IEEE TRANSACTIONS ON MULTIMEDIA, VOL. 9, NO. 2, FEBRUARY 2007 377

Header Detection to Improve MultimediaQuality Over Wireless Networks

Syed Ali Khayam, Shirish S. Karande, Muhammad Usman Ilyas, and Hayder Radha

Abstract—Wireless multimedia studies have revealed that for-ward error correction (FEC) on corrupted packets yields betterbandwidth utilization and lower delay than retransmissions. Tofacilitate FEC-based recovery, corrupted packets should not bedropped so that maximum number of packets is relayed to a wire-less receiver’s FEC decoder. Previous studies proposed to mitigatewireless packet drops by a partial checksum that ignored payloaderrors. Such schemes require modifications to both transmittersand receivers, and incur packet-losses due to header errors. Inthis paper, we introduce a receiver-based scheme which uses thehistory of active multimedia sessions to detect transmitted valuesof corrupted packet headers, thereby improving wireless multi-media throughput. Header detection is posed as the decision-the-oretic problem of multihypothesis detection of known parametersin noise. Performance of the proposed scheme is evaluated usingtrace-driven video simulations on an 802.11b local area network.We show that header detection with application layer FEC pro-vides significant throughput and video quality improvements overthe conventional UDP/IP/802.11 protocol stack.

Index Terms—Communication systems, multimedia communi-cation, video signal processing, wireless LAN.

I. INTRODUCTION

WIRELESS communication channels incur unpredictableand time-varying packet-losses due to fading, interfer-

ence, and mobility. This data loss is particularly detrimentalfor real-time communications whose delay constraints generallydo not allow retransmission-based recovery of lost packets. Tocombat wireless errors and losses, emerging multimedia stan-dards have introduced enhanced error-resilience and conceal-ment features, e.g., slices in JVT/H.264 [1] and reversible VLCin MPEG-4 [2]. For an error-resilient application, distortion inmultimedia quality can be decreased by reducing the amountof data loss at a wireless receiver, i.e., by relaying maximumnumber of error-free and corrupted packets to the multimediaapplication. It is then up to the application to drop or retain thecorrupted packets.

Due to the high error-rates of wireless media, many errorsare not corrected by the physical layer. These errors causechecksum failures at higher layers, consequently leading to

Manuscript received February 18, 2005; revised May 5, 2006. The associateeditor coordinating the review of this manuscript and approving it for publica-tion was Dr. Chang Wen Chen.

The authors are with the Department of Electrical and Computer Engi-neering, Michigan State University, East Lansing, MI, 48824 USA (e-mail:khayamsy@egr.msu.edu; karandes@egr.msu.edu; ilyasmuh@egr.msu.edu;radha@egr.msu.edu).

Color versions of one or more of the figures in this paper are available onlineat http://ieeexplore.ieee.org.

Digital Object Identifier 10.1109/TMM.2006.889051

a significant number of packet drops on a link employinga conventional (e.g., TCP-UDP/IP) protocol stack. Previouswireless multimedia studies proposed to reduce packet dropsby employing a partial checksum which only covers packets’headers while payload errors are ignored [3]–[9]. Payload er-rors are subsequently corrected using forward error correction(FEC) at the application layer. It has been shown that partialprotection with FEC requires much lesser FEC redundancythan a conventional protocol stack that drops corrupted packets.However, schemes using partial checksum incur packet dropsdue to header errors, especially at high data rates. Also, supportof partial checksum requires changes to the standard protocolsat the multimedia transmitter and/or intermediate networknodes. In many realistic scenarios, modifications to multimediaservers and intermediate nodes cannot be dictated by theend-receivers.1

In this paper, we propose a receiver-based decision-theoreticapproach to relay packets with corrupted headers to wirelessmultimedia receivers. The proposed technique requires no mod-ifications to wireless transmitters and intermediate nodes, andonly minor modifications are required at the receiver. We iden-tify critical header fields (CHF) that can uniquely differentiatemultimedia sessions at a wireless receiver. Under the proposedtechnique, wireless multimedia receivers maintain an a prioriCHF histogram from previously received error-free packets. Wealso propose and employ a likelihood function based on thesample space of the a priori distribution. When a corruptedpacket is received, the a priori distribution and the likelihoodfunction are used to compute the a posteriori distribution. Cor-recting the CHF then reduces to the decision-theoretic problemof multihypothesis detection of known parameters in noise [10],in short referred to as header detection in this work.

We demonstrate the efficacy of the proposed scheme by trace-driven video simulations at different data rates and for varyingnumber of video sessions over an 802.11b LAN. We show thatheader detection has high accuracy and negligible false posi-tives. After header detection, we use FEC [11] at the applicationlayer to correct the corrupted data. We show that header detec-tion requires much lesser FEC redundancy than the 802.11b pro-tocol stack. We also show that for fixed FEC parameters, headerdetection provides significantly better multimedia quality thanthe 802.11b protocol stack.

The rest of this paper is organized as follows. Section II de-tails our proposed approach and its variants. Section III de-scribes the experimental setup. Sections IV–VI, respectively,

1For instance, public domain video web-casting is oblivious of the end-linkcommunication media and often identical multicast sessions serve heteroge-neous users with wired as well as wireless end-links.

1520-9210/$25.00 © 2007 IEEE

378 IEEE TRANSACTIONS ON MULTIMEDIA, VOL. 9, NO. 2, FEBRUARY 2007

evaluate the throughput, FEC, and video performances of theproposed scheme.

II. THE HEADER DETECTION METHODOLOGY

The main objective of header detection is to relay maximumnumber of (error-free and corrupted) packets to a wireless re-ceiver’s application layer. In this paper, we use 802.11b wirelessLANs as our performance evaluation test-bed. We start by iden-tifying critical header fields (CHF) that can uniquely classify amultimedia session at the receiver and are not liable to changeduring a multimedia transmission. These fields are: destinationMAC address, source and destination IP addresses, and sourceand destination ports. More fields can be concatenated to theseCHF without any modifications to the proposed methodology.

Let denote the CHF of a received packet and letbe the set of distinct correctly received CHF

of the last multimedia packets. That is, each , for, represents the CHF of session . The set ,

the set of all bins of the histogram, constitutes the sample spaceof the a priori CHF distribution. The a posteriori distribution ofthe received CHF can then be expressed as

(1)

where represents the a priori probability of CHFand the conditional probability terms, ’s, representthe likelihood function [10]. The a posteriori distribution ren-ders the probability that while the received CHF is , the actualtransmitted CHF was . Once the a posteriori distribution isascertained, the present header detection problem correspondsto the problem of multihypothesis detection of known param-eters in noise,1 [10] with CHF of the multimedia sessions rep-resenting the hypotheses under consideration. We refer to thepresent problem as header detection. The following sections ex-plain the computation of the a priori distribution and the likeli-hood function.

A. Generation of the a priori Distribution

Under the proposed approach, a multimedia receiver main-tains a CHF histogram based on previously-received error-freepackets. Normalizing this histogram renders a robust a prioridistribution of active wireless sessions. The CHF histogram canbe generated either using the error-free packets that were des-tined for the local receiver or, due to the broadcast nature ofwireless networks, using the error-free packets destined for anynode on the wireless network. This notion will be used later todefine two variants of the proposed technique.

Fig. 1 shows a generalized flowchart outlining both variantsof the proposed technique. If a received packet passes the MAClayer checksum and the packet is destined for the local receiver,then the packet is passed to the application. The receiver thendecides whether or not the a priori histogram should be updated

2The classical detection problem concludes with each of the hypotheses beingscored as detected or undetected. We, on the other hand, use the detected CHF asan estimate of the transmitted fields to perform further packet processing. Thus,the present problem can also be viewed as an estimation-theoretic problem.

Fig. 1. Flowchart of the header detection methodology.

with the CHF of the received packet. This decision is dependenton the variant under consideration. If the decision is affirmative,the histogram is updated and the receiver returns to its initialstate; otherwise the receiver returns directly to the initial statewithout updating the histogram. If the received packet fails thechecksum then the receiver has to decide whether it will attemptpacket recovery or not. This decision varies with respect to thevariant under consideration.

It should be noted that the present strategy does notcharacterize the complete prior information. Let andrepresent two events corresponding to the reception of anerror-free and a corrupted packet, respectively. We onlycater for event , but the complete is

. Probabilities involving eventcan be computed only if the transmitter or an intermediate

node at the edge of the wired channel protects the CHF usingFEC. Such a strategy, however, requires modifications to thetransmitter or intermediate nodes, thereby defying the mainpremise of this work. Our simulation results show that theheader detectors employing the partial priors (based only onevent ) incur negligible errors, implying that the partial priorsprovide a good estimate of the complete priors.

B. Computation of the Likelihood Function

Since we know the possible values of the active CHF from thesample space of the a priori distribution, we propose a likeli-hood function to exploit this knowledge. We assume a memory-less channel with a fixed probability of bit-error, . While theassumption of a memory-less channel is somewhat unrealistic,a model with memory can be employed if some side-informa-tion renders real-time channel prediction and characterization.We are unaware of any current scheme capable of providingsuch side-information. In the absence of good channel charac-terization, the memory-less channel assumption is a pragmaticalternative. It should be noted though that, while our proposedmethodology relies on the memory-less premise, the error traces

KHAYAM et al.: HEADER DETECTION TO IMPROVE MULTIMEDIA QUALITY OVER WIRELESS NETWORKS 379

used for simulations later in this paper were collected over an ac-tual 802.11b network under realistic settings. The memory-lessassumption, hence, provides a lower bound on the achievableperformance. We use a blind and fixed estimate of for the ex-periments.

To calculate the conditional likelihood probability,, for a given , we compute the Hamming dis-

tance between and . The Hamming distance will indicatethe total number of bits that are different between and .Now if we assume that the different bits are in fact the bit-errorsintroduced by the memory-less channel, then can bewritten as

(2)

where is the length of the CHF in bits, and the functionis the Hamming distance between bit sequences

and . The expression given in (2) renders the probability thatwas the transmitted CHF which, due to channel bit-errors,

was received as . By plugging (2) in (1), we obtain the aposteriori density on which different detectors can be applied.

C. AxMAP and RMeAP Detectors

We apply two detectors on the a posteriori distribution:Approximate Maximum a posteriori (AxMAP) Detector:

This detector selects the mode (i.e., the most likely ) of thea posteriori density as the detected CHF. We refer tothis detector as an approximation of the well-known MAP de-tector [10] because i) AxMAP operates in the discrete domain,and hence does not retain the optimality properties of the MAPdetector and ii) in the present setup, the a priori distribution isbased on partial priors.

Rounded Mean a posteriori (RMeAP) Detector: This de-tector selects the mean of the a posteriori density, ,rounds it to the nearest CHF value, and uses that value as thedetected CHF. We map the CHF of each multimedia sessionto a distinct integer value. Thus, the mapping yields

. The mean is computed using themapped values, , as the outcomes of the randomvariable . Once the mean is computed and rounded, we againmap the resultant integer value, say , back to its correspondingCHF, .

D. Variants of the Proposed Methodology

We use two variants of the proposed scheme.Global Statistics Variant (GSV): This variant exploits the

broadcast nature of wireless networks. The a priori CHF distri-bution is generated using all the error-free packets on the wire-less network, and not necessarily the packets that were destinedfor a particular receiver. Thus, each promiscuous node in thewireless network maintains a CHF histogram from packets thatit correctly receives.

Local Statistics Variant (LSV): Under this variant, the CHFhistogram is computed using packets that are destined for thelocal wireless node. In the LSV, a receiver first reads the desti-nation MAC and destination IP addresses of a received packet.Header detection is initiated if at least one of these addressesis the local node’s address. The LSV will be effective in securewireless networks where packet header information is encrypted

Fig. 2. Simulation setup for evaluation of GSV and LSV.

to curb network sniffing. Furthermore, LSV requires lesser pro-cessing and data maintenance overhead than GSV, which is avery important consideration for complexity- and power-con-strained wireless receivers.

It is noteworthy that, in addition to wrong detections, LSV ex-periences missed packets where both the destination MAC andIP addresses are corrupted, and thus detection is not attempted.Also, the prior information for LSV is not as robust as GSV.It would, therefore, be expected that LSV performs worse thanGSV. On the contrary, we show that at most data rates LSV’sperformance is comparable to GSV. Hence, LSV provides anaccurate, effective and low-complexity alternative to GSV.

III. EXPERIMENTAL SETUP

We use the 802.11b MAC layer bit-error traces collected in[8] to simulate the wireless channel. The bit-error traces werecollected at 2, 5.5, and 11 Mbps data rates of an operational802.11b wireless LAN under realistic settings. In [8], it wasshown that the bit-error rate at the 802.11 MAC layer is directlyproportional to the data rate at which the wireless network is op-erating. This observation will be emphasized in the subsequentperformance evaluation sections of this paper. Readers are re-ferred to [8] for details of data collection.

For our simulation setup, we assumed the 802.11b networkoutlined in Fig. 2. Our setup consisted of up to twenty mul-timedia servers with IP addresses of well-known multimediaweb servers. In order to quantify the worst-case performance,we used multiple source IP addresses from the same streamingsubnet so that only a few bit-errors can map one source IP ad-dress into another. Furthermore, we assumed that all the streamswere being received at only three wireless stations in the wire-less LAN. The wireless receivers had IP addresses which werevery close to each other so that a few bit-errors can change onedestination IP address to another. Distinct source and destina-tion ports were used for each multimedia stream.

The video streams used for the simulations were compressedusing the H.264/JVT video encoder [1]. Unless otherwisestated, all video streams have the same encoding bitrate. These

380 IEEE TRANSACTIONS ON MULTIMEDIA, VOL. 9, NO. 2, FEBRUARY 2007

Fig. 3. GSV and LSV throughputs of AxMAP and RMeAP detectors at 2 Mbps with uniform source distributions. (a) GSV and (b) LSV.

equal bitrate multimedia streams render the least informativea priori distribution. Since all the streams have the sameencoding bitrate, the amount of data transmitted over thenetwork increases linearly with an increase in the numberof multimedia streams. For each packet transmission i) 512bytes were taken from a video stream as the packet payload,ii) UDP, IP and 802.11-MAC headers were appended to thepayload, and iii) the resultant (header and payload) packetwas corrupted using the bit-error traces. The video streamswere assigned to the three wireless receivers in a round-robinmanner, , ,

, , up to stream19.Transmission of packets from each stream was also simulatedin a round robin fashion.

IV. THROUGHPUTS OF GSV AND LSV

In this section, we evaluate the throughputs provided by GSVand LSV at 2, 5.5, and 11 Mbps data rates of an 802.11b wire-less LAN. Here, the term “throughput” refers to the ratio ofthe total number of packets correctly relayed to the wireless re-ceiver’s application and the total number of packets sent by thesender’s application layer over a fixed period of time. Some ofthese packets will be corrupted, and therefore this throughputconsists of both error-free and corrupted packets. In subsequentsections, we demonstrate that errors in corrupted packets canbe corrected using much less FEC redundancy than the conven-tional UDP/IP/802.11 protocol stack. Unless otherwise stated,we use a bit-error probability of and results are reportedfor receiver-0.

A. Throughput at 2 Mbps

Fig. 3(a) outlines the performance of AxMAP and RMeAPdetectors in conjunction with GSV at 2 Mbps. The dotted linein the figure provides the total number of corrupted packets; thatis, the maximum number of corrupted packets that can be cor-rectly detected and relayed to the receiver’s application layer. Itis clear from Fig. 3(a) that both Ax-MAP and RMeAP detectorshave very high detection accuracies; for both detectors, morethan 99% of received packets are correctly detected.

Fig. 3(b) shows the performance of the detectors in an LSVscenario. The LSV only focuses on packets destined for re-ceiver-0, and therefore comparison with Fig. 3(a) shows that thetotal number of corrupted packets is always lesser than the GSVcase. Also note that the performances shown in Fig. 3(b) arethe sum of incorrect decisions (i.e., packets whose destinationwas not accurately ascertained) and missed packets (i.e., packetswhose recovery was not attempted). It is easily observed fromFig. 3(b) that even in the LSV case both AxMAP and RMeAPdetectors provide very accurate decisions; in all cases, more than99% decisions are correct. Thus, although based completely onlocal statistics, the LSV renders performance comparable to thatof GSV at 2 Mbps.

B. Throughput at 5.5 Mbps

The AxMAP and RMeAP detectors’ throughputs are shownin Fig. 4. For GSV, the results are similar to the results at 2 Mbpssince both AxMAP and RMeAP detectors provide extremely ac-curate detection of corrupted packet headers; more than 99% ofthe packets are correctly detected. It was shown in [8] that theerror-free packet throughput at 5.5 Mbps is much lower than 2Mbps. The results of Fig. 4(a) show that even for this somewhathigh error-rate 5.5 Mbps channel, the AxMAP and RMeAP de-tectors render excellent performances in a GSV scenario.

The 5.5 Mbps LSV results given in Fig. 4(b) are also consis-tent with the 2 Mbps case as both AxMAP and RMeAP havevery high detection accuracies. The overall LSV performanceis slightly inferior to the 2 Mbps case, but more than 98% ofthe corrupted packets are detected correctly. Thus, we concludethat the increased error-rate at 5.5 Mbps does not deteriorate theperformance of the detectors under consideration.

C. Throughput at 11 Mbps

Fig. 5(a) demonstrates that in a GSV scenario, both AxMAPand RMeAP detectors are very accurate for the high error-rate11 Mbps channel. The GSV performances of AxMAP andRMeAP detectors are slightly worse than at 5.5 and 2 Mbps.This performance degradation is due to the increased error-rate.

KHAYAM et al.: HEADER DETECTION TO IMPROVE MULTIMEDIA QUALITY OVER WIRELESS NETWORKS 381

Fig. 4. GSV and LSV throughputs of AxMAP and RMeAP detectors at 5.5 Mbps with uniform source distributions. (a) GSV and (b) LSV.

Fig. 5. GSV and LSV throughputs of AxMAP and RMeAP detectors at 11 Mbps with uniform source distributions. (a) GSV and (b) LSV.

Nevertheless, 98% of the corrupted packets are detected cor-rectly.

The LSV detection performance at 11 Mbps (Fig. 5(b) ))shows a clear drop in the accuracy of both detectors; approx-imately 85% of the packets are correctly detected as opposed tomore than 99% and 98% at 2 and 5.5 Mbps, respectively. Thisresult is not very surprising since we know from [8] that the per-centage of error-free packets is approximately 15% at 11 Mbpsas opposed to more than 64% and 99% at 5.5 and 2 Mbps. Dueto the high error-rate at 11 Mbps, and due in part to the localnature of the variant under consideration, there are not enougherror-free packets to build a robust a priori distribution. Never-theless, keeping in view the very high error-rate at 11 Mbps, theLSV performance is still quite good since more than 85% of thepackets are correctly detected for any number of streams.

D. Detection for Streams With Different Source Bitrates

At this point, we have established the efficacy of the proposedschemes for the case where the source bitrates of all receivedstreams are equal. In Fig. 6, we evaluate the scenario where thesource bitrates are skewed, with encoding bitrates varying from

50 Kbps to 1 Mbps. Accuracies of the AxMAP and RMeAP de-tectors were almost identical, therefore Fig. 6 only shows theAxMAP detector. Also, performance at 2 Mbps was similarand is skipped for brevity. It can be clearly observed that thethroughput of the header detection scheme is very good at 11and 5.5 Mbps. While LSV’s performance is slightly inferior toGSV, both variants relay most of the corrupted packets to the ap-plication layer. Henceforth, all results employ the (worst-case)equal bitrate streams.

E. Coping With False Positives

A packet that is not intended for a multimedia session, butgets relayed to that session represents a false positive. False pos-itive rates of the AxMAP detector for the twenty streams caseare given in Table I. It can be observed that the false positiverates at all data rates are extremely low. While these false pos-itives are almost negligible, they can desynchronize video andFEC decoders. Thus, it is important to detect the false positives.To this end, when using JVT/H.264 based encoding one canhave a single slice per packet, with the slice sequence numbersprotected with enough redundancy to ensure that these sequence

382 IEEE TRANSACTIONS ON MULTIMEDIA, VOL. 9, NO. 2, FEBRUARY 2007

Fig. 6. GSV and LSV throughputs of the AxMAP detector with skewed source distributions. (a) 5.5 Mbps and (b) 11 Mbps.

TABLE IFALSE POSITIVES OF THE AXMAP DETECTOR WITH A TOTAL OF 20 STREAMS

numbers can always be recovered at the receiver. A receivercan then drop all packets whose slice numbers are much largeror smaller than the next/expected slice number. For video en-coders that do not have a slice/packet sequence number, a smallincremental packet sequence number with parity bytes can beinserted into each packet by the sender’s application layer. Thissequence number based scheme can also provide erasure loca-tions (i.e., dropped packets) to the FEC decoder.

F. Discussion

The high throughputs rendered by LSV and GSV contain cor-rupted data and it is important to ascertain how much redun-dancy will be required to correct the erroneous bits within thecorrupted packets. Similarly, in order to draw a fair comparisonit is important to compare the FEC redundancy of the proposedschemes with a similar FEC scheme deployed on an applica-tion running on the conventional UDP/IP/802.11-MAC protocolstack. The next section provides this comparison.

V. FEC PERFORMANCE OF GSV AND LSV

The last section focused on techniques to relay maximumamount of error-free and corrupted multimedia data to a wire-less application. In this section, we evaluate the amount of FECredundancy required by the multimedia application to correct er-rors in the corrupted multimedia content. We compare the FECredundancies of GSV and LSV with the redundancy requiredto recover from losses if a conventional UDP/IP/802.11-MACprotocol stack is employed at the receiver. Since both AxMAPand RMeAP detectors relay approximately the same number ofcorrupted packets to the application, henceforth we only reportresults for the AxMAP detector.

Due to header detection, some of the packets reaching the ap-plication layer have errors, while some other packets are lost dueto missed or inaccurate detections. We consider missing bytesdue to lost packets as erasures in an FEC codeword. Thus, anFEC scheme operating over LSV and GSV should be able todecode errors and erasures simultaneously.

A. FEC Construction

We employ block-based systematic Reed–Solomon (RS)codes for FEC-based recovery. For error correction, if a code-word has number of redundant symbols then a maximum of

transmission errors in that block can be corrected. Noerror can be corrected if the number of transmission errors in ablock is greater than . Many contemporary FEC schemeshave been designed for and applied to erasure recovery, whereknowledge of error locations facilitates FEC decoding. If acodeword has a redundancy of then a maximum of erasurescan be recovered. Thus, in a present scenario, where both errorsand erasures are observed, the data recovery and correction isbounded by , where is the number of errors and

is the number of erasures. For the UDP/IP/802.11 protocolstack, we employ RS erasure decoding (Berlekamp algorithm[11]) to recover dropped packets. For LSV and GSV, we usea variant of the RS coding that is capable of joint error anderasure recovery [11].

In this section, we show FEC results for a packet block lengthof . The FEC scheme is systematic and hence a packetblock can be segregated into message packets and redundantpackets. Each FEC codeword is composed of four bytes from adifferent packet, where each packet consists of 512 bytes. Thus,each packet contributes to 128 separate FEC codewords, andeach codeword spans over 30 packets. Thus, while the packetblock-length is 30 packets, the code-length is 120 bytes (or120 symbols where each symbol is drawn from a Galois field

.) The results we present for the above FEC parametersare representative of performance trends that were observed fora wider range of packet block lengths, code lengths and packetsizes.

KHAYAM et al.: HEADER DETECTION TO IMPROVE MULTIMEDIA QUALITY OVER WIRELESS NETWORKS 383

Fig. 7. FEC performances of GSV and LSV with respect to UDP/IP/802.11-MAC from the minimum up to 100% recovery. (a) GSV (b) LSV.

Fig. 8. LSV video evaluation for fixed FEC parameters at 5.5 Mbps: (a) LSV video frame-46, (b) UDP video frame-46, and (c) PSNR of a GOP with LSV andUDP.

B. Comparison of FEC Redundancies

Fig. 7 compares the FEC redundancies of GSV and LSV tothe conventional UDP/IP/802.11-MAC protocol stack. It can beclearly seen that for all data rates the conventional schemes re-quire a higher amount of redundancy to provide 100% reliabilitywhen compared with the proposed header detection variants. Itcan be observed from Fig. 7 that the FEC performances of GSVand LSV improve with respect to an increase in the 802.11b datarate. This is congruent with our preceding discussions where weobserved that the error-rate at 2 Mbps is quite low and thereforeerror recovery does not require much redundancy. The FEC re-dundancy subsequently increases at 5.5 and 11 Mbps, and theperformance gap between header detection and the conventionalprotocol stack widens.

Here, it is noteworthy that due to their enhanced error re-silience features, emerging real-time applications can toleratea certain level of losses in the multimedia content [1], [2]. Thus,often in video streaming if the underlying source coding is error-resilient then the aim of the FEC scheme is to keep the numberof losses under a tolerable threshold instead of attempting 100%data recovery. Consequently, it is important to compare the rel-ative performance of the conventional and proposed schemeseven over the region where neither conventional UDP/IP/802.11

nor GSV/LSV provides 100% reliability. In this context, theutility of the proposed schemes can be further appreciated bynoting that the throughputs of the proposed schemes are equalto or better than the conventional protocol stack for all redun-dancies.

VI. VIDEO PERFORMANCE OF GSV AND LSV

The results presented so far show throughput improvementsrendered by GSV and LSV. For fixed FEC parameters (i.e.,possibly less than 100% recovery), the video received by theapplication after header detection and FEC decoding containserror-free, partially corrupted, and completely dropped packets.It is then up to the multimedia application to recover from thesechannel impairments. We decompress the video after FECdecoding to substantiate the advantages of using the proposedGSV and LSV schemes.

We evaluated video for varying values of quantization param-eters, frame frequencies, frame sizes, types of video sequences,FEC parameters and data rates. In all the video performanceevaluations, the video quality rendered by GSV and LSV wasconsistently better than the conventional UDP/IP/802.11 pro-tocol stack. It was observed that at 2 Mbps the total packetdrops, even for a conventional protocol stack, were very small

384 IEEE TRANSACTIONS ON MULTIMEDIA, VOL. 9, NO. 2, FEBRUARY 2007

Fig. 9. GSV video evaluation for fixed FEC parameters at 5.5 Mbps: (a) GSV video frame-70, (b) UDP video frame-70, and (c) PSNR of a GOP with GSV andUDP.

Fig. 10. Identical GSV and LSV video performance for fixed FEC parameters is shown in (a), (b), (d), (e), and (g). (a) Frame-15 with GSV, (b) frame-15 withLSV, (c) frame-15 with UDP, (d) frame-29 with GSV, (e) frame-29 with LSV, (f) frame-29 with UDP, and (g) PSNR of two GOPs with GSV/LSV and UDP.

and the margin for improvement was small. On the other handat 11 Mbps, while the proposed scheme easily outperformed theconventional stack, the channel errors (even for GSV and LSV)were quite profound. Therefore, in this section we focus on the5.5 Mbps data rate.

The JVT/H.264 video encoder/decoder was used to com-press/decompress the video streams. We used a group ofpictures (GOP) size of 15 frames. The video frame sequencewas , where is an intra-coded frame and is aunidirectional predictive frame. The video sequences had a CIFframe size and were encoded at a frequency of 30 frames/s. Thesource encoding parameters were adjusted to provide a sourcebitrate of 700 kbps. A JVT video packet size of 512 bytes wasused. For FEC recovery, we use the RS-based scheme describedin the previous section with packet block length of .Each block consists of 21 message packets.

Figs. 8 and 9 provide subjective and temporal comparisons onthe basis of video frames and peak signal-to-noise ratio (PSNR).

In both figures there were many video frames that were com-pletely lost due to the UDP/IP/802.11 protocol stack. Theseframe losses are represented by a PSNR of zero dB. It can beseen that the video quality provided by the header detectionvariants is significantly better than the conventional protocolstack. It should be highlighted that for the UDP/IP/802.11-MACprotocol stack, a number of picture frames are completely lost,leading to considerable block distortion and motion jerkiness;these artifacts are not evident in the provided results. Note alsothat while the video sequence in Fig. 9 had a high amount oftemporal redundancy, the losses for the conventional stack wereso high and prolonged that both block distortions and motionjerkiness artifacts were observed in the multimedia content.

In general, our video simulation results indicated a slightlyinferior yet comparable LSV performance as opposed to GSV.There were scenarios in which GSV and LSV provided iden-tical performances. One such scenario is shown in Fig. 10 forthe Stefan video sequence, where both LSV and GSV provide

KHAYAM et al.: HEADER DETECTION TO IMPROVE MULTIMEDIA QUALITY OVER WIRELESS NETWORKS 385

identical performance improvement over the conventional pro-tocol stack. Thus, LSV provides an effective low-complexity al-ternative to GSV.

VII. CONCLUSION

We proposed a receiver-based approach to improve wirelessbandwidth utilization for multimedia applications. Two variantsof the proposed approach were investigated. Comparison witha conventional UDP/IP/802.11 protocol stack showed thatboth variants of the proposed scheme significantly improvethroughput and multimedia quality.

REFERENCES

[1] Draft ITU-T Recommendation and Final Draft International Standardof Joint Video Specification (ITU-T Rec. H.264/ISO/IEC 14496-10AVC), , Mar. 2003, ISO/IEC JTC 1/SC29/WG11 and ITU-T SG16Q.6, Doc. JVT-G050.

[2] Text of ISO/IEC 14496-2:2001 (Unifying N2502, N3301, N3056, andN3664, , Jul. 2001, ISO/IEC JTC 1/SC29/WG11, Doc. N4350.

[3] L. Larzon, M. Degermark, and S. Pink, “UDP lite for real time mul-timedia applications,” in Proc. IEEE Int. Conf. Communications, Jun.1999.

[4] ——, “Efficient use of wireless bandwidth for multimedia applica-tions,” in Proc. IEEE Int. Workshop on Mobile Multimedia Commu-nications, Nov. 1999.

[5] A. Singh, A. Konrad, and A. D. Joseph, “Performance evaluation ofUDP lite for cellular video,” in Proc. ACM Int. Workshop on Networkand Operating Systems Support for Digital Audio and Video, Jun. 2001.

[6] H. Zheng and J. Boyce, “An improved UDP protocol for video trans-mission over internet-to-wireless networks,” IEEE Trans. Multimedia,vol. 3, pp. 356–365, Sep. 2001.

[7] H. Zheng, “Optimizing wireless multimedia transmissions throughcross layer design,” in Proc. IEEE Int. Conf. Multimedia and Expo.,Jul. 2003.

[8] S. A. Khayam, S. Karande, H. Radha, and D. Loguinov, “Performanceanalysis and modeling of errors and losses over 802.11b LANs for high-bitrate real-time multimedia,” Signal Process. Image Commun., vol. 18,no. 7, pp. 575–595, Aug. 2003.

[9] S. A. Khayam, S. Karande, M. Krappel, and H. Radha, “Cross-layerprotocol design for real-time multimedia applications over 802.11b net-works,” in Proc. IEEE Int. Conf. Multimedia Expo., Jul. 2003.

[10] H. L. Van Trees, Detection, Estimation, and Modulation Theory. NewYork: Wiley, 2001, pt. I.

[11] R. E. Blahut, Theory and Practice of Error Control Codes. Reading,MA: Addison-Wesley, 1984.

Syed Ali Khayam received the B.E. degree in com-puter systems engineering from National Universityof Sciences and Technology (NUST), Pakistan, in1999, and the M.S. and Ph.D. degrees in electricalengineering from Michigan State University, EastLansing, in 2003 and 2006, respectively.

He has accepted an Assistant Professor position atNUST, Pakistan, starting February/March 2007. Heworked at Communications Enabling Technologiesfrom October 2000 to August 2001. His research in-terests include analysis and modeling of statistical

phenomena in computer networks, network security, cross-layer design for wire-less networks, and real-time multimedia communications.

Shirish S. Karande received the B.E. degree inelectronics and telecommunications from Universityof Pune in 2000, the M.S. degree in electricalengineering from Michigan State University (MSU),East Lansing, in 2003 and he is currently pursuingthe Ph.D. degree at MSU.

His research interests include multimedia com-munication, wireless networking, and source andchannel coding.

Muhammad Usman Ilyas received the B.E. degreein electrical engineering from the National Univer-sity of Sciences and Technology (NUST), Pakistan,in 1999, the M.S. degree in computer engineeringfrom Lahore University of Management Sciences in2005, and is currently pursuing the Ph.D. degree inelectrical engineering at Michigan State University,East Lansing.

He was with East-West Infiniti as an ElectronicsEngineer, at Engro Checmical as a Systems Engineer,and at Communications Enabling Technologies as a

Design Engineer. He is currently working with Quartics, LLC. His research in-terests include cross-layer solutions in IEEE 802.11 and 802.15.4 wireless net-works, wireless sensor networks, and computer architecture.

Hayder Radha received the B.S. degree (withhonors) from Michigan State University (MSU),East Lansing, in 1984, the M.S. degree from PurdueUniversity, West Lafayette, IN, in 1986, and thePh.M. and Ph.D. degrees from Columbia University,New York, in 1991 and 1993, respectively (all inelectrical engineering).

He joined MSU in 2000 as Associate Professorin the Department of Electrical and ComputerEngineering. From 1986 to 1996, he was with BellLaboratories. From 1996 to 2000, he worked at

Philips Research USA and became a Philips Research Fellow in 2000. Hisresearch interests include wireless and multimedia communications and net-working, stochastic modeling, and image and video coding and compression.He has more than 25 patents in these areas. He served as cochair and editor ofthe ATM and LAN Video Coding Experts Group of the ITUT in 1994–1996.

Dr. Radha is a member of the IEEE Signal Processing Multimedia TechnicalCommittee. He is a recipient of the Bell Labs Distinguished Member of Tech-nical Staff Award (1993), the Withrow Distinguished Scholar Award (2003), andthe Microsoft Research Content and Curriculum Award (2004).