Accurate wireless channel modeling for efficient adaptive Forward Error Correction in JPEG 2000 video streaming systems

Transcription

1 International Journal of Engineering Research and Develoment e-iss: X, -ISS: X, olume 0, Issue (December 04), PP Accurate wireless channel modeling for efficient adative Forward Error Correction in JPE 000 video streaming systems Max Agueh, Magaye Dio, ACSC ECE, 37 Quai de renelle, 7575 Paris Cedex 5, France ESP, Université Cheikh Anta Dio, P 5085 Dakar, Sénégal Abstract:- In this aer, we evaluate the imact of accurate 80. based wireless channel modeling on the efficiency of dynamic Forward Error Correction (FEC) schemes in Motion JPE 000 video streaming systems. We derive a comromise on the suitable trace length for ractical estimation of Packet Error Rate (PER) at decoder side. We demonstrate the validity of the derived trade-off using a real JPE 000 based video streaming system. Index Terms:- Wireless channel modeling; trace length; analysis window length; dynamic Forward Error Correction; Packet Error Rate estimation; Motion JPE 000 video streaming. I. ITRODUCTIO With the develoment of wireless multimedia streaming systems there is an increasing demand on alication level tools which could increase the robustness of video streaming over wireless networks. JPE 000 [], the newest image reresentation standard, addresses this issue by redefining error resilient tools in his core encoding system (art ) and going straightforward by defining in its th art called wireless JPE000 ( JPW) [] a set of error resilient techniques to imrove the transmission of JPE000 code streams over errorrone wireless channel. One of the mains recommendations of JPW final draft is the use of Forward Error Correction (FEC) with Reed-Solomon. Hence, in the literature, authors [3][4] roosed FEC rate allocation schemes for robust JPE 000 images and video streaming. These schemes statically [3] or dynamically [4] allocate JPE 000 code streams channel coding rate based on the estimation of resectively it Error Rate (ER) and Packet Error Rate ( PER ). In [4], at the decoder side, error rate metrics are derived after frame decoding and these metrics are used, at the encode side, to derive an alication level channel model. The FEC rate allocation scheme relies on the derived channel model to select the aroriate channel codes for next frames rotection. Hence, in wireless images and video streaming systems such as the one considered in this work (see Figure ), the effectiveness of FEC rate allocation schemes deends on the accuracy of error rate estimation and channel modeling in the system. Figure. Wireless multimedia system In this aer, we evaluate the imact of accurate Packet Error Rate (interchangeably it Error Rate) on the effectiveness of JPE 000 based FEC rate allocation schemes. Then, we define an analysis window length as the length of the used trace for PER estimation at decoder side. Ideally, the analysis window length should be the highest ossible for accurate channel modeling, however due to the real time constraint this value must not be too high. y the other side, a very short analysis window length leads to bad channel modeling because the estimated PER is not statistically reresentative. Finally, we derive a trade-off oint for ractical estimation of analysis window length and we validate this comromise using an alication of Motion JPE 000 video streaming over real ad-hoc networks traces. To the best of our knowledge the resent work is the first to give ractical clues (comromise on analysis window length) for accurate wireless channel modeling in order to enhance the effectiveness of the FEC rate allocation schemes used in Wireless JPE 000 images and video streaming systems. 30

2 Accurate wireless channel modeling for efficient adative Forward Error Correction in JPE 000 video The aer is arranged as follows. Section II resents the two FEC rate allocation schemes considered in this work for robust JPE 000 images and video streaming over wireless networks. In section III, the loss atterns generation is described and an analysis of the considered real wireless channel traces is rovided. A comromise on the suitable analysis window length for efficient FEC rate allocation schemes is also derived in section III. The results validating the derived emirical trade-off are resented in section I. Finally, some conclusions are rovided in section. II. FEC RATE AOCATIO SCHEMES In this aer we considered two low-comlexity Forward Error Correction schemes used resectively in [3] and [4] for robust JPE 000 based images/video streaming. A. ayered Unequal Error Protection Scheme Z. uo and al roosed in [3] a slightly comlex layered unequal error rotection scheme for robust Motion JPE 000 streaming over wireless network. This FEC rate allocation scheme relies on the contribution of each layer to the imrovement of image quality to select the suitable RS codes. Hence owerful RS codes are set to most imortant layers such as the base layer and the other layers are rotected by decreasing order of imortance. However, this algorithm is not JPW comliant and was designed based on the assumtion that the channel is a memoryless inary Symmetric Channel (uncorrelated error occurrence) which is not realistic because wireless channels have correlated errors sequences.. Dynamic FEC rate allocation scheme The second considered FEC rate allocation scheme is resented in [4] and is a dynamic layered based unequal error rotection FEC rate allocation methodology for efficient JPE 000 streaming over MAET. This scheme is based on the assumtion that transmitted JPE 000 image quality is linked to the amount of correctly decoded ackets at the receiver. Hence, goal of this scheme is to maximize the overall throughut in the system. The dynamic FEC rate allocation scheme imroves the erformance by about 0% comared to a riori selection of channel coding. The drawback of both methodologies is that their effectiveness deends on the accuracy of error rate estimation at decoder side. This aer addresses this issue by relying on real 80. traces analysis to derive a trade-off on the estimation window length for accurate Packet Error Rate evaluation. In the following section we analyse the real MAET traces. III. APPICATIO EE CHAE MODEI A. oss Patterns eneration The latform used to generate the loss atterns is resented in Figure. It consists of a client/server software air running on two Windows XP latos connected in ad-hoc network using two PCMCIA IEEE 80. b/g cards (at,4 Hz). Figure. oss atterns generation latform The set of loss atterns generated covers different transmission scenarios (mobile or static). Each attern corresonds to a secific Carrier to oise ratio C ( C is the ratio between the desired signal and the total received noise ower).the used mode at the hysical layer of the wireless link is the mode 4 where the modulation is QPSK, the coding rate is 3/4 and the ominal Data Rate R min is 8Mbit/s. In the considered loss atterns, C varies between 0 d and d which corresonds to a Packet Error Rate ranging from 3 0 to enerated traces are available in [5]. o al 3

3 Accurate wireless channel modeling for efficient adative Forward Error Correction in JPE 000 video. Modeling oss Patterns with ilbert Model The ilbert model was first introduced by ilbert in [9]. Elliot rooses an extension of the ilbert model in [0], the last model is commonly known as the ilbert-elliot (E). In E model, the wireless channel is modelled to have two states: good and bad. In the good state ( g ), the channel rovides a constant and low error robability ( P ) whereas in the bad state ( b ), the channel exeriences a high error robability ( P ). Hence we have P << P for E, and P 0 and P for the ilbert channel, in other words the ilbert model is a simlified E model. In this work, we consider an 8-bit symbol oriented model to emulate the correlated error characteristics of wireless channel. Therefore, our wireless channel is modelled as a two state Markov rocess (see figure 3).With this model, the channel roduces error bursts. gg bb Figure 3. Two state Markov rocess scheme et and be resectively the mean length of error free and erroneous sequences, we have: and gg bb Alying Markov rocess at symbol level, we derive the Symbol Error Rate ( SER ) for ilbert model [9]: gg SER bb gg A comrehensive descrition of the Markov modeling for wireless channel is exlained in []. C. Trace Analysis under ilbert Framework It is worth noting that in the considered traces each acket had a fixed length of 8 symbols (bytes). In our work we do not consider a cross layer estimation of transmission errors, so the Symbol Error Rate ( SER ) is equal to the Packet Error Rate ( PER ). Therefore, acket oriented ilbert models derived from traces had the same characteristics and same arameters than the 8-bit symbol oriented ilbert models used to emulate the wireless channel at alication level. As loss atterns are alied on ackets, for coherence reason, a acket oriented analysis of the traces is resented. In the loss atterns ood state () and bad state () are reresented resectively by 0 and. Hence 0 corresonds to a well received acket and to an erroneous acket. The distribution of error bursts is resented in figure 3-a for different loss atterns. From figure 3-a, it can be noticed that, in most cases, the error burst length is less than 0 ackets. So max 0 is considered as the uer bound of the error burst length. The error free burst length distribution in figure 3-b shows that the uer bound max 00 is ten times suerior to the error bursts. This is due to the fact that aart from the moment when the wireless channel exerience fading (bursts of errors), the transmission is often successful. We also notice that the number of error free bursts is lower than the number of error bursts, but this ga is comensated by the stay time in error free state (error free burst length) which is much longer than the one in error state (error burst length). So in the models used to emulate the traces, the mean time in the good state should be sensibly greater than the mean time in the bad state. 3

4 Accurate wireless channel modeling for efficient adative Forward Error Correction in JPE 000 video Figure 3-a. Error bursts distribution Figure 3-b. Error free bursts distribution Figure 4-a resents the mean error burst length evolution for different traces. An exonential increase of the mean error burst length with the Packet Error Rate can be observed. It is worth noticing that even when the channel is highly noised, tyically0 PER 0, the mean error burst does not vary, and is about.05 ackets. It can be deduced that our channel exeriences relatively fast fading eriods. In figure 4-b, the error burst length standard deviation is resented, in order to evaluate the fluctuation of the error burst length around the mean error burst length of the considered traces. Figure 4-a. Error burst mean length versus PER Figure 4-b. Error burst length standard deviation versus PER 33

5 Accurate wireless channel modeling for efficient adative Forward Error Correction in JPE 000 video Figure 4-b indicates that for slightly noised channel, tyically 0 PER 0 standard deviation is low, meaning that error burst length is closed to the mean value, the error burst length, whereas for highly noised channel, tyically PER 0, the error burst standard deviation is high, leading to a sensible fluctuation of error burst length around the mean value. Similarly figure 5-a shows that the mean error free burst length decreases exonentially with the Packet Error Rate. Hence this metric varies between 00 and 00 ackets for PER 0, and 00 and 0 ackets for 0 PER 0. For PER 0 the mean error free burst length is 0 ackets, which is ten times higher than the mean error burst length, demonstrating that even for highly noised channel the ratio between error free burst length and error burst length is still higher. It comes under view that the channel exeriences longer error free sequence than error sequence which confirms the conclusions derived from figure 3-a and figure 3-b. Figure 5-a. Error free burst mean length versus PER Figure 5-b. Error free burst length standard deviation versus PER In figure 5-b the standard deviation of the error free burst length versus the Packet Error Rate is resented. We notice that for PER 0, error free burst length standard deviation is high meaning that error free burst length fluctuates a lot around the mean value whereas for PER 0 this metric is sensibly reduced making the error burst length closer to the mean value. Thanks to loss atterns analysis, interesting information can be derived about the wireless channel such as the mean error free burst length and the mean error burst length. In revious section, we shown that, and its associated standard deviation increase exonentially with PER whereas 34 and its associated standard deviation behave inversely. Those metrics are then used to derive the ilbert model arameters and using the relation verified by R. Jain in []: and. In our work each loss attern had a fixed length of 0000 ackets. et evaluate, the imact of the considered traces length on the accuracy of its associated ilbert model. D. Imact of Analysis Window length on the Accuracy of ilbert Models In order to estimate the imact of the trace length on the accuracy of generated ilbert models, the analysis window and the arameters and are fixed. Then, a Monte-Carlo simulation is conducted using 000 artificial traces of same length. For each artificial trace, the arameters ˆ and ˆ are estimated. et and be the relative ga between the arameters estimated from artificial traces and from the

6 Accurate wireless channel modeling for efficient adative Forward Error Correction in JPE 000 video fixed ilbert model arameters, we have: variance resectively of resented in figure 6-a. and ˆ and. The arameter ˆ. et and be the versus the analysis window length is Figure 6-a. ariance of delta P_ versus analysis window length Figure 6-b. ariance of delta P_ versus analysis window length We notice that increasing the analysis window length decreases the variance of increases the accuracy on the estimation of the estimated and so. In other words, the more the trace length is long the more accurate is the estimated ilbert model. This is due to the fact that the estimation of ilbert arameter, relies on mean error free burst length whose value becomes statistically reresentative with the increase of the trace length. Moreover, we observe that increasing PER leads to a sensible reduction of the error made in the estimation of. Indeed, for PER.66*0 and with an analysis window of 000 ackets, we have.5*0 3 whereas for the same 000 and PER 5.*0 we have close to 0. In the last case, the imrovement in the accuracy of the model is about 60 times higher than in the revious case. Hence, the more the channel is erroneous, the more the estimation of is accurate. This is exlained by the fact that increasing PER decreases the error free bursts length standard deviation as shown in figure 5-b. As the accuracy of is linked to statistically reresentative values of, it comes under view that the considered arameter estimation changes with the PER whose sensitively imacts. However, as shown in figure 6-b this is no longer the case for the estimation of arameter. From figure 6-b, it can be noticed that for analysis 35

7 Accurate wireless channel modeling for efficient adative Forward Error Correction in JPE 000 video window length of 000, estimation of is slightly sensitive to PER, because we do not have enough error bursts to derive sufficiently reresentative measures of in which we rely to estimate. However, as far as significantly increases, we have a better estimate of and its estimation became less imacted by the PER increase for a considered analysis window length contrarily to 36.This is exlained by the fact that increasing PER increases the number of error bursts but does not increase their length whose mean remains constant as shown in figure 4-a. Finally, we can deduced from this section that for accurate ilbert model arameters estimation, an analysis window length of at least 000 ackets (corresonding to 0% of the overall trace length) should be considered. oing straightforward, we extend this trade-off to video streaming systems. Hence, denoting as the video length, an interesting window length for accurate PER estimation is 0.. As an examle, for a Motion JPE 000 video constituted by 00 frames, the PER at receiver should be estimated at least every twenty JPE 000 frames in order to guaranty the efficiency of the FEC rate allocation scheme. In the following section, we validate this emirical comromise using a Motion JPE 000 video streaming over real MAET traces. I. JPE000 IDEO STREAMI RESUTS The interest of this section is to validate the comromise on PER estimation window length in a real dynamic FEC rate allocation system. The video sequence used is the seedway.mj [3] containing 00 JPE000 frames at an overall quality of 0. b (bit er ixel) with 0.05 b for base layer, 0. b for the second layer and 0. b for the third layer. As error occurrence in the transmission channel is a random rocess, different runs were made for each simulation and the Mean Square Error ( MSE ) between the original image ( I o ) and the decoded image ( I d ), is averaged over all the runs in order to have statistically reresentative metrics. The measured Peak Signal to oise Ratio (PSR) is obtained as follows: M MSEI o, I d I o ( x, y) I d ( x, y) M. x y MSE and MSE PSR 0*log 0 frames Where MSE is the Mean Square Error over all the streaming, 55 MSE images considered. In the case of Motion JPE000 frames reresents the 00 JPE000 frames constituting the video sequence Each PSR measure is frames associated to a decoding rate metric which corresonds to crash estimation on the basis of frames transmission trials. The FEC rate allocation schemes used are the one resented in section. Thanks to the wireless multimedia system resented in figure, the effectiveness of both data rotection schemes is evaluated by comuting the PSR at the outut of the system. Through a client/server alication the JPE000 frames extracted from the Motion JPE000 (MJ) file are transmitted to the receiver which reresents the wireless client. The Packet Error Rate is estimated at the decoder and sent back to the encoder via the ulink in order to imrove next frames rotection levels. Thanks to a better knowledge of the channel, the encoder selects a set of RS codes to rotect each layer of JPE 000 frames. In our work, the wireless channel is emulated using real loss atterns [5] resented in section 3. The considered loss atterns have C 8 d and the Motion JPE 000 file (seedway.mj) length is 8704 bytes. Figure 7 illustrates the selected RS( n, k) codes error correcting ability n k t versus the PER estimation window length while using the FEC rate allocation schemes resented resectively in [3] and [4]. Since the best set of selected RS codes are achieved with the maximal window length, we notice that PER estimation (analysis) window length of 0. is enough to yield the suitable rotection level for JPE 000 code streams. Using a PER estimation window length higher than the roosed trade-off leads to the same set of RS codes selection but decreases the reactivity of the system and by the way its effectiveness because PER values are

8 Accurate wireless channel modeling for efficient adative Forward Error Correction in JPE 000 video udated less frequently. On the over side, using a PER estimation window length lower than the roosed tradeoff leads to inaccurate channel modeling which reduces the efficiency of the rotection schemes. Figure 7. Correcting ability of the RS codes used for MJPE 000 data rotection andwidth 8 Mbits/s Table I and table II show the PSR of decoded video sequence versus estimation window length while using the FEC rate allocation schemes resented resectively in [3] and [4]. TAE I: PSR OF IDEO SEQUECE AD IMAES SUCCESSFU DECODI RATE ERSUS PER ESTIMATIO WIDOW ETH (Z. UO) Window length PSR (d) Successful (bytes) Decoding Rate (%) (8704 bytes) (87 bytes) (88 bytes) TAE II: PSR OF IDEO SEQUECE AD IMAES SUCCESSFU DECODI RATE ERSUS PER ESTIMATIO WIDOW ETH (DYAMIC FEC) Window length PSR Successful (bytes) (d) Decoding Rate (%) (8704 bytes) (87 bytes) (88 bytes) 4. 9 We notice from table I and table II, that the PSR of decoded video sequence is almost the same for PER estimation window length suerior or equal to 0. ( 46 d ). However, the PSR decreases about 5 d under the roosed trade-off oint which imlies a reduction of the video quality decoded by the wireless client. It is worth noting that the successful decoding rate is often higher than 90% which means that in our work even in case of inaccurate channel modeling, both FEC rate allocation schemes select sufficiently owerful RS codes to avoid decoder crashes. We conclude that the roosed PER estimation window length is a valid trade-off for ractical imlementation of FEC rate allocation schemes in Motion JPE 000 video streaming systems. An interesting extension to this work could be the comarison of the results obtained using the roosed emirical window threshold to those achievable while using the sliding window algorithm roosed in [4]. 37

9 Accurate wireless channel modeling for efficient adative Forward Error Correction in JPE 000 video. COCUSIO In this aer we evaluate the imact of the accuracy of the channel models on the effectiveness of alication level FEC rate allocation schemes. We start by resenting the considered FEC rate allocation schemes for robust Motion JPE 000 video streaming alication over wireless channels. After analysing real 80. based traces we derived a comromise for the PER estimation window length at receiver side. The roosed estimation window length is about ten time less than the transmitted video length. We then show the validity of the roosed trade-off using an alication of Motion JPE 000 video streaming of JPE 000 traces. REFERECES []. D.S. Taubman et M.W. Marcellin, JPE 000 Image Comression Fundamentals, Standards and Practice, Kluwer Academic Publishers, The etherlands 00 []. JPE 000 art Final Draft International Standard, ISO/IEC JTC /SC 9/W 3797 [3]. Z. uo, Y. ishikawa, R. Y. Omaki, T. Onoye and I. Shirakawa, A ow-comlexity FEC Assignment Scheme for Motion JPE 000 over Wireless etwork, IEEE Transactions on Consumer Electronics, ol. 5, Issue, Feb. 006 Page(s): 8 86 [4]. M. Agueh, J-F Diouris, M. Dio and F-O Devaux, Dynamic channel coding for efficient Motion JPE 000 streaming over MAET, Proc. Mobimedia007 conf, August 007, afaktos, reece [5]. oss atterns acquired during the WCAM Annecy 004 measurement camaigns IST WCAM, Wireless Cameras and Audio-isual Seamless etworking, roject website: htt:// [6]. M. Agueh, J-F Diouris, Alication level channel modelling in wireless network: examle of JPE 000 images/video streaming over wireless channel, In Proc. IRAMUS Worksho, Jan 007,al thorens, France [7]. Almudena Konrad, en Y. Zhao, Anthony D. Joseh, and Reiner udwig, A Markov-ased Channel Model Algorithm for Wireless etwork, Proc. International Worksho on Modeling, Analysis and Simulation of Wireless and Mobile Systems (MSWiM),00 [8]. Ping Ji,. iu, D. Towsley, Zihui e, J. Kurose, Modeling Frame-level Errors in SM Wireless Channels, Internet erformance symosium. 55, issue - (Jan 004) [9]. E.. ilbert, Caacity of a burst noise channel, ell Syst. Tech. J., vol. 398, , Set 960 [0]. O.Elliot, Estimates of error rates for codes on burst-noise channel, ell Syst.Tech. J., vol. 4, , Set 963 []. Julio Arauz, rashant Krisnamurthy, Markov Modeling of 80. channels, In ehicular Technology Conférence, 003.TC 003-Fall. 003 IEEE 58 th. ol., , 6-9 Oct. 003 []. R. Jain, The art of comuter systems erformance analysis, [3]. Seedway video sequences have been generated by UC. Available:htt://eutere.tele.ucl.ac.be/WCAM/ublic/Seedway%0Sequence/ [4]. Z. i, J. Chakareski, X. iu, Y. Zhang and W. u, Modeling and Analysis of Distortion caused by Markov-model burst acket losses in ideo transmission, submitted to IEEE Trans. On Circuits and Systems for video Technology, 008, Ref: TS-article Available: htt://infoscience.efl.ch/record/389?ln=fr 38