J. Vis. Commun. Image R. 25 (2014) 1493–1506

Contents lists available at ScienceDirect

J. Vis. Commun. Image R.

journal homepage: www.elsevier .com/ locate/ jvc i

Engineering wireless broadband access to IPTV

http://dx.doi.org/10.1016/j.jvcir.2014.06.0131047-3203/� 2014 Elsevier Inc. All rights reserved.

⇑ Corresponding author.E-mail address: fleum@essex.ac.uk (M. Fleury).

Laith Al-Jobouri, Martin Fleury ⇑, Mohammed GhanbariUniversity of Essex, Wivenhoe Park, Colchester CO4 3SQ, United Kingdom

a r t i c l e i n f o a b s t r a c t

Article history:Received 18 October 2013Accepted 20 June 2014Available online 30 June 2014

Keywords:H.264/AVCIPTVVideo codecVideo streamingWiMAXApplication-layer FECData-partitioningError resilience

IPTV is now extending to wireless broadband access. If broadband video streaming is to achieve compet-itive quality the video stream itself must be carefully engineered to cope with challenging wireless chan-nel conditions. This paper presents a scheme for doing this for H.264/AVC codec streaming across aWiMAX link. Packetization is an effective tool to govern error rates and, in the paper, source-codeddata-partitioning serves to allocate smaller packets to more important data. A packetization strategy isinsufficient in itself, as temporal error propagation should also be addressed by insertion of intra-codeddata. It may be necessary to include redundant packets when channel conditions worsen. The wholeshould be protected by application-layer rateless coding. Therefore, the contribution of the paper is acomplete scheme comprised of various protection measures aimed at robust IPTV streaming. Due to com-putational overheads, the scheme is aimed at the new generation of smartphones with GHz CPUs.

� 2014 Elsevier Inc. All rights reserved.

1. Introduction

Due to the flexibility of network delivery, Internet ProtocolTeleVision (IPTV) is attractive as an alternative to digital TV overterrestrial channels, though it may suffer from: delays due to con-gestion; and packet losses leading to fluctuations in video quality.It does represent an intrusion into personal privacy, as individualviewing habits can be tracked, although advertisers may benefitfrom this facility. Whereas terrestrial TV is limited to a fixed chan-nel broadcast schedule, i.e. linear TV, in contrast IPTV can offerpause-TV (when a partially-viewed program is cached for laterviewing), catch-up TV [1] and time-shifted TV (when a programis re-aired in a different time zone), as well as varieties of Video-on-Demand (VoD). It also opens up the possibility of interactiveTV and hybrid TV (a combination of terrestrial broadcasting withnetwork delivery).

A crucial aspect of engineering IPTV is configuring the videocodec in order to achieve good performance over a wireless linkfor flexible access. In this paper, we demonstrate effective codecconfiguration for wireless IPTV video streaming. In this way, thepaper also provides a reference for content providers to pre-encodetheir video using a suitable codec configuration for broadbandwireless networks, especially WiMAX. The paper also shows thatmultiple protection measures, as detailed in the paper, are neces-sary as a complete protection scheme for such an IPTV service

and we show how our scheme achieves robust delivery of video.As bursty error conditions present a major problem to compressedvideo stream, the research reports performance over broadbandwireless access links experiencing this type of error pattern. Thatproblem arises [2] because isolated errors are less harmful thanthe same number of errors appearing as a contiguous burst, asvideo compression relies on prediction from prior data. The paperwill be of interest to those charged with deploying a broadbandwireless IPTV service, as it contains an assessment of the currentprospects for wireless IPTV.

In this paper, H.264/Advanced Video Codec (AVC) [3] standardcodecs are assumed, as though the High Efficiency Video Codec(HEVC) was standardised at the beginning of 2013, experienceshows that it will take many years to be deployed, if indeed it isextensively used for streaming over wireless, given the absenceof error resilience and concealment features [4]. The absence oferror protection features may arise either because HEVC isintended to be used with Dynamic Adaptive Streaming over HTTP(DASH), a multi-streaming system for reliable TCP [5] or becauseHEVC’s considerable coding gain is achieved by abandoning Macro-blocks (MBs) for tree-structured Coding Units. HEVC is suitable for720p High Definition (HD) network delivery, where the up to 50%lower bitrates over H.264/AVC compensate for the high framerates, at least 50 frame/s at that resolution. Indeed, HEVC deliverychains have made quicker progress [6] than expected, though eval-uation has shown that, without persistent HTTP connections,streaming is subject to delays and interruptions when using DASH[7].

1494 L. Al-Jobouri et al. / J. Vis. Commun. Image R. 25 (2014) 1493–1506

The underlying transport protocol for HTTP is TCP, whichretransmits packets when acknowledgments fail. This implies thatthere is no need for the channel coding used in our paper. How-ever, over a wireless channel, TCP has difficulty distinguishingbetween packet loss due to congestion and packet loss due totransmission errors, implying that packets will be retransmittedeven if they may be again lost through a persistently poor wirelesschannel. DASH compensates for TCP’s behavior by allowing the cli-ent to control the bit-rate of the stream being downloaded (byselection between a set of streams with differing bit-rates). Unfor-tunately, during periods of heavy network congestion, DASHdefaults to predominantly TCP control rather than client control.Study of a Netflix application [8] showed service interruptions ofaround 300 s, which is larger than a typical Windows wired devicebuffer of 240 s. The risk to display interruption on an Androiddevice is still greater, as the typical buffer size is only 30 s. Fromthe evidence of the performance studies referred to in [8], the per-formance of DASH has until recently been relatively poorly under-stood. The main reason [9] that commercial companies haveimplemented such multi-stream systems is not delivered videoquality, which is difficult to monitor in a DASH-based system,but the compatibility of DASH with off-the-shelf web servers andexisting content-delivery networks. The simpler UDP transportsystem, which is the subject of this paper: reduces the risk of videoservice interruption; does not rely on network over-provisioning[9]; and as a result potentially represents a greener solution.

The authors’ robust scheme for streaming video over wirelessbroadband access has been simulated for IEEE 802.16 (WiMAX)systems [10]. WiMAX may have recently lost the technologicalcompetition with Long Term Evolution (LTE) [11] in developedWestern countries but it is still being deployed in rural areas[12] and in countries where 3G cellular phone coverage is poor,including Africa and the Middle East. WiMAX is also attractive[13] for backhauling from local IEEE 802.11 networks. Thoughthe authors’ scheme as a whole has been analysed in previousworks such as in [14,15], the individual codec settings have notbeen analysed in isolation. In fact, it has been suggested to the firstauthor that a very helpful service to the system developer commu-nity would be to analyze codec settings independently of eachother, which is what this paper sets out to do.

The authors’ scheme combines data-partitioning as a form oferror resilience [16] with the addition of Forward Error Correction(FEC) using rateless channel coding [17] at the application layer,together with retransmission of extra redundant data whenrequired. Retransmission is limited to one round to avoid accumu-lating latencies. The current authors have introduced into theirvideo streaming scheme various error resilience measures [18]that exist in the literature: different rates and types of Intra-Refresh (IR) Macroblocks (MBs) [19], Constrained Intra Prediction(CIP) [20] and where necessary, redundant data-partitioned Net-work Abstraction Layer (NAL) units (codec level packets) [21].The impact of each of these measures is analyzed in isolation fromeach other, for Constant Bit Rate (CBR) as well as Variable Bit Rate(VBR) streaming. CBR allows storage and bandwidth capacity to beplanned in advance, at a cost in video quality fluctuations. VBRenables greater compression efficiency relative to CBR, which iswhy it is generally used for disc storage. VBR can benefit fromtwo or even three-pass encoders, which are unsuitable for livevideo compression. The relative merits of CBR and VBR are furtherdiscussed in [22]. In general, we have noticed that many research-ers on similar topics have not explored the effect of video settingson their schemes and the authors believe that it is important toinvestigate this aspect of any protection scheme in order to under-stand the overall outcome of the scheme. In particular, it is impor-tant to determine how critical the video codec settings are to the

success of a protection scheme. Is the scheme robust to changesor should certain settings be set within a given range?

The remainder of this paper is organized as follows. The follow-ing Section adds further information on how IPTV can be deliveredto the consumer, as well as supplying background technical con-text. Section 3 selects research from the literature that revealsthe motivation behind the codec-based approach to video stream-ing used in this paper. Particular attention is paid to video stream-ing over WiMAX in that section. Section 4 then outlines theapproach taken to protect IPTV streams. As the main focus of thispaper is the configuration of the scheme rather than the schemeitself, the description is necessarily brief. Section 5 continues bydetailing how the video codec settings were modelled in order toprovide the evaluation that appears in Section 6. Section 7 summa-rizes the findings and rounds off with some observations aboutvideo streaming in this environment.

2. Context

Readers requiring further information on: IPTV, technical con-text of H.264/AVC data-partitioning, and/or Raptor codes can con-sult this Section.

2.1. IPTV delivery

There are two basic forms of IPTV: (1) That delivered overclosed or proprietary managed networks, including cable TV, oftento set-top boxes (STBs) that perform decoding and possibly decryp-tion; and (2) That delivered over the unmanaged public Internet,often displayed directly on desktop or laptop computer screens.The latter may be called Web or Internet TV [23] or latterlyOver-the-Top (OTT) TV, as a way of distinguishing it from the ser-vice that the telecommunications companies hoped to challengeterrestrial broadcasters with by means of a high-quality servicedelivered over closed networks. Increasingly the industry trend,apart from cable TV, is towards the latter model of IPTV, i.e. OTTTV. There is no place for STBs when mobile IPTV [24] is delivered,hence, our interest in OTT TV. The two forms of IPTV may differ in:the picture quality of the stream offered (lower for unmanaged);the amount of content available (higher for unmanaged); the for-mats that the video is compressed to (the need to update the STBis an issue); the way content is secured (unmanaged deliverymay not be encrypted or employ selective encryption [25]instead); and the frame resolutions presented (lower resolutionsfor unmanaged delivery). A flexible way to deliver IPTV is to set-up (and tear-down) connections [26] using the Real-Time Stream-ing Protocol (RTSP) (over reliable TCP transport), with subsequentdata delivery using the Real-time Transport Protocol (RTP) overUDP transport. The Real-time Control Protocol (RTCP) can alsoallow end-to-end feedback to the IPTV server to control the bitrate. RTSP also can support so-called trick mode functionality suchas rewind and pause.

To allow unmanaged delivery, OTT TV, to compete with Stan-dard TV (SDTV) Quality-of-Experience (QoE), TV and video materialcan be locally cached [27], reducing latency to the access link,which in our paper is represented by a WiMAX base to subscriberstation link. Whether streaming over managed or unmanaged net-works, packets will be lost at the access network. This is the case,whether over broadband wireless or xDSL (Digital Subscriber Line),as the latter also suffers from burst errors [28]. Prior to that,whether managed or unmanaged networks are in use, videostreams are aggregated over high-capacity optical networks, suchas in Swisscom’s Bluewin IPTV service [29]. Only Passive OpticalNetworks (PONs) can reduce the error rates at the access link but

+ + + + + + ++ +++

codingOuter

LTcoding

Redundantnodes

k

m

Fig. 2. Raptor encoding scheme.

L. Al-Jobouri et al. / J. Vis. Commun. Image R. 25 (2014) 1493–1506 1495

currently PONs are not widely deployed. The approach adopted inthis paper is potentially applicable across a range of access linktypes, as it provides protection against error bursts and causes onlymoderate delays. If a multicast service is required, it can also bereadily adapted from its unicast form by dropping acknowledg-ments to the base station and increasing the level of FEC. This pos-sibility was tested by us in [30] and, therefore, this paper confinesitself to unicast delivery.

2.2. H.264 codec

Features of the H.264 codec are used in this work. H.264/Advanced Video Coding (AVC) standardization was initiated bythe Video Coding Experts Group (VCEG), which is a working groupof the International Telecommunication Union (ITU-T). The JointVideo Team carried out the final task of developing H.264/AVC[3], a co-operative effort of both Moving Picture Expert GroupMPEG and VCEG in 2003. Fig. 1 shows the video frame structure.A video-frame in H.264/AVC is divided into MBs [31], which arethe basic unit for motion prediction. Each MB can be further sub-divided into blocks that are transform-coded in order to de-corre-late the data. For transmission, a video frame can be split intoslices, each of which occupies a network packet. When data-parti-tioning is enabled, each slice can be further sub-divided in up tothree partitions, data-partitions-A, -B and -C, each of which canoccupy a separate network packet. It will be necessary to split anetwork packet if its size exceeds the Maximum Transport Unit(MTU) of the network.

2.3. Rateless channel coding

We selected Raptor codes [32] as our rateless code. Fig. 2 is asimple diagram of how a Raptor encoder works. k representssource symbols and m represents received symbols. A systematicRaptor code is arrived at [32] by first applying the inverse of theinner code to the first k symbols before the outer pre-coding step.Systematic channel codes separate the redundant coding data fromthe information data, allowing the information data to be passeddirectly to the source code decoder without channel decoding ifno errors are detected (usually by means of checksums).

For a rateless code, even if there were no channel errors, there isa very small probability that the decoding will fail (the probabilityconverges to zero in polynomial time in the number of input sym-bols), which failure can be modeled statistically by the schemeintroduced in [28]. In compensation, these codes are completelyflexible, have a theoretical linear decode computational complex-ity, and generally their overhead is considerably reduced comparedto fixed erasure codes. Furthermore, if the packets are pre-encodedwith an inner code, as in Raptor codes to achieve a systematic code,a weakened Luby Transform (LT) [33], with reduced computationalcomplexity compared to the original LT [28], can be applied to thesymbols and their redundant symbols.

In [32], an implementation of Raptor code is reported decodingat several gigabits per second on a 2.4 GHz Intel Xeon processor. By

Macro block

Frame

Frameslices

DP-A DP-B DP-C

DP-A DP-B DP-C

DP-A DP-B DP-C

Fig. 1. Video frame structure, DP-A = data-partition-A and similarly for DP-B andDP-C.

way of comparison, in 2013 smartphones generally had a four-coreprocessor running at around 1.5 GHz, though high-end smart-phones had 64-bit processor clock speeds of around 2.2 GHz. As ameasure of its acceptability for mobile multimedia, Raptor codinghas been adopted for two wireless standards, Digital Video Broad-casting (DVB) for handheld devices (DVB-H) [34] and Third Gener-ation Partnership Project (3GPP) Multimedia Broadcast/MulticastServices (MBMSs) [35], though neither uses the codes for unicastor accepts repair packet requests, as occurs in this paper. Impor-tantly though, using Raptor code in these standards at the block-level rather than the byte-level, as herein, can significantly impedeits throughput.

Notice that in these standards [34,35], application-layer Raptorcoding is applied at the packet-erasure correction level, whilephysical-layer channel coding occurs at the bit-level. If physical-layer correction fails according to checksum detection [36] then apacket is marked as an erasure to be corrected at the applicationlayer. However in our paper, upon physical-layer correction failure,packets are not marked as erased but their data are passed to theapplication layer for byte-level Raptor code correction. A relatedprocedure is proposed for mobile phones in [37], when, after phys-ical-layer correction failure, partially decoded packets are passedto the application layer for block-level correction.

In a Raptor code, a belief-propagation (BP) algorithm can beemployed to decode the inner LT code. A BP implementation in[38] treated decoding as a soft decoding task rather than correctionof erasures. The alternative is a hardware implementation of Rap-tor’s inner code using Gaussian elimination. Gaussian eliminationhas a total complexity of order O(nk2) to solve a linear equationsystem with n output symbols and k unknowns (the original datasymbols), resulting in a linear per data symbol complexity ofO(nk). The outer code in [28] is selected as a Low-Density Parity-check Code (LDPC), which is also an optional physical-layer codein IEEE 802.11n, either using Gaussian elimination or in [39] a BPalgorithm in hardware. As Gaussian elimination is at the heart ofRaptor decoding, optimizations of the algorithm intended for hard-ware implementation on mobile phones [40] are available.

3. Related research

A number of research papers have considered some aspects ofthe video streaming method used by us. In [41] a packetizationmethod was presented for robust H.264 video transmission overan IEEE 802.11 wireless local area network (WLAN) configured ashome network. Video robustness was enhanced by using smallNAL units and by retrieving possible error-free IP packets fromthe received MAC frame. An aggregation scheme with a recoverymechanism was deployed and evaluated via simulation. For fixedphysical-layer resources, the system provided a 2.5-dB gain invideo quality (PSNR) compared to making no NAL packetizationadjustments for similar throughput efficiency. Equally an 80%

1496 L. Al-Jobouri et al. / J. Vis. Commun. Image R. 25 (2014) 1493–1506

improvement in throughput was achieved for a similar video qual-ity. However, data partitioning as a way of varying NAL unit sizeswas not used. The work in [41] was tailored to IEEE 802.11 WLANsusing the Distributed Coordination Function (DCF) for negotiationof access to the wireless channel. As a result, data throughputwas strongly influenced by the data frame size. Broadband wirelesssystems tend to employ centralized packet scheduling, typicallythrough Time Division Multiple Access (TDMA) with pre-config-ured frame sizes into which packets from each node (subscriberstation) are packed. Consequently, data partitioning for such sys-tems is a more appropriate way of managing packetization issues.

Research in [42] used Forward Error Correction (FEC) and Auto-matic Repeat re-Quest (ARQ) to support streaming over WiMAX,exploiting features of the WiMAX standard. In particular, channelstate information held at WiMAX stations served to dynamicallyconstruct the MAC Protocol Data Unit (MPDU). The size of theseunits was thus determined such that the packet dropping probabil-ity was minimized without compromising goodput. Simulationresults showed that the ARQ-enabled adaptive algorithm wasalways better than the non-adaptive algorithm. The scheme in[42] anticipated our work but no particular application was cateredfor in [42] other than it should be a real-time streaming one and,indeed, the authors of [42] were particularly concerned with voiceover IP (refer to [43]). Because our work is specialized for videostreaming it can apply aspects of video source coding to streaming.We also utilize adaptive channel coding rather than the blockcodes of [42]. While we utilize ARQ to request additional FEC dataaccording to channel conditions, ARQ in [42] was employed tochange the MPDU size, as the possibility of altering the channelcoding rate was not available to the authors of [42].

In [16] the researchers compared non-scalable video codingwith data partitioning using H.264/AVC under similar applicationand channel constructions for conversational applications overmobile channels. The experimental results showed that by usingthe data-partitioning scheme the number of entirely lost framescan be lowered and the probability of poor-quality decoded videocan be reduced. In the data-partitioning scheme of [16], differentialprotection was achieved by selecting from a set of discrete channelcoding rates, through punctured convolutional codes. However, inorder to determine the protection level, an optimization procedurebecame necessary to minimize potential distortion. This proceduredepended on the quantization parameter (QP) and the coding ratefor each partition. The wireless channel characteristics also had tobe known in advance by the encoder. However, leaving aside thecomputational complexity of the optimization search in [16], thereis another key difference between the scheme of [16] and thispaper. In [16] no feedback occurred, so that it was not possible torequest additional redundant data. In fact, when using puncturedconvolutional codes in [16] (rather than the rateless codes usedherein) it was not possible to generate additional redundant data.

Data-partitioning can be viewed as a simplified form of SNR orquality layering. Extended quality layering can also be applied tovideo streaming across WiMAX. In [44], adaptive multicast stream-ing was proposed for WiMAX using the Scalable Video Coding(SVC) extension for H.264. WiMAX channel conditions were mon-itored in order to vary the bitrate accordingly. Unfortunately, thesubsequent decision of the JVT standardization body for H.264/AVC not to support fine-grained scalability (FGS) implied it willbe harder to respond to channel volatility in the way proposed in[44], which is one reason why we do not employ SVC and a numberof other issues surrounding SVC are mentioned in the rest of thisparagraph. Other work concerned with video streaming overWiMAX links has also investigated: combining scalable video withmulti-connections in [45]; and comparing [46] H.264/SVC withH.264/AVC for WiMAX. However, the data-dependencies betweenlayers in H.264/SVC medium-grained scalability remain a concern.

Unlike in FGS, enhancement layer packets may successfully arrivebut not be able to be reconstructed if key pictures also fail to arrive.Besides, for commercial one-way streaming, simulcast is nowlikely to be preferred to H.264/SVC for the reasons outlined in[47]. In [47], it was found that the extra overhead from sendingan SVC stream compared to an H.264/AVC stream meant that thecost of bandwidth consumption outweighed the reduced storagecost of SVC once more than 64 sessions had occurred (assuming16 simulcast streams or 16 video layers per session). In anothercomparison [48], it was proposed that scalable video with unequalerror protection cannot provide any advantage over H.264/AVCwith equal error protection in a wireless environment, due to theoverhead of scalable video coding compared to that of single-layercoding.

4. Outline of streaming system

This Section outlines an effective video streaming system forWiMAX that provides error resilience through source-code datapartitioning [16] and which works without the need to apply priv-ileged protection to the high-priority partitions. As already men-tion in Section 1, the main point of this paper is to drawattention to the importance of correctly configuring the videocodec parameters. Therefore, this Section provides a basic outlineonly and the interested reader is referred to the authors’ otherworks, such as [14] and [15], for more details and variants of ourstreaming system.

The H.264/AVC QP is set by us in such a way that temporally-coded texture data occupies a larger part of a frame’s data thanthat occupied by data in each of the other two partitions. If thistexture data should be dropped, it can be replaced more easilythrough error concealment at the decoder than data in the othertwo partitions. Error concealment (backward error correction)[49] is a non-normative feature [50] of H.26/AVC that neverthelessis present in the H.264/AVC JM reference code (found at http://iphome.hhi.de/suehring/tml/). The texture data in partition-C ispacketized in a WiMAX MAC Service Data Unit (MSDU), within aMAC Protocol Data Unit (MPDU) [10,51].

In contrast to the authors’ counter-intuitive approach, the intu-itive approach is to give special protection to the partition-A,which for predictively-coded frames includes motion vectors aswell as other important parameters. Partition-B, bearing intra-coded MBs, may also be given special protection, as it contains spa-tially-coded data whenever suitable references in other videoframes are not available. As an example of the Unequal Error Pro-tection (UEP) approach, in [52] hierarchical modulation wasemployed to favor those partitions with more important data forthe reconstruction of the video frame. For example, partition-Amotion vectors can be used for motion copy error concealmentwhen partition-C data are lost and, therefore, the intuitiveapproach is to protect partition-A (possibly combined with parti-tion-B). Readers should consider for themselves from the evalua-tion of this paper whether UEP measures are necessary.

Though no special protection is given, it is still necessary to pro-tect the bit-stream (without privileging partitions A and B) againstthe risk of packet loss. This was achieved through Equal Error Pro-tection (EEP) of FEC provision. Application-layer rateless coding[53] was selected for its flexibility and its linear computationalcomplexity at both the decoder and the encoder. To avoid longlatencies, which would occur if packet-level FEC were to beapplied, redundant data were added to the packets themselves,treating the bytes within each packet as the data symbols. Againto reduce latency, just a single Automatic Repeat reQuest (ARQ)was made if the available data were insufficient to reconstruct acorrupted packet. Additional rateless redundant data were added

L. Al-Jobouri et al. / J. Vis. Commun. Image R. 25 (2014) 1493–1506 1497

to the next available packet to be transmitted, according to theamount calculated in the way specified by us elsewhere [14].

In an extension to the basic IPTV streaming scheme, we alsoadded redundant NAL unit packets. A discussion of alternativeredundant packet schemes is postponed to Section 6.4.

5. Simulation model

This Section details how we set about modeling the WiMAXsystem for the evaluation of Section 6.

5.1. Simulation system

Fig. 3 shows the overall data flow of the simulation system. Rawvideo (.YUV) is encoded by H.264/AVC into compressed form. Thecompressed file is passed for later extraction of any dropped pack-ets. At the same time, a video trace file is generated, which willbecome an input into the ns-2 simulator [54]. The trace file con-tains the size of each video packet and the transmission schedule.However, the trace file does not contain any video data. After sim-ulation using the wireless channel model, the simulator outputs alist of sent and dropped packets. This is used to filter from the ori-ginal compressed data file any data that did not arrive at the des-tination. The file is subsequently decoded by the H.264/AVCdecoder. The decoder outputs a raw video file in YUV (YCbCr) for-mat. This file would normally be displayed but, for the simulation,it is compared with the original raw input video. This allows theobjective video quality to be calculated as Peak Signal-to-NoiseRatio (PSNR), through a pixel-by-pixel comparison.

Two video clips with different source-coding characteristicswere employed in the tests in order to judge the dependency ofthe results upon video source-coding complexity. The first testsequence was Paris, which is a studio scene with two upper bodyimages of presenters and moderate motion. The background is ofmoderate to high spatial complexity leading to larger slices. Theother test sequence was Football, which has rapid movementsand consequently has high temporal coding complexity. Bothsequences were encoded at Common Intermediate Format (CIF)(352 � 288 pixel/picture). CIF resolution was used for ready com-

Video file .YUV

JM 14.2 H.264/AVC

encoder

Generate video trace

Received trace packets

Sent trace packets

Checking dropped packets

Filtered video stream

Reconstructed YUV file

PSNR calculation

JM 14.2 H.264/AVC

decoder

Sending node

Receiving node

Networkchannel model

Fig. 3. Simulation system.

parison with the prior work of others on video communication tomobile devices.

5.2. Wireless channel model

The Gilbert–Elliott channel mode employed in this work hasbeen used by researchers in the wireless field [55,56] because ofits ability to model error burst patterns as experienced at the recei-ver. This channel model was introduced into the ns-2 event-drivensimulator [54]. The Gilbert–Elliott channel model itself is a two-state Markov chain. It is based on: good and bad states; the prob-abilities of these states; and the probabilities of the transitionstates between them. In the case of the bad state, losses happenwith higher probability while in the good state losses happen withlower probability. PGG refers to the probability of being in the goodstate and PGB is the probability of a transition from the good stateto the bad state. PBB is the probability of being in the bad state andPBG refers to the probability of a transition from the bad to goodstate. PGG (PBB) can be interpreted as the probability of remainingin the good (bad) state, given that the previous state was good(bad). Conversely, PGB represents the probability that given thatthe previous state was good, a transition is made from the goodto the bad state. By the law of total probability, all probabilitiessum to one (certainty). Therefore, we have PGG + PGB = 1, resultingin (1). A similar argument for the bad state leads to (2).

PGG ¼ 1� PGB ð1Þ

PBB ¼ 1� PBG ð2Þ

For the stochastic process to remain stationary in time,

pGpGB ¼ pBpBG ð3Þ

where pG and pB are the steady state probabilities of being in a goodor bad state respectively. Again by the law of total probability,pB = 1 � pG. Substituting this expression for pB into (3) easily leadsto:

pG ¼pBG

pBG þ pGBð4Þ

Similarly, pG = 1 � pB. Substituting this expression for pG into(3) easily leads to:

pB ¼pGB

pBG þ pGBð5Þ

The Gilbert–Elliott model good and bad states have their ownerror distributions that are independent of the process of arrivingor leaving those states, i.e. forming a Hidden Markov Model. Sup-pose the probability of packet loss is pG and pB in the good andbad states respectively. Then the average packet loss rate producedby a Gilbert–Elliott channel is given in (6) by the usual expressionfor the expectation of a probability distribution.

p ¼ pGpG þ pBpB ð6Þ

To model the effect of slow fading at the packet-level, as the Gil-bert–Elliott model’s parameters, the PGG was set to 0.96,PBB = 0.95, PG = 0.01 and PB = 0.02. Additionally, it is still possiblefor a packet not to be dropped in the channel but nonetheless to becorrupted through the effect of fast fading (or other sources ofnoise and interference). This byte-level corruption was modelledby the second Gilbert–Elliott model, with the same parameters(applied at the byte level) as that of the packet-level model exceptthat PB (now probability of byte loss) was increased to 0.165. Effec-tively, this second model emulates fast fading between good andbad conditions.

The slow fading parameters in the Gilbert–Elliott model wereestablished empirically by a long series of trial simulations aimed

Table 1IEEE 802.16 (WiMAX) parameter settings.

Parameter Value

PHY OFDMFrequency band 5 GHzBandwidth capacity 10 MHzDuplexing mode TDDFrame length 5 msMax. packet length 1024 BRaw data rate 10.67 MbpsIFFT size 1024Modulation 16-QAM 1/2Guard band ratio 1/8DL/UL ratio 3:1Channel model Gilbert–ElliottSS transmit power 245 mWBS transmit power 20 WApprox. range to SS 1 kmAntenna type Omni-directionalAntenna gains 0 dBDSS antenna height 1.2 mBS antenna height 30

PHY = Physical layer, IFFT = Inverse Fast Fourier Transform, DL/UL = Downlink/Uplink, QAM = Quadrature Amplitude Modulation, SS = Subscriber Station,BS = Base Station, TDD = Time Division Duplex, OFDM = Orthogonal FrequencyDivision Multiplexing.

1498 L. Al-Jobouri et al. / J. Vis. Commun. Image R. 25 (2014) 1493–1506

at establishing packet loss rates (PLRs) that were known fromexperience to be critical to the acceptable reconstruction of a videostream by a decoder. Specifically, PLRs below 5% can be toleratedwithout protection measures, while above 10% it is difficult forerror resilience measures and/or channel coding to reconstruct avideo sequence. This observation can be generalized in the abovedual Gilbert–Elliott model to achieve data loss rates in the range5–10%. It is only realistic to specify a data loss range because ofthe statistically nature of the model’s behaviour. In this paper,we have presented results in the data loss range that was selectedas critical. It is perfectly possible for a link to be benign and no datalosses to occur but clearly one should engineer a streaming solu-tion to cope with critical conditions. As the channel coding protec-tion in this paper is adaptive, the problem of transmittingredundant overhead in a benign channel is reduced. As extraredundant data cannot be transmitted more than once, the possi-bility of attempting video repair when no amount of protectionwill work is also avoided. Fig. 4 is an example from preliminarytests in which the Gilbert–Elliott PB parameter is systematicallyvaried while the other Gilbert–Elliott parameters were set asabove. This allows the relationship to the PLR for four referencevideo clips to be found. By observation, the relationship is approx-imately linear.

In [57] for an IEEE 802.15.4 wireless link, the authors takeextensive measurements of a bursty link to arrive at a metric forburstiness, which is then found to be a generalisation of the burs-tiness in a Gilbert–Elliott model. However, the Gilbert–Elliottmodel differed from ours in representing the hidden bad state toresult in the loss of all packets and likewise the good state to resultin all packets surviving. An interesting observation of the authorsof [57] is that burstiness tends to occur when a mobile device isat the limits of the range of a radio link, so that small changes inthe physical channel cause the signal to drop below the receiver’ssensitivity. In [56], the estimation of model parameters from chan-nel measurements is also treated for wired Internet examples.

5.3. WiMAX simulation configuration

The physical (PHY) layer settings selected for WiMAX simula-tion on ns-2 [58] are given in Table 1. The antenna was modelledfor comparison purposes as a half-wavelength dipole. The TimeDivision Duplex (TDD) frame length was set to 5 ms, as this isthe only value supported by the WiMAX Forum [59]. As mentionedunder modulation, the physical-layer FEC convolutional code rate

0

5

10

15

20

25

0.03 0.06 0.09 0.12 0.15 0.18 0.21 0.24 0.27 0.3PB

Pack

et lo

st %

Foreman

Car Phone

Highway

Paris

Fig. 4. Example plot showing the relationship between Gilbert–Elliott PB parameterand packet loss rate for several CIF video sequences at 30 Hz, with data-partitioningapplied.

was ½. Video was transmitted over the downlink with UDP trans-port (see Section 2.1). In order to introduce sources of traffic con-gestion, an always available FTP (File Transfer Protocol) sourcewas modelled with TCP transport to a Subscriber Station (SS) fromthe WiMAX Base Station (BS). Likewise a CBR source with packetsize of 1000 byte and inter-packet gap of 0.03 s was downloadedto another SS.

6. Evaluation of codec settings

This Section evaluates in turn the codec configurations intro-duced in Section 1.

6.1. Intra-Refresh Macroblocks

As previously mentioned, in H.264/AVC data partitioning[16,60], motion vectors (MVs) are packed into partition-A bearingNAL units, allowing motion copy error concealment at the decoderto partially reconstruct a picture, despite missing partition-C NALunits (containing quantized transform coefficient residuals). Parti-tion-B NAL units contain intra-coded (spatially encoded) MBs,which are substituted for inter-coded MBs according to encoderimplementation (only the decoder input format is standardizedin H.264/AVC). Therefore, when Intra-Refresh (IR) MBs areincluded alongside naturally intra-encoded MBs, partition-B slicesgrow in size. This means that data-partitioned video compressionprovides a convenient way to examine the effect of variousamounts of IR MB provision. Once the H.264/AVC encoder hasformed a NAL unit, it can also provide a RTP header prior to encap-sulation by IP/UDP network protocol headers.

A point to note is the different way that random IR MBs arespecified in the H.264/AVC JM 14.2 implementation compared tothat of cyclic IR line intra update. In random IR MB, a maximumpercentage of IR can be specified, which percentage includesalready encoded IR MB. If the given quota of IR MB is already lar-gely occupied by naturally encoded MBs (those encoded to covernewly revealed objects or to improve the quality of the video upto a given bit budget or for some other encoder-dependent reason),then only a small amount of extra randomly inserted MBs will beadded. In contrast, if a line of IR MBs is inserted then these MBs

Table 2Mean size of different partitions in bytes for Football at various QPs.

QP 2% Intra-refresh MB 5% Intra-refresh MB 6% Intra refresh MB

A B C A B C A B C

20 1842 2678 3889 1845 2767 3867 1846 2810 385025 1687 1697 2533 1690 1763 2511 1696 1793 250230 1459 1047 1496 1463 1082 1482 1467 1098 147935 1117 572 688 1120 595 682 1123 604 681

QP 25% Intra-refresh MB MB line intra update

A B C A B C

20 1893 3450 3669 1885 3385 368325 1746 2216 2379 3683 2160 240030 1505 1346 1405 1498 1312 141435 1146 729 646 1143 716 652

L. Al-Jobouri et al. / J. Vis. Commun. Image R. 25 (2014) 1493–1506 1499

are added in addition to those intra-coded MBs that have alreadybeen included by the encoder, as shown in Fig. 5.

Football was VBR encoded with a Group-of-Pictures (GoP) struc-ture of IPPP. . . at 30 frame/s. (Refer forward to Section 6.3 to seethe tests that led to choice of the more favourable IPPP. . . GoPstructure.) From Table 2, it is apparent that, as the percentage ofIR MBs is increased, the size in bytes of partition-B increases forthe same QP. Because more MBs are assigned to partition-B, thesize of partition-C reduces. And because of the large amount of nat-urally-encoded intra MBs, this effect is gradual until 25% of randomIR MBs are added. 25% of random IR MBs is shown in Table 2, asthat amount approximately corresponds to the total partition-Bsize if cyclic line intra update is turned on instead (with approxi-mately the same number of MBs).

Fig. 6 illustrates the range of metrics we extracted from ourcodec configuration tests. However, it should be noted that, inFig. 6, increasing the provision of IR MBs from 2% to 5%, then to6% and finally 25% (in the case of MB line intra update), increasesthe throughput and, hence, the bandwidth requirements in respectto co-existing traffic. The 25% IR commitment tested is large inpractice, due to the coding inefficiency of spatial reference coding.It results in larger packets and the size of packets is the mostimportant factor affecting the percentage of dropped packets, asis also evident from the decrease in dropped packet percentages,Fig. 6a, when the QP increases (and video quality decreases). FromFig. 6a and c, higher-quality video (QP = 20, 25) benefits from ran-dom insertion (and this effect is reversed for low-quality video).From the objective video quality (PSNR) resulting, Fig. 6c, it canbe seen that reducing the IR MB percentage to 2% actuallyimproves the PSNR at QPs 30 and 35. From this one can concludethat there is little if anything to be gained by raising the IR percent-age. This is not surprising, as at lower QPs the encoder has a greaterbit budget and, hence, can include more naturally-encoded intraMBs. In addition, the main effect of reducing the percentage of IRMBs is that the size of partition-B-bearing packets is reduced. Inturn, this makes these packets less likely to be affected by channelconditions, especially burst errors arising from fast fading. Conse-quently, more partition-B packets survive intact, again causing arelative gain in video quality.

Packet end-to-end delay, Fig. 6d, is the mean delay of thosepackets unaffected by channel conditions. The results show thatthis is generally small in duration, though with a tendency toincrease due to the propagation delay of larger packets at lower

Fig. 5. Differences between random intra-refresh MBs (upper fram

QPs. Significantly, a lower percentage of IR MBs, in addition to bet-ter quality video in many cases, actually results in lower delay.

Turning to corrupted packets, there are larger percentages ofcorrupted packets at higher QPs, Fig. 6b. These are packets thathave not been repaired completely by the adaptive channel codingscheme. Because of the additional transmission time, the meanend-to-end delay of corrupted packets is higher than other packets,Fig. 6e. In fact, it is the extent of the delay that is the main contri-bution of corrupted packets to the quality-of-service. However,though the percentage of corrupted packets increases at higherQPs, the packet sizes decrease due to the reduced coding efficiency.Smaller packets take less time to re-transmit compensating for theincrease in the number of corrupted packets.

6.2. Constrained intra prediction setting

In order to decode partition-B and -C, the decoder must knowthe location from which each MB was predicted, which impliesthat partitions B and C cannot be reconstructed if partition-A islost. Though partition-A is independent of partitions B and C, CIPshould be set in the codec configuration [20] to make partition-Bindependent of partition-C. (Though reference [20] refers to a pro-posal to add constrained inter prediction to H.264/AVC, it alsodescribes constrained intra-prediction, which is already a part ofthe codec standard.) By setting this option, partition-B MBs areno longer predicted from neighboring inter-coded MBs. This isbecause the prediction residuals from neighboring inter-coded

es with 6%) and MB cyclic line intra update (lower frames).

0

6

12

18

24

30

20 25 30 35

Dro

pped

pac

kets

%

QP

2% Intra refresh MB5% Intra refresh MB6% Intra refresh MBMB Line Intra Update

0

6

12

18

24

30

20 25 30 35

Cor

rupt

ed p

acke

ts %

QP

2% Intra refresh MB

5% Intra refresh MB

6% Intra refresh MB

MB line Intra Update

(a) (b)

0

8

16

24

32

40

20 25 30 35

Mea

n PS

NR

(dB

)

QP

2% Intra refresh MB5% Intra refresh MB6% Intra refresh MBMB Line Intra Update

0

0.03

0.06

0.09

0.12

20 25 30 35Pack

et m

ean

end -

to-e

nd d

elay

(s)

QP

2% Intra refresh MB5% Intra refresh MB6% Intra refresh MBMB Line Intra Update

(c) (d)

0

0.03

0.06

0.09

0.12

0.15

20 25 30 35

Cor

rupt

ed p

acke

ts m

ean

dela

y (s

)

QP

2% Intra refresh MB

5% Intra refresh MB

6% Intra refresh MB

MB Line Intra Update

(e)

Fig. 6. Mean performance metrics for Football with 2%, 5%, 6% IR MBs and MB line intra update.

Table 3Paris video sequence: Mean NAL unit size in bytes according to partition type andvideo quality.

QP Without CIP (B) With CIP (B)

A B C A B C

20 740 495 4097 738 504 415425 601 336 2058 601 356 209330 473 210 838 473 238 85735 331 124 281 330 149 291

1500 L. Al-Jobouri et al. / J. Vis. Commun. Image R. 25 (2014) 1493–1506

MBs reside in partition-C and cannot be accessed by the decoder ifa partition-C packet is lost. There is a by-product of increasingpacket sizes due to a reduction in compression efficiency but theincrease in size may be justified in error-prone environments.

The two video clips (Paris and Football) were VBR encoded atCIF, with a GOP structure of IPPP(for choice of GoP structure againrefer forward to Section 6.3) . . . at 30 Hz and with 5% IR MBs(choosing a lower percentage as a result of the tests in Section 6.1).Table 3 presents NAL unit sizes for the Paris sequence with andwithout CIP, showing the increase in partition-B and -C sizes,which results from the loss in encoding efficiency. Notice thatany NAL units that are above the maximum packet size of 1024B in Table 3 are later constrained by the encoder to the MTU whenforming an RTP packet prior to encapsulation by network headers.The reason for this precaution is to avoid these larger NAL unitsbeing segmented at the link layer, avoiding the possible separationof header information from NAL data.

At lower QPs, i.e. higher-quality video, the relative size of parti-tion-C NAL units means that the more important partition-A and -Bpackets are less likely to suffer channel error. However, the largerpacket sizes mean that congestion may have more of an impact,because longer packets take longer to transmit and free the

channel. At higher QPs, the advantage of differential packet sizesis lost but the generally smaller packet sizes compensate to someextent. There is also a small growth in partition-B and -C packetsize when CIP is turned on.

From Table 4 for Football one sees that though the relative rank-ing of sizes between the partition types is similar, the actual sizesare larger than those for Paris (see Table 3). The larger sizes are dueto the temporal coding complexity of Football. For high QP, the rel-atively larger size of partition-A NAL units compared to the otherpartitions NAL units may create a problem, as it does not result

Table 4Football video sequence: Mean NAL unit size in bytes according to partition type andvideo quality.

QP Without CIP (B) With CIP (B)

A B C A B C

20 1845 2767 3867 1845 2870 432625 1690 1763 2511 1681 1845 287330 1463 1082 1482 1431 1080 175435 1120 595 682 1092 494 925

0

7

14

21

28

35

20 25 30 35

Cor

rupt

ed p

acke

ts %

QP

Paris with CIP

Paris without CIP Football with CIP

Football without CIP

Fig. 8. Paris and Football video sequences: Protection scheme corrupted packets,with and without CIP.

0

5

10

15

20

25

30

20 25 30 35

Mea

n PS

NR

(dB

)

QP

Paris with CIP

Paris without CIP

Football with CIP

Football without CIP

Fig. 9. Paris and Football video sequences: Protection scheme video quality (PSNR),with and without CIP.

L. Al-Jobouri et al. / J. Vis. Commun. Image R. 25 (2014) 1493–1506 1501

in a relatively reduced risk of loss of packets bearing partition-ANAL units. Also of concern is the number of NAL units that areabove the maximum packet size, causing more than one packetto be sent.

Fig. 7 shows the effect of the various schemes on packet dropswhen streaming Paris and Football. The Fig. assesses the effect ofturning on CIP. As turning on CIP can increase packet sizes, dueto less efficient source coding, there is a risk of more channelerrors. From Fig. 7, the larger packet drop rates at QP = 20 for Foot-ball will have a significant effect on the video quality. However, thepacket size changes with and without CIP have little effect on thepacket drop rate in the case of the less active Paris. In fact, the pat-tern of channel errors can actually result in a decrease in packetdrops when Paris is streamed with CIP turned on.

Fig. 8 shows the numbers of corrupted packets arising fromsimulated fast fading. Nevertheless, the effect of the corruptedpackets on video quality only occurs if a packet cannot be recon-structed after application of the adaptive retransmission scheme.That is if the packet bearing the repair data is itself corrupted ordropped, then the original corrupted packet cannot be recon-structed. Fig. 8 also shows that turning CIP on generally resultsin more corrupt packets and, hence, additional delay due toretransmissions.

Examining Fig. 9 for the resulting PSNR, one can see that thevideo quality is generally below 31 dB, and, hence, would probablybe ranked as ‘fair’ (not ‘good’) according to the ITU P.800’s recom-mendation, originally intended for subjective testing but used in anobjective as guide to quality in papers such as [61]. This video qual-ity is similar in general terms to the video quality reported in [61],though for multicast streaming without feedback. Nevertheless,one observes from Fig. 9 that, in all cases, the inclusion of CIP inour scheme results in improved objective video quality at all QPstested and both for Paris and the more active Football sequence.Therefore, CIP is to be recommended.

As previously remarked, the impact of corrupted packets, giventhe inclusion of retransmitted additional redundant data, is largely

0

2

4

6

8

10

20 25 30 35

Dro

pped

pac

kets

%

QP

Paris with CIP Paris without CIP Football with CIP Football without CIP

Fig. 7. Paris and Football sequences: Protection scheme packet drops, with andwithout CIP.

seen in additional delay. There is an approximate doubling in perpacket delay between the total end-to-end delay for normal packettransmission, Fig. 10, and that of corrupted packets, Fig. 11. Never-theless, the delays remain in the tens of millisecond range, exceptfor when QP = 20, i.e. broadcast quality video, when large packetsizes imply longer transmission times. This type of delay range isacceptable even for interactive applications but may create unac-ceptable latency if it forms part of a longer network path.

6.3. Group of Pictures structure

Attention should also be given to the Group of Pictures (GoP)structure. For the purposes of comparison, we tested two different

0

0.003

0.006

0.009

0.012

0.015

20 25 30 35

Pack

et m

ean

end-

to-e

nd d

elay

(s)

QP

Paris with CIP

Paris without CIP

Football with

Football without

CIP

CIP

Fig. 10. Paris and Football video sequences: Protection scheme mean end-to-endpacket delay, with and without CIP.

0

0.005

0.01

0.015

0.02

0.025

20 25 30 35

Cor

rupt

ed p

acke

ts m

ean

dela

y (s

)

QP

Paris with CIP

Paris without CIP

Football with CIP

Football without CIP

Fig. 11. Paris and Football video sequences: Protection scheme mean delay ofcorrupted packets, with and without CIP.

Table 5Mean P-picture size (bytes) according to QP for two different GoP structures.

QP Football Paris

IPPP. . . IBBP. . . IPPP. . . IBBP. . .

20 8905 15,520 5102 759025 6301 10,311 2824 424930 4185 6381 1398 223835 2444 3756 647 1100

0

7

14

21

28

35

20 25 30 35

Mea

n PS

NR

(dB

)

QP

Football IPPP...

Football IBBP...

Paris IPPP...

Paris IBBP...

1502 L. Al-Jobouri et al. / J. Vis. Commun. Image R. 25 (2014) 1493–1506

GoP structures: IPPP. . . (i.e. one initial I-picture and all P) andIBBPBBP. . . (i.e. insertion of bi-predictive B-pictures for greatercoding efficiency but still with one initial I-picture). Before startingthose tests, we examined the effect of the two different GoP struc-ture types on the sizes of video frames. One potential impact of theGoP structure is that B-pictures increase coding efficiency at a costin coding complexity due to the need to reference at least twoother pictures. Unfortunately, with the inclusion of B-pictures,the mean size of P-pictures increases, as a result of the increasedreference distance between P-pictures. For example, for QP = 20,the IBBPBBP. . . mean P-picture size is as high as 15 kB, whichmay well result in a series of large packets.

The tests were again performed on Paris and Football. Bothsequences were VBR encoded video at 30 Hz, CIF, with 2% IR MBsrandomly inserted. Recall from Section 6.1 that inserting 2% of IRMBs gave the best video quality and the lowest delay, which iswhy this percentage of IR MBs was selected for these tests. CIPwas configured and the two different GoP types (IPPP. . . andIBBP. . .) generated the video traces.

The effect of adding B-pictures is evident in Fig. 12, in whichmany more packets are dropped for Football at QP = 20. ‘Droppedpackets’ includes buffer overflow and outright channel drops.Repeated simulations of the IBBP. . . configuration of Football forQP = 20 confirmed the packet drop spike at this setting. The effectderives from the larger P-pictures that are the result of inserting B-pictures into the GoPs, when P-pictures are already large becauseof the lower QP setting. The result is a series of large P-picture-bearing packets. As previously mentioned, from Table 5, the Foot-

0

4

8

12

16

20

20 25 30 35

Dro

pped

pac

kets

%

QP

Football IPPP... Football IBBP... Paris IPPP... Paris IBBP...

Fig. 12. Dropped packets for differing GoP structure and content.

ball mean P-picture size at QP = 20 for IBBP. . . is almost twice thesize of that when an IPPP. . . GoP structure is configured and simi-larly twice the size of P-pictures for either of the Paris GoP config-urations at QP = 20 in Table 5. The IBBP. . . GoP structureparticularly impacts the more temporally complex sequence,resulting in large B-partition packets. A series of large packetsresults in more transmission errors and more packet drops if thesend buffer becomes saturated. The general reduction in packetdrop numbers in Fig. 12 for decreasing QP, results in an increasein video quality in Fig. 13. Moreover, an IPPP. . . GoP structure ispreferable except at QP = 35. However, at QP = 35 all GoP struc-tures result in a PSNR of over 25 dB.

Corrupted packet levels are generally high, Fig. 14, but inrespect to GoP structure it seems that an IPPP. . . structure maybe more favorable for temporally complex content (Football) andvice versa for less active sequences. As was previously remarked,the main consequence of higher levels of packet corruption, afterthe application of the proposed adaptive scheme, is in greaterdelay for a greater percentage of packets, Fig. 15. Apart from theanomaly at QP = 20 for Football with an IBBP. . . GoP structure, as

Fig. 13. Video quality (PSNR) for two different GoP structures and content.

0

7

14

21

28

35

20 25 30 35

Cor

rupt

ed p

acke

ts %

QP

Football IPPP...

Football IBBP...

Paris IPPP...

Paris IBBP...

Fig. 14. Percentage of corrupted packets for two different GoP structures andcontent.

0

0.02

0.04

0.06

0.08

20 25 30 35

Cor

rupt

ed p

acke

ts m

ean

dela

y (s

)

QP

Football IPPP...

Football IBBP...

Paris IPPP...

Paris IBBP...

Fig. 15. Corrupted packet mean delay for two different GoP structures and content.

0

0.015

0.03

0.045

0.06

20 25 30 35

Pack

ets M

ean

end-

to-e

nd d

elay

(s)

QP

Football IPPP...

Football IBBP...

Paris IPPP...

Paris IBBP...

Fig. 16. Mean end-to-end packet delay for differing GoP structure and content.

0

1

2

3

4

5

20 25 30 35

Dro

pped

pac

kets

%

QP

Data partition

Redundant A

Redundant A,B

Redundant A,B,C

Fig. 17. Paris sequence protection schemes’ packet drops.

L. Al-Jobouri et al. / J. Vis. Commun. Image R. 25 (2014) 1493–1506 1503

a result of many more large packets each resulting in more trans-mission delay (refer back to the discussion of Fig. 12 and Table 5),corrupted packet delay is approximately twice that of normalpacket end-to-end delay, Fig. 16, reflecting the single retransmis-sion of extra redundant that is permitted. Again for Football,QP = 20 the mean packet delay is much higher in Fig. 16 as a resultof many more large P-picture packets each with more transmissiondelay when using an IBBP. . . GoP structure. It was also confirmedthat for a moderate increase in mean packet size, making parti-tion-B completely independent of partition-C through CIP resultedin a small (a few dB) improvement in video quality, whenever theQP setting allowed sufficient packets to be delivered.

6.4. Redundant NAL units

To improve video quality it is possible to provide redundantNAL unit packets. There is a variety of ways of providing redundantpackets.

It is possible to use redundant picture slices [62] or duplicatepicture slices [63]. Redundant picture slices employ a higher QPand, hence, coarser quantization than the original slices. Thoseinterested in investigating the use of redundant slices furthershould notice that methods to refine the selection of redundantslices [64] have also been designed. However, in our scheme usingredundant picture slices will cause additional drift between theencoder and the decoder, if (say) a partition-A packet was matchedwith a partition-C packet’s data with a different QP. Therefore, wetested the inclusion of duplicate picture slices/data partitions usingthe same QP as the original slice.

Even so, there is an implementation issue when employingduplicate picture slices. Though both redundant/duplicate slices

and data-partitioned slices co-exist in the Extended profile, theyare not jointly implemented in the JM implementation of H.264/AVC [65] and, in fact, appear to not to be implemented at all inmost other software codec implementations such as QuickTime,Nero, and LEAD randomly to name a few. However, when employ-ing data-partitioning, it is also possible through repeated runs ofthe encoder to create an additional stream of all partition-A slicepackets or an additional stream consisting of partition-A and par-tition-B packets or, indeed, a complete duplicate version of the ori-ginal stream.

To test the use of duplicate packets, the same two video clips(Paris and Football) with different source coding characteristicswere employed in the tests to judge content dependency. Bothsequences were VBR encoded, with a GoP structure of IPPP. . . at30 Hz. 5% IR MB data were added to each picture, increasing thesize of partition-B packets (refer to Table 2). Again a low percent-age of IR MBs was added as a result of the gains in video qualityexperienced in Section 6.1’s tests from avoiding higher percentagesand, as before, the favorable IPPP. . . GoP structure was selected as aresult of the tests in Section 6.3.

The duplicate NAL units do not amount to a change in bitratebecause the packets are simply replicated. However, the end-to-end packet delay will obviously increase because of the interleav-ing of the duplicate slice packets. In the redundant/duplicate NALunit scheme, retransmission of extra redundant data was sched-uled for all corrupted packets, even if two packets duplicated eachother. This is because it is not possible to know in advance whetherthe extra redundant data will arrive for any one of the two packets.This provision has a significant effect in improving the video qual-ity at higher QPs. The reason is that retransmitting extra redundantdata by two alternative means increases the chance that a packetcan be reconstructed.

Fig. 17 shows the effect of the various schemes on packet dropswhen streaming Paris. ‘Data-partition’ in the Figure legend refers tosending no redundant/duplicate packets. ‘Redundant X’ refers tosending duplicate redundant packets containing data-partitionsof partition type(s) X in addition to the rateless coded data-parti-tion packets. From Fig. 17, the larger packet drop rates atQP = 20, due to the larger packets, will have a significant effecton the video quality.

Fig. 18 shows the pattern of corrupted packet losses arisingfrom simulated fast fading. There is actually an increase in the per-centage of packets corrupted if a completely redundant/duplicatestream is sent (partitions A, B, and C), though this percentage istaken from corrupted original and redundant/duplicate packets.However, the effect of the corrupted packets on video quality onlyoccurs if a packet cannot be reconstructed after application of theadaptive retransmission scheme.

0

6

12

18

24

30

20 25 30 35

Cor

rupt

ed P

acke

ts %

QP

Data partition

Redundant A

Redundant A,B

Redundant A,B,C

Fig. 18. Paris sequence protection schemes’ corrupted packets.

0

7

14

21

28

35

20 25 30 35

Mea

n PS

NR

(dB

)

QP

Data partition

Redundant A

Redundant A,B

Redundant A,B,C

Fig. 19. Paris sequence protection schemes’ video quality.

0

0.03

0.06

0.09

0.12

0.15

20 25 30 35

Cor

rupt

ed p

acke

t mea

n de

lay

(s)

QP

Data partition

Redundant A

Redundant A,B

Redundant A,B,C

Fig. 20. Paris sequence protection schemes’ mean delay for corrupted packets.

0

0.03

0.06

0.09

0.12

20 25 30 35

Pack

et m

ean

end-

to-e

nd d

elay

(s)

QP

Data partition Redundant ARedundant A,BRedundant A,B,C

Fig. 21. Paris sequence protection schemes’ mean end-to-end delay for normalpackets.

Table 6Football redundant/duplicate NAL unit results.

QP Data-partition Redundant A Redundant A, B Redundant A, B, C

Dropped packets (%)20 6.92 2.69 13.55 23.1225 4.23 2.50 1.38 4.8830 3.97 1.44 0.46 0.1235 1.66 1.44 0.38 0.00

Corrupted packets (%)20 30.76 31.37 22.72 14.5125 30.64 30.79 30.27 25.1130 27.56 30.79 27.42 26.9735 21.92 16.55 24.73 22.35

PSNR (dB)20 19.98 23.65 23.11 19.6525 20.96 26.47 27.58 33.9230 21.16 29.45 29.16 34.5035 23.27 27.65 27.59 30.73

1504 L. Al-Jobouri et al. / J. Vis. Commun. Image R. 25 (2014) 1493–1506

Examining Fig. 19 for the resulting video quality (PSNR), onesees that data partitioning with FEC protection, when used withoutredundant/duplicate packets, is insufficient to bring the videoquality to above 31 dB, that is to a ‘good’ quality. However, it isimportant to notice that sending duplicate redundant partition-Apackets alone (without redundant packets from other partitions)is also insufficient to raise the video quality to a ‘good’ rating(above 31 dB). Therefore, to raise the video quality to a ‘good’ levelrequires not only the application of the adaptive rateless channelcoding scheme but also the sending of duplicate data streams.

The impact of corrupted packets, given the inclusion of retrans-mitted additional redundant data, is again largely seen in addi-tional delay. As before, there is an approximate doubling in perpacket delay between the total end-to-end delay for corruptedpackets Fig. 20 and normal packet end-to-end delay, Fig. 21. Itmust be recalled, though, that for the redundant schemes there isup to twice the number of packets being sent. Therefore, delay isapproximately further doubled, still though with end-to-enddelays remaining in the tens of millisecond range. As previouslynoted, this type of delay range is acceptable even for interactiveapplications, but may contribute to additional delay if the WiMAXlink forms part of a longer network path.

The experimental results for Football are summarized in Table 6.Table 6 shows how packet drops and losses are reflected in videoquality. Very large numbers of packets are dropped at QP = 20because of the larger packet sizes. However, there is a thresholdeffect, as the numbers of dropped packets decline quickly withincreasing QP (as packet sizes reduce). The protection pattern forredundant/duplicate packets is accentuated compared to Paris, inthe sense that providing duplicate redundant versions of morethan just partition-A packets is now clearly seen to be preferable.Given that in Quality-of-Experience subjective tests for mobile

devices [66], news scenes rather than sport are preferred by view-ers, it may be advisable to favor content without rapid motion,especially if small footballs or similar sports’ balls need to betracked by the viewer.

6.5. Discussion

As is normal in a research environment and in papers of thisnature, we have made a set of simulations of IPTV video streamingrather than a testbed. The disadvantage of a testbed, whatever itsmerits in terms of verisimilitude is a lack of flexibility. Therefore,

L. Al-Jobouri et al. / J. Vis. Commun. Image R. 25 (2014) 1493–1506 1505

in an ideal world both types of tests should be conducted. In theabsence of a testbed then detailed simulation can be indicative ofthe streaming performance if not a conclusive demonstration. Bothtypes of test environments may not anticipate all the physical con-ditions that may arise such as distances from the base station,types of channel fading, and multi-path effects. It should benoticed though that the Gilbert–Elliott channel model used in thetests is not a physical model of channel conditions but a modelof the burst errors experienced at the receiver. In other words, itmodels the error patterns rather than the physical conditions thatgiven rise to error conditions. This distinction may explain the pop-ularity of this type of channel model for studies of video commu-nication over wireless networks.

Certain aspects of the proposed scheme such as data-partition-ing, CIP, IR MB percentage, and GoP structure can be configured atencoding time, allowing pre-storage of TV programs within serverbanks. Unfortunately, the wireless environment is notoriously vol-atile, especially if mobile stations are present. It is for this reasonthat we include the rateless channel coding element that allowsthe video server, herein collocated at the WiMAX base station, toadapt the rateless channel coding rate to adapt to channel condi-tions. Compared to (say) dropping frames the adaptive channel cod-ing solution is more fine-grained, as it occurs at the packet level.Consequently, the user experience is less likely to be disrupted.Other adaptive solutions are possible that take advantage offlexibilities within the hardware. For example, the authors havethemselves experimented with a form of WiMAX adaptive modula-tion in [67]. The disadvantage of such approaches is that they aretechnology dependent in the sense that the firmware within a basestation may need to be modified, sometimes through cross-layerintervention. A software adaptive approach such as the one usingadaptive rateless coding in this paper can more easily be deployedand modified for differing broadband wireless technologies.

7. Conclusion

Both FEC and data-partitioning of IPTV video streams are a wayof providing graceful quality degradation in a form that will workin good and difficult wireless channel conditions. This papershowed that video configuration also affect the video quality,dropped packets and delay. In that respect, it was shown that itis better to include a small percentage of IR MBs that can buildtheir effect over time than employ the cyclic IR line update scheme.Packet size, which is determined by content, video quality, and GoPstructure is an important determinant of packet drops. The use ofequal error protection is a way of taking advantage of the naturalpacket size differential, which is in inverse order of the priorityof the data partitions. Thus, smaller packet lengths already confera lower risk of channel error. However, the inverse size order ofdata partitions (larger partition-A and -B) was seen to occur whensmaller QPs were chosen. It was also confirmed that for a moderateincrease in mean packet size, making partition-B completely inde-pendent of partition-C resulted in a small but significant improve-ment in video quality.

An interesting observation is that there is a need to reducepacket size to reduce packet loss, despite the combined effect ofredundant packets and application adaptive channel coding. Thisis because during ‘bursty’ error conditions (as was simulated bythe Gilbert–Elliott channel model) it is possible that both the origi-nal packet and its redundant counterpart may be dropped or cor-rupted. However, this effect is dependent on choice of QP, as alow QP can lead to high packet drop rates with poor video quality.In general, in poor channel conditions with both slow and fast fad-ing, it is not sufficient to employ just application-layer FEC unlessstream replication also takes place.

The evaluation tests performed in this paper employed two ‘ref-erence’ video sequences from a number of such sequences thatvideo coding experts provide as a test of codec implementations.Another valid approach, particularly when developing a specificproduct is to select from video actually requested by members ofthe public. For example, video requests can be selected from a realtrace file and used to select popular videos that can subsequentlybe used in tests. In fact, this approach represents future work forthis research. Future work will also confirm the validity of thewireless channel model by streaming video over a live WiMAXlink.

Acknowledgments

Laith Al-Jobouri, who has recently completed his doctorate atthe University of Essex, UK, would like to thank his financial spon-sors at the Ministry of Science and Technology, Baghdad, Iraq.

References

[1] D. de Veeschauwer, Z. Avramova, S. Wittevrongel, H. Bruneel, Transportcapacity for a catch-up television service, in: Proc. of 7th European Conf. onInteractive Television, 2009, pp. 161–170.

[2] Y.J. Liang, J.G. Apostolopoulos, B. Girod, Analysis of packet loss for compressedvideo: effect of burst losses and correlation between error frames, IEEE Trans.Circuits Syst. Video Technol. 18 (7) (2008) 861–874.

[3] T. Wiegand, G.J. Sullivan, G. Bjontegaard, A. Luthra, Overview of the H.264/AVCvideo coding standard, IEEE Trans. Circuits Syst. Video Technol. 13 (7) (2003)560–576.

[4] T. Schierl, M.M. Hannuksela, Y.-K. Wang, S. Wenger, System layer integration ofhigh efficiency video coding, IEEE Trans. Circuits Syst. Video Technol. 22 (12)(2012) 1871–1884.

[5] I. Sodagar, The MPEG-DASH standard for multimedia streaming over theInternet, IEEE Multimedia 18 (4) (2011) 62–67.

[6] S. Pecheron, Is HEVC ready for prime time?, Comms, Eng., Design (CED) Mag.39 (9) (2013) 22–24.

[7] S. Lederer, S. Müller, C. Timmerer, Dynamic adaptive streaming over HTTPdataset, in: Proc. of ACM Utimedia Conf., 2012.

[8] J. Martin, F.Y. Fu, N. Wourms, T. Shaw, Characterizing Netflix bandwidthconsumption, in: Proc. IEEE Consumer and Commun. Conf., 2013, pp. 230–235.

[9] M. Ryu, H. Kim, U. Ramachandran, Why are state-of-the-art flash-based multi-tiered storage systems performing poorly for HTTP video streaming? in: Proc.ACM Workshop on Network and Operating Systems Support for Digital Audioand Video, 2012.

[10] J.G. Andrews, A. Ghosh, R. Muhammed, Fundamentals of WiMAX:Understanding Broadband Wireless Networking, Prentice Hall, Upper SaddleRiver, NJ, 2007.

[11] E. Dahlman, S. Parkvall, J. Sköld, P. Beming, 3G Evolution, second ed., AcademicPress, Amsterdam, 2008.

[12] O.I. Hillestad, A. Perkis, V. Genc, S. Murphy, J. Murphy, Delivery of on-demandvideo services in rural areas via IEEE 802.16 broadband wireless accessnetworks, in: Proc. of 2nd ACM Workshop on Wireless Multimedia, 2006, pp.43–52.

[13] D. Niyato, E. Hossain, Integration of WiMAX and WiFi: optimal pricing forbandwidth sharing, IEEE Commun. Mag. 45 (5) (2007) 14–146.

[14] L. Al-Jobouri, M. Fleury, M. Ghanbari, Protecting H.264/AVC data-partitionedvideo streams over broadband WiMAX, Adv. Multimedia [online journal](2012) 11 pages (article ID 129517).

[15] L. Al-Jobouri, M. Fleury, M. Ghanbari, Intra-refresh provision for data-partitioned H.264 video streaming over WiMAX, Consumer Electron. Times 2(3) (2013) 137–145.

[16] T. Stockhammer, M. Bystrom, H.264/AVC data partitioning for mobile videocommunication, in: Proc. of IEEE Int. Conf. on Image Processing, Singapore,2004, pp. 545–548.

[17] D.J.C. MacKay, Fountain codes, IEE Proc.: Commun. 152 (6) (2005) 1062–1068.[18] T. Stockhammer, W. Zia, Error-resilient coding and decoding strategies for

video communication, in: P.A. Chou, M. van der Schaar (Eds.), Multimedia in IPand Wireless Networks, Academic Press, Burlington, MA, 2007, pp. 13–58.

[19] I.A. Ali, S. Moiron, M. Fleury, M. Ghanbari, Packet prioritization for H.264/AVCvideo with cyclic intra-refresh line, J. Vis. Commun. Image Represent. 24 (4)(2013) 486–498.

[20] Y. Dhondt, S. Mys, K. Vermeirsch, R. Van de Walle, Constrained interprediction: removing dependencies between different data partitions, in:Proc. of Advanced Concepts for Intelligent Visual Systems, Delft, 2007, pp.720–731.

[21] S. Wenger, H.264/AVC over IP, IEEE Trans. Circuits Syst. Video Technol. 13 (7)(2003) 645–655.

[22] B. Waggoner, Compression for Great Video and Audio: Master Tips andCommon Sense, Focal Press, Burlington, MA, 2010.

[23] W. Simpson, Video Over IP, second ed., Focal Press, Burlington, MA, 2008.

1506 L. Al-Jobouri et al. / J. Vis. Commun. Image R. 25 (2014) 1493–1506

[24] S. Park, S-H. Jeong, Mobile IPTV: approaches, challenges, standards and QoSsupport, IEEE Internet Comput. 13 (3) (2009) 23–31.

[25] M.N. Asghar, M. Ghanbari, M. Fleury, M. Reed, Efficient selective encryptionwith H.264/SVC CABAC bin-strings, in: Proc. of IEEE Int. Conf. on ImageProcessing, 2012, pp. 2645–2648.

[26] B. Bing, 3D and HD Broadband Video Networking, Artech House, Norwood, MA,2010.

[27] M. Verhoeyen, D. de Vleeschlauwer, D. Robinson, Public IP networkinfrastructure evolutions to support emerging digital video services, BellLabs Tech. J. 14 (2) (2009) 39–55.

[28] M. Luby, T. Stockhammer, M. Watson, Application layer FEC in IPTV services,IEEE Commun. Mag. 45 (5) (2008) 95–101.

[29] J.-N. Hwang, Multimedia Networking, Cambridge University Press, Cambridge,UK, 2009.

[30] L. Al-Jobouri, M. Fleury, M. Ghanbari, Multicast and unicast video streamingwith rateless channel-coding over wireless broadband, in: Proc. IEEE Int. Conf.on Consumer Communications and Networking, 2012, pp. 737–741.

[31] I.E. Richardson, The H.264 Advanced Video Compression Standard, second ed.,Wiley & Sons, Chichester, UK, 2010.

[32] A. Shokorallahi, Raptor codes, IEEE Trans. Information Theory 52 (6) (2006)2551–2567.

[33] M. Luby, LT codes, in: Proc. of the 34rd Annual IEEE Symp. on Foundations ofComputer Science, 2002, pp. 271–280.

[34] Digital Video Broadcasting (DVB); IP Datacast over DVB-H: Content DeliveryProtocols, ETSI Technical Specification, Rev. V1.2.1, 2006.

[35] 3GPP TS 26.346, Technical Specification Group Services and System Aspects;Multimedia Broadcast/Multicast Service (MBMS); Protocols and Codecs, 3GPPTechnical Specification, Rev. V7.4.1, June 2007.

[36] M. Luby, T. Gasiba, T. Stockhammer, M. Watson, Reliable multimediadownload delivery in cellular broadcast networks, IEEE Trans. Broadcast. 53(1) (2007) 235–246.

[37] T. Gasiba, T. Stockhammer, J. Afzal, W. Xu, System design and advancedreceiver techniques for MBMS broadcast services, in: Proc. IEEE Int. Conf.Commun., 2006, pp. 5444–5450.

[38] H. Jenkac, T. Mayer, T. Stockhammer, W. Xu, Soft decoding of LT-codes forwireless broadcast, in: Proc. IST Mobile Summit, 2005.

[39] M. Peyic, H.A. Baba, I. Hamzaoglu, M. Keskinoz, Low power IEEE 802.11n LDPCdecoder hardware, Microprocess. Microsyst.: Embedd. Hardware Des. 36 (3)(2012) 159–166.

[40] T. Mladenov, S. Nooshabadi, K. Kim, Implementation and evaluation of Raptorcodes on embedded systems, IEEE Trans. Comput. 60 (12) (2011) 1678–1691.

[41] P. Ferré, A. Doufexi, J. Chung-How, A.R. Nix, D.R. Bull, Robust videotransmission over wireless LANs, IEEE Trans. Vehicular Technol. 57 (4)(2008) 2596–2602.

[42] M. Chatterjee, S. Sengupta, S. Ganauly, Feedback-based real-time streamingover WiMAX, IEEE Wireless Commun. 14 (1) (2007) 64–71.

[43] M. Chatterjee, S. Sengupta, VoIP over WiMAX, in: S. Ahson, M. Ilyas (Eds.),WiMAX, Applications, CRC Press, Boca Raton, FL, 2008. pp. 55–77.

[44] O. Hillested, A. Perkis, V. Genc, S. Murphy, J. Murphy, Adaptive H.264/MPEG-4SVC video over IEEE 802.16 broadband wireless access networks, in: Proc. ofthe International Packet Video Workshop, 2007, pp. 26–35.

[45] H.-H. Juan, H.-C. Huang, C.Y. Huang, T. Chiang, Cross-layer mobile wirelessMAC designs for the H.264/AVC scalable video coding, Wireless Netw. 16 (1)(2008) 113–123.

[46] J. Casampere, P. Sanshez, T. Villameriel, J. Del Ser, Performance evaluation ofH.264/MPEG-4 scalable video coding over IEEE 802.16e networks, in: Proc. of

IEEE Int. Symp. on Broadband Multimedia Systems and Broadcasting, 2009, pp.1–6.

[47] H. Kalva, V. Adzic, B. Furht, Comparing MPEG AVC and SVC for adaptive HTTPstreaming, in: Proc. of IEEE Int. Conf. on Consumer Electronics, 2012, pp. 160–161.

[48] T. Stockhammer, Is fine granular scalable video coding beneficial for wirelessvideo applications? in: Proc. of IEEE Int. Conf. on Multimedia and Expo, vol. 1,2003, pp. 193–196.

[49] M. Fleury, S. Moiron, M. Ghanbari, Innovations in video error resilience andconcealment, Recent Patents Signal Process. 1 (2) (2011) 1–11.

[50] V. Varsa, M. Hannuksela, Y.-K. Wang, Non-normative error concealmentalgorithms, in: Proc. 14th Meeting of ITU-T VCEG, doc.VCEG-N62, 2001.

[51] M. Fleury, R. Rouzbeh, S. Saleh, L. Al-Jobouri, M. Ghanbari, Enabling WiMAXvideo streaming, in: U.D. Dalal, Y.P. Kosta (Eds.), WiMAX, New Developments,Vukavar, Croatia, 2009, pp. 213–238, ISBN: 978-953-7619-53-4.

[52] B. Barmada, M.M. Ghandi, E.V. Jones, M. Ghanbari, Combined Turbo coding andhierarchical QAM for unequal error protection of H.264 coded video, SignalProcess.: Image Commun. 21 (5) (2006) 390–395.

[53] D.J.C. MacKay, Fountain codes, IEE Proc.: Commun. 152 (6) (2005)1062–1068.

[54] T. Issariwakul, E. Hossain, An Introduction to the ns-2 Simulator, second ed.,Springer Verlag, Berlin, Germany, 2012.

[55] J. Ebert, A. Willing, A Gilbert–Elliott Bit Error Model and its Efficient Use inPacket Level Simulation, Technical University Berlin, TelecommunicationNetworks Group, TKN Technical Report Series, TKN-99-002, 1999.

[56] G. Haßlinger, O. Hohlfeld, The Gilbert–Elliott model for packet loss in real timeservices on the Internet, in: Proc. of 14th GI/ITG Conf. on Measurement,Modelling, and Evaluation of Computer and Commun. Systs., 2008, pp. 269–283.

[57] K. Srinivasan, M.A. Kazandjieva, S. Agarwal, P. Levis, The b-factor: measuringwireless link burstiness, in: Proc. 6th ACM Conf. on Embedded Network SensorSystems, 2008, pp. 29–42.

[58] F.C.D. Tsai et al., The design and implementation of WiMAX module for ns-2simulator, in: Proc. Workshop on ns2, article no. 5, 2006.

[59] C. So-In, R. Jain, A.-K. Tamini, Capacity evaluation of IEEE 802.16e WiMAX, J.Comput. Syst., Networks, Commun. (2010) 12 pages [online].

[60] S. Mys, P. Lambert, W. De Neve, SNR scalability in H.264/AVC using datapartitioning, in: Proc. Pacific Rim Conf. in Multimedia, 2006, pp. 329–338.

[61] G. Muntean, P. Perry, L. Murphy, A new adaptive multimedia streaming systemfor all-IP multi-service networks, IEEE Trans. Broadcast. 50 (1) (2004) 1–10.

[62] I. Radulovic, P. Frossard, Y.K. Wang, M.H. Hannuksela, A. Hallapuro, Multipledescription H.264 video coding with redundant pictures, IEEE Trans. CircuitsSyst. Video Technol. 20 (1) (2010) 144–148.

[63] S.H.G. Chan, X. Zheng, Q. Zhang, W.-W. Zhu, Y.-Q. Zhang, Video loss recoverywith FEC and stream replication, IEEE Trans. Multimedia 8 (2) (2006)370–381.

[64] P. Ferré, D. Agrafiotis, D. Bull, A video error resilience redundant slicesalgorithm and its performance relative to other fixed redundancy schemes,Signal Process.: Image Commun. 25 (3) (2010) 163–178.

[65] A.M. Tourapis, K. Sühring, G. Sullivan, H.264/14496-10 AVC reference softwaremanual, in: Proc. 31st Meeting of the Joint Video Team, London, 2009.

[66] F. Agboma, A. Liotta, Addressing user expectations in mobile content delivery,Mobile Information Syst. 3 (3–4) (2007) 153–164.

[67] R. Razavi, B. Tanoh, M. Fleury, M. Ghanbari, Unequal protection using adaptiveburst profile selection for WiMAX video streaming, Electron. Lett. 44 (4) (2008)867–868.