninadvance, it foows that the system design shoud optimay baance between the T u and T d (T u + T d = T T gof = const.) The time diagram of the system mode, ignoring the propagation and data processing deays, is iustrated in Figure 8. Upoad phase The upoad phase is represented by N u parae and independent singe-ink transmission processes, each characterized by different message/ayer sizes (K (i), {k (i), k(i) 2,..., k(i) L }) and channe state pairs (R(i),ɛ (i) ). Assuming that the user knows d (R (i),ɛ (i) ) and that the vaue of T u is fixed in advance by the BS, the set of ayer decoding probabiities P (i BS) d, (T u ) of the i-th user message at the BS can be cacuated for any seected EW RLC parameter Ɣ (i) (ξ). Inthefoowing,wefocusonasimpe user upoad strategy where the user appies standard RLC over the argest window such that the decoding probabiity P (i BS) d, (T u )>P th (ifsuchexists),wherep th is a (cose to one) vaue of threshod probabiity seected in advance. More formay, the i-th user wi appy RLC ony over the (i) -th window, where (i) is obtained as: { } (i) = max : P (i BS) d, (T u )>P th. (0) Note that appying RLC ony over the (i) -th window is equivaent to the specia case of appying UEP RLC with the window seection distribution Ɣ (i) (ξ) = ξ (i) (i.e., the one which paces probabiity one on the (i) -th window). Overa, the set of N u users wi upoad the subset of their ayers, jointy described by vector = { (), (2),..., (Nu) }, within the upoad phase of duration T u. The probabiity P th can be seected so as to keep the overa probabiity P (BS) d, (T u ) P N u th that the BS wi recover the set of users ayers described by during the upoad phase of duration T u sufficienty high. Broadcast phase Duringthebroadcastphase,theBSappiestheEWRLC code defined by Ɣ (BS) (ξ) over the BS message x (BS) (), which is determined by the set of upoaded user ayers. From, one can easiy obtain the BS message size parameters (K (BS), {k (BS), k (BS) 2,..., k (BS) L }). The broadcast phase can be aso anayzed using the singe-ink anaysis appied on the parameters of the broadcast transmission, as seen by each of the system users. In other words, given Ɣ (BS) (ξ), x (BS) () and the BS-to-user-i (BS-i) transmission ink parameters (R (BS),{ɛ BS i } i Nu ), the singe-ink P (BS i) d, anaysis provides the set of ayer decoding probabiities (T d ), L, describing the i-th user capabiity to recover the ayers of the BS message x (BS) () after the broadcast phase. Thus the broadcast phase reduces to the EW RLC design probem for muticast/broadcast setup that aims to simutanousy satisfy heterogeneous user ink conditions ({ɛ BS i } i Nu ). This probem has been recenty addressed for expanding window fountain (EWF) code design in video muticast setup [27], however, with the difference that in this artice, instead of broadcasting a singe stream, the BS simutaneousy broadcasts amixtureofn u user streams. Each user simutaneousy receives N u video streams originating at the remaining system users. The average received video quaity D (i) perceived by the i-th user is obtained by averaging over the received video quaities of a N u video streams: D (i) = N u N u j=,j =i D (i) j, ()

9 Vukobratović and Stanković EURASIP Journa on Wireess Communications and Networking 202, 202:28 Page 9 of 3 where D (i) j is the average received PSNR of the j-th user video content as perceived by the i-th user. D (i) j can be obtained by combining the resuts of the upoad and the broadcast phase anaysis: D (i) j (j) = P (BS) d, (T u ) = P (BS i) d,: (T d ) D j,:, (2) where the sum is taken over the set of (j) L ayers of the j-th user incuded in x (BS) (). In the above expression, P (BS-i) d,: (T d ) is the probabiity that exacty the first ayers of the BS message x (BS) () are recovered at the user i: P (BS i) d,: (T d ) = P (BS i) d, (T d ), = 0 P (BS i) d, (T d ) ( P (BS i) d,+ (T d)), 0 < < L P (BS i) d, (T d ) = L, (3) and D j,: is the average received PSNR of the j-th user video content after recovery of the first ayers (averaged over a the frames of the transmitted GOF). Finay, the average received PSNR, averaged across a the users of the muti-user video streaming session, is equa: D = N u D (i). (4) N u i= System parameters and design From () (4), by factoring out P (BS) d, (T u ),wenotethat the distortion-optimized system design aows for independent design of the upoad and the broadcast phase, given the duration T u and decoding probabiity threshod P th are fixed. In other words, by fixing and informing the users on the vaues of T u and P th,thesetofayers (T u ) that can be reiaby upoaded by users in the upoad phase can be determined by corresponding users. Consequenty, the BS message x (BS) () and T d = T T gof T u is aso determined, which reduces the broadcast phase design to optimization of the EW RLC code parameter Ɣ (BS) (ξ) such that the average received video quaity D is maximized after the target system deay T. Overa, for the distortion-optimized system design, the BS shoud optimay baance between the upoad and the broadcast phase by seecting appropriate T u, appropriate threshod probabiity P th, and optimay satisfy heterogeneous user requirements by seecting optimized Ɣ (BS) (ξ). The optima soution weights between the number of ayers that coud be upoaded to the BS with reiabiity P th after T u and their quaity of reconstruction at the set of heterogeneous users after T d. System optimization and resuts System optimization For the system mode and distortion-optimized design discussed above, the system optimization process is performed centray, e.g., at the video conference server coocated with the centra BS node. Given the parameters of a the user messages (K (i), {k (i), k(i) 2,..., k(i) L }), upink channe conditions (R (i),ɛ i ) and the broadcast channe conditions (R (BS),{ɛ BS i } i Nu ), the BS shoud provide the duration of the upoad phase T u, the threshod probabiity P th and the EW RLC code design parameter Ɣ (BS) (ξ), such that the average received PSNR D is maximized after the target system deay T. In other words, the BS soves the foowing probem: max D, (5) T u,p th,ɣ (BS) (ξ) where 0 T u min{t gof, T T gof } and for Ɣ (BS) (ξ) we have 0 Ɣ (BS), L and L = Ɣ (BS) =. Assuming that the BS knows the channe conditions (e.g., by measurements and user reporting), it sti needs to know the user message parameters (K (i), {k (i), k(i) 2,..., k(i) L })tobeabetoperformtheabove optimization. Since these data cannot be obtained instantaneousy at the BS, to avoid deays, we assume that the BS uses information avaiabe from recent GOF exchanges (e.g., the ast GOF or the average over ast severa GOFs). This way, the BS is abe to perform system optimization prior to the start of the upoad phase and to broadcast the required parameters T u and P th back to the users. The users then determine the number of ayers (i) to upoad to the BS and start the upoad phase. In genera, the compexity of cacuation of the set of ayer decoding probabiities in Sections Performance anaysis of EW RLC and 2 grows exponentiay, due to an exponentia number of terms in sums given in Equations (4) and (5), as K and L grows. However, in practica appications, the cacuations are tractabe due to the fact that K, L and N u are usuay sma. For exampe, K is aready bounded by GE decoding compexity and shoud not exceed K 00; the number of scaabe video ayers is typicay sma, e.g., L < 5; and for comfortabe use of reatime muti-user video conferencing system, N u shoud aso be sma, e.g., N u < 5. (note that N u can be arger as ong as each user dispays ony a sma subset of active user streams). With the restrictions on K, L and N u,theoptimization probem can be evauated at the BS side server with acceptabe deay. Aternativey, the BS may run optimization ess frequenty then on a GOF-by-GOF basis, using accumuated averages of channe conditions and GOF message enghts and periodicay update the users and the BS transmitter with the new vaues of (T u, P th ) and Ɣ (BS) (ξ),respectivey.

10 Vukobratović and Stanković EURASIP Journa on Wireess Communications and Networking 202, 202:28 Page 0 of 3 Tabe Parameters of H.264/SVC sequences (L = 2, N gof = 4) Sequence/ayers Number of packets Bit rate Y-PSNR b = 3, 200 [bits] [kbps] [db] Stefan BL k () = Stefan BL + EL K () = Foreman BL k (2) = Foreman BL + EL K (2) = News BL k (3) = News BL + EL K (3) = Coast BL k (4) = Coast BL + EL K (4) = Design exampes The muti-user video streaming system design proposed in this artice is iustrated using numerica exampes. Exampe 3. In this exampe, we present a distortionoptimized UEP NC soution for the muti-user video conferencing system with N u = 4 users (Figure 4). We assume users perform rea-time exchange of H.264/SVC compressed CIF Stefan, Foreman, News and Coastguard sequences ( , N fps = 30), each user sharing a different video sequence. Users encode the sequences into L = 2 quaity ayers (BL and one EL) using the coarse-grain scaabe (CGS) coding feature. The GOF size is set to very ow vaue of N gof = 4 in order to reduce the start-up coding deay T gof = 4/30 = 33 ms to the acceptabe vaue. The parameters of the obtained ayered source messages, after H.264/SVC compression and averaged over the frames of a sampe GOF we use for optimization, are given in Tabe. For the upink channe parameters, for each user, we seect rate vaues around 2 Mbps and erasure probabiities in the range ɛ = : (R () =.5 Mbps, ɛ = 0.07), (R (2) =.8 Mbps, ɛ 2 = 0.5), (R (3) =2.3 Mbps, ɛ 3 = 0.05) and (R (4) =.5 Mbps, ɛ 4 = 0.2), to account for the variations in particuar upink conditions. The BS broadcast rate is set to R (BS) =6 Mbps and, for simpicity, the broadcast erasure rates towards each user are set equa to the erasure rates of the corresponding upink channes, i.e., ɛ BS i = ɛ i. Given the system parameters above, we seek for the optima system parameters (T u, Ɣ (BS) (ξ)) such that the average received PSNR D across a system users is maximized after the target deay T = 250 ms. For simpicity, we fix P th = The soution is iustrated in Figure 9 where average PSNR is potted as a (two-dimensiona) function of (T u, Ɣ (BS) ). The system achieves the best Γ BS T u = T u PSNR Γ BS Figure 9 Two-ayer muti-user video conferencing optimization exampe.

11 Vukobratović and Stanković EURASIP Journa on Wireess Communications and Networking 202, 202:28 Page of 3 Tabe 2 Expected deays in Exampe Upoad E [ T] [ms] Broadcast E [ T] [ms] transmission transmission U -BS E [ T] = BS-U E 2 [ T] = 4.29 U 2 -BS E [ T] = BS-U 2 E 2 [ T] = U 3 -BS E 2 [ T] = BS-U 3 E 2 [ T] = U 4 -BS E [ T] = BS-U 4 E 2 [ T] = average performance for T u = 66 ms where a users are abe to share at east their BL, whie user 3 is abe to upoad both ayers to the BS, i.e., = (,, 2, ). For optima T u = 66 ms, a separate (ower) graph shows the system performance for different EW RLC codes at the BS. Athough the maximum of D is achieved over the range of first window seection probabiities Ɣ (BS), it is favourabe to seect as arge Ɣ (BS) as possibe to reduce the decoding deay for the first ayer, whie sti maintaining high probabiity of recovery of the second ayer of x (BS) at a users. e Tabe 2 iustrates the average decoding deays for upoad and broadcast phase transmisssions for the set of upoaded ayers = (,, 2, ) and the soution point (T u, Ɣ (BS) )=(0.066,0). We can easiy note that the sum of maximum deays experienced during the upoad/broadcast phase cosey satisfies the deay imits imposed by the system: E (2 BS) 2 [ T] +E (BS 3) 2 [ T] = < 7 = T T gof, where maximum upoad deay is beow seected upoad duration E (2 BS) 2 [ T] = < 66 = T u. This points out to the possibiity of approximated system design using expected deay cacuations. Exampe 4. Additiona fexibiity in the system design is obtained if the users compress their video streams into arger number of ayers. In this exampe, we observe the performance of the distortion-optimized system design for the same transmission parameters as in the previous exampe, but where the ayered source message is compressed into L = 4 quaity ayers (see Tabe 3 for the message parameters). The system performs optimay for T u = 64 mswhereusersareabetoupoadthesetofayers = (2, 3, 3, 2), wherep th = 0.99 is assumed fixed. The EW RLC broadcast phase parameters that achieve the optima vaue D = are for the window seection distribution Ɣ (BS) (ξ) = 0.5ξ + 0.5ξ 3.Wenotethatthegain obtained in average system distortion D is not arge, due to the fact that compressing video into arger number of ayers introduces sma performance penaties, but the system fexibiity refected through better ayer resoution provides more options for the system design process. Decode-and-broadcast versus buffer-and-broadcast In the proposed system, we appy decode-and-broadcast operation in centra muti-user video streaming point: the upoading streams are firsty decoded and then broadcasted within the non-overapping broadcasting stage. Ceary, this approach simpifies appications of our anaytica toos and enabes simpe and eegant system design, however, improvements are possibe if the broadcast phase is initiaized before the incoming user messages are competey recovered. A possibe improvements are shorty commented beow. Layer-by-ayer decode-and-broadcast LetusassumetheupoadphasewhereageneraUEPRLC is appied instead of the specific RLC case that encodes the argest window decodabe within T u.inthiscase, the unequa recovery time (URT) property enabes the centra point to decode user ayers sequentiay over time, starting from the BL onwards [8]. Thus the centra point is abe to produce encoded packets as soon as the BL of the message x (BS) is decoded and incude additiona ayers as soon as they become avaiabe whie updating the broadcast EW RLC code parameter Ɣ (BS) (ξ) on the fy, as iustrated in Exampe 2. We note that this scenario introduces a trade-off between increase in the upoad deays of higher ayers and decrease in the beginning of the broadcast phase, which has to be baanced by the optima soution. Unfortunatey, the distortion optimized system design for this scenario woud resut in tedious optimization probem, which is why we eave it out of consideration. However, we note that expected deay anaysis, simiar to the one presented in Tabe 2, coud Tabe 3 Parameters of H.264/SVC sequences (L = 4, N gof = 4) Sequence/ Number of packets Bit rate [kbps] Y-PSNR ayers b = 3, 200 [bits] [db] Stefan BL K () = Stefan BL + EL K () 2 = Stefan BL + 2EL K () 3 = Stefan BL + 3EL K () 4 = Foreman BL K (2) = Foreman BL + EL K (2) 2 = Foreman BL + 2EL K (2) 3 = Foreman BL + 3EL K (2) 4 = News BL K (3) = News BL + EL K (3) 2 = News BL + 2EL K (3) 3 = News BL + 3EL K (3) 4 = Coast BL K (4) = Coast BL + EL K (4) 2 = Coast BL + 2EL K (4) 3 = Coast BL + 3EL K (4) 4 =

12 Vukobratović and Stanković EURASIP Journa on Wireess Communications and Networking 202, 202:28 Page 2 of 3 be used as a simpe approximation for the ayer-by-ayer decode-and-broadcast system design. Buffer-and-broadcast Finay, the simpest buffer-and-broadcast soutionfoows the standard NC approach in which a the received encoded packets are buffered, and new encoded packets produced by appying RLC over the buffer content [6,7]. In the proposed UEP RLC case, the centra point maintains L separate buffers, each coecting encoded packets of different users produced over one of the L windows. As soon as the upoad phase starts fiing the buffers, the broadcast phase starts producing encoded packets where each encoded packet resuts from appying RLC over one of the buffers seected independenty by the appropriate window (i.e., buffer) seection distribution Ɣ (BS) (ξ). Athough very efficient, this soution acks efficient anaysis and distortion-based optimization toos. In addition, the probem of broadcasting ineary dependent encoded packets may become significant as the upoad user rates decrease and broadcast rate increases (i.e., the rate of encoded packets generation exceeds the rate of incoming source data). Concusions Rea-time sharing of video content among mutipe users over wireess networks is underying a number of existing and upcoming mobie mutimedia services. For robust, fexibe and efficient impementation of such services, this artice considered a combination of scaabe video coding and UEP NC. We have presented anaytica toos capabe of producing the vaues of key system design parameters that resut in the distortion-optima system performance. The appications of the proposed toos are iustrated through severa exampes invoving a simpe singe access point muti-user scenario. Endnotes a For compactness, we denote R (n) as R. b Note that, due to the probabiistic encoding, the decoding performance is independent of the packet erasure process in the channe and depends ony on the number N of received packets. c This mode roughy captures the behaviour of adaptive moduation and coding (AMC) at the physica ayer of ceuar systems where, depending on the channe quaity feedback avaiabe at the BS, different AMC modes coud be approximatedby different (R,ɛ) pairs. We assume sowy-varying channes where AMC mode changes are of the order of T gof. d In state-of-the-art wireess ceuar broadband systems such as LTE or WiMAX, channe quaity indicators (CQI) are continuousy fed back by user equipment to the BS. e For presentation purpose, Figure 9 is obtained by bruteforce cacuation over a grid of points in (T u, Ɣ (BS) ) space. In genera, (one of) the optima soution(s) can be obtained by appying noninear programming methods such as sequentia quadratic programming (e.g., using MATLAB). Competing interests The authors decare that they have no competing interests. Acknowedgements Dejan Vukobratovic was supported by a Marie Curie European Reintegration Grant FP7-PEOPLE-ERG-200 MMCODESTREAM within the 7th European Community Framework Programme. Author detais Department of Power, Eectronics and Communication Engineering, University of Novi Sad, Trg D. Obradovića 6, Novi Sad, Serbia. 2 Department of Eectronic and Eectrica Engineering, University of Strathcyde, 204 George Street, G XW, Gasgow, UK. Received: 27 February 202 Accepted: 20 June 202 Pubished: 3 Juy 202 References. H Shiang, M van der Schaar, Muti-user video streaming over muti-hop wireess networks: a distributed, cross-ayer approach based on priority queuing. IEEE J. Se. Areas Commun. 25(4), (2007) 2. X Zhu, P Agrawa, J Pa Singh, T Apcan, B Girod, Rate aocation for muti-user video streaming over heterogenous access networks. in ACM MULTIMEDIA 07, (Augsburg, Germany, 2007), pp R Ahswede, N Cai, S yen Robert Li, RW Yeung, Network information fow. IEEE Trans. Inf. Theory. 46(4), (2000) 4. S yen Robert Li, RW Yeung, N Cai, Linear network coding. IEEE Trans. Inf. Theory. 49(2),37 38 (2003) 5. T Ho, M Medard, R Koetter, DR Kargerm, M Effros, J Shi, B Leong, A random inear network coding approach to muticast. IEEE Trans. Inf. Theory. 52(0), (2006) 6. PA Chou, Y Wu, K Jain, Practica network coding. in Aerton 2003 Conference, (2003) 7. DS Lun, M Medard, R Koetter, M Effros, On coding for reiabe communication over packet networks. Phys. Commun., 3 20 (2008) 8. C Gkantsidis, P Rodriguez, Network coding for arge scae content distribution. in IEEE INFOCOM 2005, (Miami,FL,USA,2005),pp P Chou, Y Wu, Network coding for the internet and wireess networks. IEEE Signa Process. Mag. 24(5),77 85 (2007) 0. E Magi, P Frossard, An overview of network coding for mutimedia streaming. in IEEE ICME 2009, (New York, NY, USA, 2009), pp J Zhao, F Yang, Q Zhang, Z Zhang, F Zhang, LION: ayered overay muticast with network coding. IEEE Trans. Mutimed. 8(5), (2006) 2. M Wang, B Li, R 2 : random push with random network coding in ive peer-to-peer streaming. IEEE J. Se Areas Commun. 25(9), (2007) 3. H Seferogu, A Markopouou, Video-aware opportunistic network coding over wireess networks. IEEE J. Se. Areas Commun. 27(5), (2009) 4. N Thomos, P Frossard, Network coding of rateess video in streaming overays. IEEE Trans. Circ. Syst. Video Techn. 20(2), (200) 5. H Wang, R Chang, CCJ Kuo, Wireess muti-party video conferencing with network coding. in IEEE ICME 2009, (New York, NY, USA, 2009), pp M Ponec, S Sengupta, M Chen, J Li, PA Chou, Muti-rate peer-to-peer video conferencing: a distributed approach using scaabe coding. in IEEE ICME 2009, (New York, NY, USA, 2009), pp H Zhang, J Zhou, Z Chen, J Li, Minimizing deay for video conference with network coding. in ACM SIGCOMM 2009, (Barceona, Spain,2009) 8. D Vukobratović, V Stanković, Unequa error protection random inear coding strategies for erasure channes. IEEE Trans. Commun. 60(5), (202)

13 Vukobratović and Stanković EURASIP Journa on Wireess Communications and Networking 202, 202:28 Page 3 of 3 9. Y Wu, P Chou, SY Kung, Information exchange in wireess networks with network coding and physica-ayer broadcast. in Proc. CISS 2005, (Batimore, MD, USA, 2005) 20. S Katti, H Rahu, W Hu, D Katabi, M Medard, J Crowcroft, XORs in the air: practica wireess network coding. in ACM SIGCOMM 2006, (Pisa, Itay, 2006), pp U Horn, K Stuhmuer, M Link, B Girod, Robust internet video transmission based on scaabe coding and unequa error protection. Signa Process. Image Commun. 5,77 94 (999) 22. V Stankovic, R Hamzaoui, Live video streaming over packet networks and wireess channes. in IEEE Packet Video 2003, (Nantes, France, 2003) 23. E Maani, AK Katsaggeos, Unequa error protection for robust streaming of scaabe video over packet ossy networks. IEEE Trans. Circ. Syst. Video Tech. 20(3), (200) 24. H Shojania, B Li, Random network coding on the iphone: fact or fiction? in ACM NOSSDAV 2009, USA, (Wiiamsburg, VA, USA, 2009), pp P Vingemann, F Fitzek, M Pedersen, J Heide, H Charaf, Synchronized mutimedia streaming on the iphone patform with network coding. in IEEE CCNC 20, USA, (Las Vegas, NV, USA, 20), pp L Wihemsson, LB Mistein, On the effect of imperfect intereaving for the Gibert Eiott channe. IEEE Trans. Commun. 47(5), (999) 27. D Vukobratovic, V Stankovic, D Sejdinovic, L Stankovic, Z Xiong, Scaabe video muticast using expanding window fountain codes. IEEE Trans. Mutimed. (6), (2009) doi:0.86/ Cite this artice as: Vukobratović andstanković: Muti-user video streaming using unequa error protection network coding in wireess networks. EURASIP Journa on Wireess Communications and Networking :28. Submit your manuscript to a journa and benefit from: 7 Convenient onine submission 7 Rigorous peer review 7 Immediate pubication on acceptance 7 Open access: artices freey avaiabe onine 7 High visibiity within the fied 7 Retaining the copyright to your artice Submit your next manuscript at 7 springeropen.com