H.

263-Based Wireless Video Transmission in Multicode CDMA Systems

Mario R. Hueda and Carlos A. Marqu6s
Laboratorio de Comunicaciones Digitales - Universidad Nacional de C6rdoba Casilla de Correo 755 - Correo Central - C6rdoba (5000) - Argentina Email: mhueda @com.uncor.edu maximal ratio combiner (MRC) and convolutional codes (CC) with soft-decision Viterbi decoding (SD-VD)). These important topics are addressed in this paper. We focus our study on H.263/H.263+ video transmission at low data rates (i.e., c32 kbps) in an IS-95B based system over slow fading Rayleigh channels. In order to evaluate performance of video transmission in MC-CDMA systems, a novel, simple, and accurate semi-analytic methodology is proposed. Based on the new approach, as well as on entire system simulations, we derive a variety of peak signal-to-noise ratio (PSNR) results for various H.263 video test sequences. In addition, we I. INTRODUCTION investigate the effectiveness of retransmission error correction As a result of the evolution of wireless communications and error resilience techniques to enhance the robustness of towards transmission of multimedia services, multicode direct H.263/H.263+ video transmission over IS-95B systems. sequence code division multiple access (MC-CDMA) has 11. BACKGROUND been proposed to support variable data rates [l]. In MCCDMA (such as IS-95B [2]), high speed is provided through code aggregation. Up to eight codes may be assigned in IS- A. H.263 Based Video Compression The picture resolution is often Q C F (Quarter Common 95B (one fundamental code channel (FCC) and seven supplemental code channels (SCC)), providing a maximum Intermediate Format, 176x144 pixels), which is the most bit rate of 115.2 kilobits per second (kbps) [2]. However, it common input format at low bit-rates. At QCIF resolution, has been shown that multimedia services in MC-CDMA must each picture is divided into 11x9 macroblocks (MBs), which operate at very low bit rates (c32 kbps), in order to provide an comprise 16x16 luminance samples, and two corresponding 8x8 blocks of chrominance samples. A fixed number of adequate capacity for voice users [3]. Video transmission is one of the most important successive MBs is usually grouped into a group of blocks requirements for next generation of wireless systems. H.263 (GOB) and side information that is appropriate for a larger is a compression standard developed to transmit video using number of MBs, but not for an entire frame ($rametrefers to the phone network at data rates less than 64 kbps [4]. A video), can be communicated efficiently on that level. typical H.263 video stream is composed of inter-coded B. Error Concealment and Video Coding Mode frames (I-frames) and predictive-coded frames (P-frames). H.263 provides means to insert synchronization words at An I-frame is an independently coded video frame that can be the picture and optionally at the GOB layers. We use the latter decoded by itself. A P-frame is composed of changes in the current image of the video stream relative to the last I-frame option to allow resynchronization in case of errors. We also encoded. In general, an H.263 video stream is usually an I- assume that GOB sync words are inserted at the beginning of frame followed by many P-frames, with an I-frame re- each macroblock row. This is exploited by the error introduced to restore image quality in case of transmission concealment technique employed in the decoder, which errors. H.263+ is an extension of H.263 and its aim is to discards corrupted GOBS and replaces the corresponding improve upon the compression efficiency of H.263 and image content with data from the previously decoded frame. broaden the range of applications for wireless transmission. This technique works almost perfectly for non-moving parts However, additional protection schemes must be incorporated of the sequence, but introduces severe distortion for moving in transmission over slow fading channels as a result of the image regions. In transmission over mobile channels, transmission errors long block error bursts characteristic of those channels. degrade the video quality at the receiver. In order to mitigate Recently, some papers have analyzed video transmission over MC-CDMA systems (e.g., [5]). However, none of them the effects of error transmission, periodic I-frame refresh has have considered MC-CDMA transmission over slow fading been proposed by several authors. However, since the overall channels, as well the effects of important ingredients of the image quality decreases due to the Iower efficiency of Iexisting system architecture (e.g., RAKE receivers with frames in comparison to P-frames, and the MC-CDMA system has a severe bandwidth limitation, no periodic I-frame refresh is applied in the simulations in order to achieve the This work was supported in part by SeCyT - UNC - Argentina
0-7803-7005-8/01/$10.00 0 2001 IEEE

Abstract - We present a simple and accurate semi-analytic methodology for performance evaluation of wireless video transmission in multicode DS-CDMA (MC-CDMA). Based on the novel approach (as well as on entire system simulations), we analyze performance of H.263 in an IS-95B system over slow fading channels at low bit rates. Peak signal-to-noise ratio (PSNR) values for various video test sequences demonstrate that important improvements of the video quality can be achieved by using feedback error control schemes. In addition,we investigate the effectiveness of the combined use of error correction and error resilience techniques for reliable video transmission in MC-CDMA systems.

433

Authorized licensed use limited to: UNIVERSIDAD CORDOBA. Downloaded on August 27, 2009 at 14:35 from IEEE Xplore. Restrictions apply.

highest coding performance for the error free case, which serves as a baseline. Loss of Picture Quality The average peak signal-to-noise ratio (PSNR) has been widely adopted as a distortion measure. For error-free transmission, the PSNR of the video reproduced at the receiver (i.e., the baseline) is given by
psN~cor,ec,) = 10lOg10 (255 1Dsc ( t ) ) . (t Dsc ( t ) is the video source coding distortion given by

matrix I , ( x ) = I, (1) , where I, I

C.

(7)

(1)

with

m,,,(x)

u , v = 0,1,2,...,2 M -1,

representing

the

probability of B, = v in slot i given that the superblock in slot i-x was Y l--x = U . The Markov model defined by (6) and (7) is called the super-block Markov model (SBMM). A in method to estimate the elements mu, transmission over u=l slow fading channels is presented in [6]. where U is the total number of samples of the picture at t , while 0,( t ) and r, ( t ) are the u-th amplitudes of the original B. A Model f o r the Loss o Picture Quality f and reconstructed pictures, respectively. Owing to Next w: derive an analytical expression for the loss of Pc transmission errors, an additional distortion DcH ( t ) occurs at frames picture quality owing to transmission errors, DcH(t). the video decoder. Thus, the overall decoded video distortion Towards this end, we assume that an en-or signal is (i.e.. after the error concealment stage) results in introduced at t = O (there is no error in the video frames or D c ~ ( t ) = 0 t < 0 ). The resulting error sequence is treated, for Do ( t ) = DSC (t> DCH (t> + . (3) before the video decoder, by the error concealment technique Then, the PSNR of the video sequence at the decoder can described previously. This introduces distortion for moving be expressed as image regions, which propagates spatially arid temporally until an I-fame refresh is applied at t = T,, . !$tuhlmiiller et PSNR,,, ( t ) = 10loglo(255 / D o( t ) ) . (4) al. [7] have shown that the variance of the prclpagated error In this paper, the loss of picture quality APSNR(t), signal over a P-frame sequence owing to the m o r introduced defined by at t = 0 can be well approximated at low error riates by
ApSNR(t) = PSNRlo,, (0- PSNRcorrec, 0) = lolog10 [Dsc ( t )1Do 011,
(5)
of ( t ) = f;r(i

+ et ) - l ,

Olt<T-.

(8)

is used as a measure of the video degradation.

In.PERFORMANCE EVALUATIONH.263 IN MC-CDMA OF

TRANSMISSION ANALYTICAL USING MODELS

A.

A Model for Block Error Process in MC-CDMA Our performance studies are based on a novel Markov model for the block error process in MC-CDMA transmission we introduced in [6]. Unlike previous work, this takes into account the presence of: (a) multicodes, (b) MRC RAKE receivers, (c) CC with SD-VD and (6)non-ideal interleaver performance (owing to transmission over slow fading channels). Consider a certain user that employs M

multicodes in an MC-CDMA system. Let

Pi,n be

a binary

process such that Pi, = 1 if the data block i at the Viterbi decoder output of the n-th multicode is in error, and 0 otherwise. We define a new process Y I as

We showed in [6] that the process

Pi is well

approximated by a 2 M - state Markov model, with transition

PB is the block error probability and depends on the transmission system and channel characteristics (e.g., channel code, numlxr of RAKE fingers, etc.). Parameter r represents the sensitivity of the video decoder to an increase in error rate, and its value depends on several implementation issues, such as packetization, resynchronization, and error concealment, as well as the encoded video sequence. It can be considered as a constant that does not depend on the other model parameters. Note that P,T is the error variance introduced in the video sequence at t = 0 [7]. The leakage 0 describes tlhe efficiency of explicit and/or implicit (e.g., due to sub-pel motion compensation, overlapped block transmission) loop filtering to remove the introduced error. Its value depends on the strength of the loop filtering as well as on the shape of the power spectral density of the introduced error. The range of typical values is given by 0 < 0 .:1. From (8) note that the energy of the error signal decays over time due to spatial filtering in the prediction loop. Unlike 1:7], in this work we are interested not only on the time avera,ged video distortion owing to transmission errors, but also c e the temporal progression of {he distortion value, DcH ( t ) . This model is very useful because it allows,

434

Authorized licensed use limited to: UNIVERSIDAD CORDOBA. Downloaded on August 27, 2009 at 14:35 from IEEE Xplore. Restrictions apply.

To perform our simulation experiments, we use the University of British Columbias H.263+ Reference codec. Moreover, we utilize the rate control method discussed in TMN-8. All of the results represent around of 160 pictures of test sequences having QCIF resolution, coded at 6 frames per second ( f p s ) . Although numerous simulations have been performed, only a few results are discussed in this paper, because of space constraints. Specifically, we analyze the loss for 0 5 t < T,, ,with of picture quality for the test sequences MotherdDaughter and Foreman. These sequences are selected because of their @ ( t )= 8 ln(1+ et)+ OS(2 + et11 + et)- . (10) different characteristic in motion and spatial detail [7]. (to derive (10) we approximated the sum in (9) by an integral We use the parameters, interleaver/deinterleaver, and using trapezoid based numerical integration). convolutional code of the downlink of the IS-95B standard PE can be easily estimated from the SBMM (this analysis [2]. Furthermore, we use soft-decision decoding of the rate %, is not included in this paper). For a given packetization and constraint length 9, convolutional code (i.e., rate set 2 of ISvideo sequence, we estimate parameters 8 .and r by fitting 95B). We set the number of RAKE fingers, L , to four (5) (with D,,(t) and D,,(t) given by (2) and (9), ( L = 4 ). We assume that RAKE fingers have equal power. respectively) to the measurement points obtained from The block rate is l/T, = 50 block/s. A 16-bit cyclic SBMM based fast simulation. From experimental results we redundancy code (CRC) is used for block error detection. have verified that r and 8 are approximately constant over Carrier frequency is 1800 M H z and the maximal Doppler the range of the video rates considered in this work (i.e., c32 frequency is f, = 2 Hz. Since the rate variation of the kbps). channel is small, we assume ideal coherent demodulation at the RAKE receiver. The Rayleigh channel is simulated using C. A Model for the Overall Video Distortion Jakes model. We set the number of multicodes to two (M=2). Using (9) in (3), the total distortion Do ( t ) results in FCC and an SCC are used to transmit video. The rate of SCC is 14.4 kbps. The transmission rate of FCC is assumed 7.2 D, ( t ) = D,, ( t )+ p B r o ( t ) , O I t <T,,. (11) kbps (14.4 kbps is its maximal rate). 2.4 kbps are spent for Then, from (11) we can derive the time averaged total overhead (i.e., packet headers + CRC + tail block) and the video distortion: video net bit rate results in 19.2 kbps. Three channel states are - analyzed in this paper: (a) Bad, (b) Good and (c) Very Do = D,, + P B r O , (12) Good, which have an SCC average block error of 0.11,0.035 and 0.017, respectively. These values correspond where approximately to O - (T,, + I ) - ~ . ~ - ~ [ ( ~ + B T , , ~ ~ ~ ( ~ + B T , , ) - I ) + ~ J code per user) in a15, 10 and 8 data users per cell (with one typical multicell environment. (13) + +e- ln(1+ BTm,)]+ 0.5[0(0)+ O(T,,)]}. A. Accuracy of SBMM We use SBMM to evaluate H.263 video degradation in Note that the minimum value, 0 =1, occurs when there is MC-CDMA transmission over slow fading channels. We also no error propagation (T,, = O ) . From (12) and (13) we can compare simulation results derived from both SBMM and conclude that the minimum distortion (i.e., 0 =1) of a given entire system simulations in order to verify the accuracy of video sequence achieved by error resilience techniques (such our channel modeling. It is important to realize that, although as error-tracking [9]) in transmission over a channel with Markov models have been previously used to evaluate block error probability PB,is given by performance of wireless video transmission, comparisons with results obtained from simulation of the entire system r n i n E } = D,, + P B r . (14) have not been reported so far. These comparisons are The distortion rate (DR) model is used to approximate important because it has been demonstrated that the firstD,, . Thus, the average video source coding distortion can order Markov approximation to model slow fading channels cannot be useful in several cases of interest (e.g., applications be expressed by requiring a large number of consecutive samples)[8]. Fig. 1 shows the loss of picture quality for various video test sequences obtained from simulations. Fast simulationwhere R, is the video source rate, while A. ,R, , and CO are based performance evaluation is achieved by using the the parameters of the DR model which depend on the encoded SBMM defined by (7). We also present results from sequence [7]. simulation of the entire system. The coded sequences are

for example, to evaluate analytically the performance of numerous error resilience techniques. Assuming that the error process is stationary over successive frames (i.e., the effect of lost blocks is approximately constant for each transmitted data block) and the error signals are uncorrelated [7], from (8) we obtain:

IV.NUMERICAL RESULTS

os^,,

435

Authorized licensed use limited to: UNIVERSIDAD CORDOBA. Downloaded on August 27, 2009 at 14:35 from IEEE Xplore. Restrictions apply.

transmitted 350 times using different: ( i ) realizations of the block error process (in SBMM simulations), and (ii) starting points in the fading simulator (in entire system simulations). The average MSNR(t) over all the runs is presented. We can verify the excellent accuracy of the SBMM. Results also show that video quality is severely degraded in transmission over MC-CDMA. This is because, in slow fading environments, the probability that all M received data blocks are in error in a deep fading is high [6],therefore long blockerror bursts occur at the input of the video decoder. If we take into account that P-frames are being transmitted, it is easy to infer that the error propagation problem in the reproduced video is severe, as it can be seen in Fig. 1. Furthermore, we have verified that several modes incorporated in H.263+ to improve performance in wireless channels are not effective (these results are not included in this paper). From the above, we conclude that extra protection must be provided in order to improve the reliability of H.263 transmission in MC-CDMA.
Analysis of Error Correction Techniques From Fig. 1 we noted that video quality is severely degraded in transmission over MC-CDMA. Improvements of video quality can be achieved by reducing the packet loss PB through error correction techniques such as automatic repeat request (ARQ) protocols. In this paper, we consider the NonSelective Variable Bandwidth Retransmission (NSVBR) scheme, a novel bounded delay modified GBN retransmission-based scheme that uses FCC to transmit, at a proper rate (lower than its maximal), video information instead o voice. Because real-time services require a bounded f delay, only one retransmission is allowed. Moreover, depending on the round trip delay and the number of MCCs, only a limited number of packets can be retransmitted. To retransmit data blocks with low time delay, the proposed approach increases the FCC data rate. Thus, our scheme not only avoids underutilizing and wasting the scarce bandwidth, but it also significantly reduces the interference and the complexity of the bandwidth allocation algorithm. Since slow fading and low data rates are assumed, it can be inferred that NSVBR provides a good tradeoff between efficacy and complexity. B.

Let N, 111 and R,,, denote the round trip delay in slots, the number of MCCs assigned to a given user, and the total user bit rate for video, respectively. In "normal conditions" (i.e., no errors), the FCC works at a data rate, RFccno, less than its , maximal RFCcma.The value of RFCCno, will depend on the link quality (e.g., the system load), and is determined at the beginning of the communication. For example, if the average error rate of the link is moderately high, a low value of RFccn0,is selected (i.e., the video bandwidth (quality) is reduced), thus a "high" bandwidth ( RFccma-- RFCc,, ) is reserved fix retransmissions. When the receiver (e.g., a mobile station) detects only one of the M received blocks in error, a negative acknowledgment (NAK) is sent to the transmitter (e.g., the base station). This approach simplifies implementation at the expense of efficiency. IHowever, we have verified that the degradation is not imporlant since the block error processes among MCCs are highly correlated [ 6 ] . Then, when the transmitter receives an NAK, the data rate of FCC is increased to RFccmm and the retransmission (only one) of MCC blocks begins. Unlike classical GB", not all the NM blocks are sent again since the bandwidth assigned for retransmission is limited. Only a fraction of the information lost is retr,uwnitted. For example, consider N = 6 and the transmissice rates defined in Section IV (i.e., R,, = RFccno,+ Rscc =21.6 kbps). Then, when an NAK is received, Ionly 33 % of the NM = 12 blocks can be retransmitted. When the retransmission process finishes, the system returns to the "normal" mode. The ability of NSVBR to improve the reliability of H.263 wireless video transmission in MC-CDMA is analyzed in Fig. 2. We adopt M, N ,R,,, , RFcc,, and RFCC,,, as given in the previous example. Moreover, we assume that no feedback errors OCCIK. We consider the channel state "Very Good'. Results obtained from entire system simulation and theoretical values are presented. Residual PB can, be estimated by using SBMM (this analysis is not included in this paper). We verify the excellent accuracy of both the distortion modeling approach (9) and the parameter estimation method based on SBMM. Moreover, note that NSVBR achieves significant gains for both test sequences with small additional complexity. These gains depend on the residud block error rate and the amount of motion present in the sequence. The biggest gain can be observed for Foreman [9].

C Analysis of Error Resilience Techniques .

Residual transmission errors cannot be avoided with a mobile radio channel. Additional schemes to stop the error propagatioin have to be considered. In this sense, NSVBR can be efficiently combined with error resilience techniques to further improve the video quality in MC-CDMA transmission.

Fnm. Numkr

Fnm. N u m k r

Fig. 1. Results for QCIF Foreman and Mother&Daughter sequences in IS-95Bfor different channel states.

436

Authorized licensed use limited to: UNIVERSIDAD CORDOBA. Downloaded on August 27, 2009 at 14:35 from IEEE Xplore. Restrictions apply.

Several cases are considered: no protection (NF), I-ET (only), and I-ET+NSVBR. The values are analytically derived from (12) with D,, given by (15). 0 is calculated from (13) for NP, while 0 = O(Td) for I-ET and I-ET+NSVBR (see (17)). We also include the theoretical limit defined by (14), which represents the minimum distortion of the video sequence that can be achieved by NSVBR. Results show that significant gains can be achieved when NSVBR is used in combination with I-ET. We also verify that a residual average distortion around 0.75 dB is added to the theoretical minimum value achievable by NSVBR. This result suggests that new error resilience schemes could be designed to improve even more the performance of NSVBR in video transmission over error-prone channels. V. CONCLUDING REMARKS This paper has presented a novel semi-analytical methodology to estimate performance of wireless video transmission in MC-CDMA systems. Based on this approach, we analyzed H.263 video transmission in IS-95B at low bit rates over slow fading channels. Comparisons with values derived form entire system simulation have demonstrated the excellent accuracy of our technique. We found that extra protection must be provided in order to improve the video quality of the end user. We also introduced a new reduced complexity feedback error control scheme, which has been shown to significantly improve the reliability of video transmission. Furthermore, we investigated the effectiveness of error correction and error resilience techniques, and discussed theoretical limits of the video quality that can be achieved in transmission over error-prone channels. The methodology we proposed in this paper can be extended to performance evaluation of several video codecs in MCCDMA transmission over generalized fading channels.

Fmm. Number

Fram. Number

Fig. 2. Results for QCIF Foreman and MotherBDaughter sequences in IS-95B for the channel state Very Good. Here we evaluate video performance achieved by NSVBR combined with a theoretical stop error propagation technique we called ideal-error-trucking (I-ET). This approach is similar to that one presented in [9], that is, it utilizes intracoded MBs (I-MBs) refresh to stop temporal error propagation. Using a feedback channel, the temporal and spatial Occurrence of an error is reported to the transmitter. Thus, the location and extent of propagated errors is reconstructed at the encoder. However, unlike [9] we assume that in I-ET the video source coding distortion D s c ( t ) is not affected by the I-MBs refresh, and the error signal at t = O is successfully canceled at t =Td . It can be verified that the video performance obtained from I-ET constitutes a theoretic upper bound of the video quality that can be achieved by any error resilience technique using the same conditions of operation (e.g., time delay Td ). The loss of the picture quality achieved by I-ET results in

while the time averaged video distortion is given by

DCH-Im

( t ) pBr@(Td) .

(17)

Table I presents results of the average distortion 0, for NSVBR and I-ET. Channel state Good is used and the time delay between the video coder and decoder is assumed Td = 0.8 s.

Video Sequence
M&D
Foreman

Average Video Distortion (dB)

NP
18.58 23.86

I-ET

17.73
21.83

I IET+NSVBR I I 16.93 I
2 1.09

Min.Dist

16.21 20.30

REFERENCES Chih-Lin I, R. D. Gitlin, Multicode CDMA wireless personal communications networks, IEEE lCC95, v01.2. TWEIA-95B. Mobile station-base station compatibility standard for dual-mode wideband spread spectrum cellular system. July 1997. D. Ayyagari and A. Ephremides, Cellular multicode CDMA capacity for integrated (voice and data) services, IEEE JSAC, vol. 17, pp. 928-938, May 1999. ITU,Tel. Stand. Sector ITU,Video coding for narrow tel. channels at< 64 kbit/s, lTU-T Recom. H.263, 1996. B. Deep and W.Feng, Adaptive code allocation in multicodeCDMA for transmitting H.263 video, Proc of IEEE WCNC, pp. 1003-1007, New Orleans, September 1999. M. R. Hueda, A Markov model for performance evaluation in multimedia CDMA wireless transmission Proc. IEEE-VTC, Boston, Sept. 2000. K. Stuhlmiiller, N. F;irber, M. Link, and B. Girod, Analysis of video transmission over lossy channels, IEEE JSAC., vol. 18, pp. 1012-1032, June 2000. C. Tan and N. Beaulieu, First-order Markov modeling for the Rayleigh fading channel, IEEE Globecom98, pp.3669-3674. E. Steinbach, N. F&ber, and B. Girod, Standard compatible extension of H.263 for robust video transmission in mobile environments, IEEE T. CSVT., vol.7.p~. 872-881. Dec. 1997.

437

Authorized licensed use limited to: UNIVERSIDAD CORDOBA. Downloaded on August 27, 2009 at 14:35 from IEEE Xplore. Restrictions apply.