University of Southern California Dissertations and Theses

Adaptive video transmission over wireless fading channel

(USC Thesis Other)

Adaptive video transmission over wireless fading channel

PDF

Download a page range

Download transcript

Copy asset link

Request this asset

Transcript (if available)

Content ADAPTIVE VIDEO TRANSMISSION OVER WIRELESS FADING CHANNEL by W uttipong Kumwilaisak A Dissertation Presented to the FACULTY OF THE GRADUATE SCHOOL UNIVERSITY OF SOUTHERN CALIFORNIA In Partial Fulfillment of the Requirements for the Degree DOCTOR OF PHILOSOPHY (ELECTRICAL ENGINEERING) May 2003 Copyright 2003 W uttipong Kumwilaisak Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. UMI Number: 3103921 UMI UMI Microform 3103921 Copyright 2003 by ProQuest Information and Learning Company. All rights reserved. This microform edition is protected against unauthorized copying under Title 17, United States Code. ProQuest Information and Learning Company 300 North Zeeb Road P.O. Box 1346 Ann Arbor, Ml 48106-1346 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. UNIVERSITY OF SOUTHERN CALIFORNIA THE GRADUATE SCHOOL UNIVERSITY PARK LOS ANGELES, CALIFORNIA 90089-1695 This dissertation, written by ) 1 ! under the direction o f h dissertation committee, and approved by all its members, has been presented to and accepted by the Director of Graduate and Professional Programs, in partial fulfillment of the requirements fo r the degree of DOCTOR OF PHILOSOPHY Director Date May 1 6 , 2 0 0 3 Dissertation Committee Chair Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. D ed ication To my advisor Dr. C.-C. Jay Kuo, for helping me get to where I am today. To my parents, my wife KyoungMi Youn. and my country, for pushing me to finish this thesis. Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. A cknow ledgem ents First, I would like to acknowledge the continuous guidance and support from my thesis advisor, Professor C.-C. Jay Kuo. W ith his insightful advice and ideas, he has made my research experience enjoyable and invaluable and developed my full potential to become a m ature researcher. I would also like to thank Professor Antonio Ortega, Professor Roger Zimmer- mann, Professor Srikant Narayanan and Dr. Jongwon Kim for serving on my com mittee. A special thanks should go to my mentor, Dr. Jongwon Kim, for his great support and guidance for this research during his stay at USC. I am also grateful to Dr. Frank Hartung of the Ericssion Research Laboratory in Germany, and Dr. Qian Zhang and Dr. Wenwu Zhu of the Microsoft Research Asia in China for offering me opportunities to work in these two prestigious Labs in the Summer of 2001 and 2002, respectively. The two summer internships enriched my research and life experiences significantly. Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. I would also like to express my gratitude to all members of my research group at USC for creating a pleasant research environment. The good time that we spent together is unforgettable in my memory. Finally, I gratefully acknowledge the generous support from my family. I owe my deepest gratitude to my father and mother for their dedication to my education. I am also much indebted to my wife, KyoungMi Youn, for her endless understanding and love. My appreciation to her cannot be adequately expressed in words. Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. C o n te n ts D edication ii A cknow ledgem ents iii List o f Tables viii List of Figures ix A bstract xiv 1 Introduction 1 1.1 Significance of the R e s e a rc h .......................................................................... 1 1.2 Issues of Wireless Video Transm ission.......................................................... 5 1.2.1 Wireless Channel Modeling ............................................................. 5 1.2.2 Channel C o d in g ..................................................................................... 6 1.2.3 Source C oding........................................................................................ 8 1.2.4 Joint Source Channel C o d in g .......................................................... 10 1.2.5 Quality of S ervice................................................................................. 11 1.2.6 Cross-Layer D e sig n .............................................................................. 12 1.3 Contribution of the R esearch.......................................................................... 12 1.4 Outline of the d isse rta tio n .............................................................................. 15 2 W ireless V id eo Transm ission: Background R eview 16 2.1 In tro d u ctio n ......................................................................................................... 16 2.2 Wireless Video Communication S y s te m ...................................................... 17 2.2.1 Video Transmission System: ITU-H.324 18 2.3 A daptation Techniques in Wireless Communication S y ste m s................................................................................ 23 2.3.1 Review of Previous W o rk .................................................................... 25 2.4 Optimization T echniques................................................................................. 28 2.4.1 Lagrangian optimization approach................................................... 29 2.4.2 Dynamic programming approach ................................................... 29 2.4.3 Review of Previous W o rk .................................................................... 30 v Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 2.5 Review of Robust Wireless Visual Communications .............................................................................................. 32 2.5.1 Multi-resolution Based Video T ransm ission................................. 32 2.5.2 DCT-Based Video T ransm ission..................................................... 34 2.5.3 Application A daptation and Cross-Layer D e sig n ....................... 35 3 N on-station ary Fading C hannel M odeling w ith A daptive Variable L ength M arkov C hains (VLM C) 36 3.1 In tro d u ctio n ....................................................................................................... 36 3.2 Overview'of the Proposed Adaptive Channel M o d e lin g ....................... 41 3.3 Mapping from Physical Measurements to Dynamic Variable Length Markov Chains (V L M C ).............................................................................................................. 45 3.3.1 Fading Channel SNR Distribution E s tim a tio n .......................... 45 ■3.3.2 Discrete Markov M o d elin g ............................................................... 49 3.4 Channel Model U p d a te ................................................................................... 62 3.5 Applications of Derived VLMC ................................................................... 65 3.5.1 Fading Parameter E s tim a tio n ......................................................... 66 ■3.5.2 Adaptive Packet-Size T ra n sm issio n ............................................... 68 3.6 Experimental R e s u lts ....................................................................................... 71 3.6.1 Stationary Performance ................................................................... 72 3.6.2 Adaptive Channel M o d e lin g ............................................................ 82 3.7 C onclusion........................................................................................................... 87 4 D ynam ic QoS M anagem ent for P rioritized W ireless V ideo D elivery 89 4.1 In tro d u ctio n ........................................................................................................ 89 4.2 Review^ of Previous W o rk ................................................................................ 94 4.3 Dynamic QoS Management Framework for Time-Varying Wireless Channels ............................................................... 95 4.4 Rate Constraint Derivation for Priority Networks.............................................................................................................. 98 4.4.1 Time-Varying Service-Rate Wireless C hannel.............................. 98 4.4.2 Effective Bandwidth and Capacity ..................................................101 4.4.3 Derivation of Transmission Rate Constraint .................................104 4.5 QoS Mapping Mechanism ................................................................................109 4.5.1 Solution to the Optimization Problem ........................................... 114 4.6 Interaction between Prioritized Video and Priority N etw orks................................................................................................ 118 4.6.1 QoS B o u n d ............................................................................................. 118 4.6.2 QoS A daptation via Video-Network I n te r a c tio n .......................... 119 4.7 Experimental R e su lts ..........................................................................................123 4.7.1 Transmission Rate C o n s tra in t............................................................123 vi Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 4.7.2 Adaptive Video Transmission with Dynamic: QoS A d ju s tm e n t..............................................................................................128 4.8 C onclusion............................................................................................................. 131 5 Prioritized V ideo Transm ission w ith A daptive FEC P ro tection 134 5.1 In tro d u ctio n ..........................................................................................................134 5.2 System Configuration and Video Packetization.......................................... 137 5.3 Corruption Model for Priority A ssig n m e n t.................................................139 5.4 Wireless Channel M o d e l.................................................................................. 141 5.4.1 Estimating and predicting channel s ta t u s .........................................145 5.5 Concatenated Code Utilization ..................................................................... 146 5.5.1 Rate-Compatible Punctured Convolutional (RCPC) Code . . . 147 5.5.2 Reed-Solomon (RS) Code and Concatenated C o d e ........................ 149 5.5.3 Concatenated Code Rate S e le c tio n ................................................... 152 5.6 Optimized Adaptation M echanism ..................................................................153 5.6.1 Problem F orm ulation .............................................................................. 153 5.6.2 Dynamic Programming Solution to Adaptation and Packeti zation 157 5.7 Simulation R e s u l t s .............................................................................................161 5.8 C onclusion............................................................................................................. 166 6 C onclusion and Future W ork 169 6.1 C onclusion..............................................................................................................169 6.2 Future W o rk ..........................................................................................................171 R eference List 174 vii Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. List of Tables 1.1 Different service classes and applications of 3G systems......................... 2 3.1 The parameters of the VLMC channel model for the Rayleigh fading channel................................................................................................................... 73 3.2 The param eters of the VLMC channel model for the log-normal shad owing channel....................................................................................................... 74 3.3 Performance for the Rayleigh fading channel approximation with var ious channel adaptation speeds....................................................................... 86 5.1 The mean and the standard deviation of RPI for different sequences. 142 5.2 The video packet drop percentage in different RPI intervals for the Glasgow sequence with SNR equal to 16dB under the RPI-aware and the RPI-blind modes.............................................................................................163 5.3 The video packet drop percentage of the Glasgow sequence in compar ing the RPI aware system among the optimized adaptive concatenated code with other adaptive schemes with SNR equal to 16dB......................167 viii Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. List of Figures 1.1 Relationship between core-networks and their air interface alterna tives in 3G wireless communication systems 3 1.2 Adaptive cross-layer design and operation................................................... 13 2.1 The block diagram of an H.324 multimedia term inal............................... 19 2.2 The structure of an H.263 video encoder. . ........................................... 20 2.3 Components of the H.223 standard................................................................ 22 2.4 The refigurable video transceiver system...................................................... 27 2.5 Illustration of the Lagrangian optimization principle............................... 30 2.6 The structure of dynamic programming....................................................... 31 3.1 The non-stationary nature of a fading channel........................................... 42 3.2 The proposed VLMC channel modeling system for a non-stationary fading channel...................................................................................................... 44 3.3 The fading SNR distribution estimate for 5000 simulated observations from Rayleigh fading channel with different Normalized Doppler fre quency.................................................................................................................... 47 3.4 The Kullback-Liebler measure between the actual and the estimated channel SNR distribution of the Rayleigh fading as a function of the number of samples used to estimate the channel SNR distribution parameterized by normalized Doppler frequency........................................ 48 3.5 The partitioning of the channel SNR range................................................. 49 ix Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 3.6 Comparison of bit error probabilities obtained by theory and the pro posed fading partitioning m e c h a n ism ......................................................... 53 3.7 Comparison of bit error probabilities obtained by equal partitioning and the proposed iterative fading partitioning mechanism, with 16 intervals under the BPSK modulation and the Rayleigh fading envi ronment ............................................................................................................... 54 3.8 The tree structure of VLMC of order 2, where the set of contexts representing the state of fading channel SNR is {a, ba, bb, be, bd. c.d}. . 59 3.9 The finite-state representation of a fading channel..................................... 60 3.10 A daptation of the tree structure along with the changing of fading statistics................................................................................................................. 65 3.11 The autocorrelation comparison between the closed form of the log normal shadowing and those obtained from VLMC channel modeling in urban environment......................................................................................... 76 3.12 The autocorrelation comparison between the closed form of the log normal shadowing and those obtained from VLMC channel modeling in suburban environment.................................................................................. 77 3.13 The comparison of autocorrelations obtained form the closed-form for mula and the VLMC channel modeling with the normalized Doppler frequency equal to 1CT3 for Rayleigh fading................................................ 78 3.14 The comparison of autocorrelations obtained form the closed-form for mula and the VLMC channel modeling with the normalized Doppler frequency equal to 10"2 for Rayleigh fading................................................ 79 3.15 Comparison of level crossing rates obtained from the VLMC model and computed from the closed form solution for the Rayleigh fading channel................................................................................................................... 80 3.16 Comparison of the averaged fade duration computed from the VLMC model and the closed form solution for the Rayleigh fading channel. . 81 3.17 Comparison of the averaged channel capacity computed from the VLMC model and the closed form solution for the Rayleigh fading channel with the normalized Doppler frequency of 1CT2.......................... 82 x Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 3.18 Performance comparison between adaptive and non-adaptive channel model in terms of the difference of the channel capacity (top) and the KL distance (middle) under a time-varying power environment as indicated in the bottom of the figure............................................................. 84 3.19 Performance comparison between adaptive and non-adaptive channel models under a time-varying mobile speed environment.......................... 85 3.20 The throughput comparison using adaptive and fixed packet sizes over a non-stationary wireless channel................................................................... 88 4.1 The system architecture of QoS adaptation for a wireless video trans mission system..................................................................................................... 97 4.2 The discrete Markov model of a wireless time-varying service-rate channel................................................................................................................... 99 4.3 Dependency of transmission rate constraint of transm itting multiple priority classes with QoS......................................................................................109 4.4 The GOP structure of MPEG-4 PFGS scalable video................................110 4.5 Illustration of the deterministic dynamic programming approach to derive the optimal mapping algorithm............................................................. 116 4.6 Dependency of the expected distortion of the priority network............... 117 4.7 Interaction between video applications and priority networks for QoS adaptation................................................................................................................120 4.8 The transmission rate constraint of a wireless channel with two service classes, which is computed from the discrete Markov channel model corresponding to the normalized Doppler frequency = 10~2 and aver age power = 16 dB and the buffer sizes are chosen to be 250 and 500 packets...................................................................................................................... 125 4.9 The transmission rate constraint of a wireless channel with twr o service classes and a buffer size of 250 packets based on absolute priority scheduling when the packet loss rate requirement of class 1 is equal to 10”2 and 10~4.................................................................................................... 126 xi Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 4.10 The transmission rate constraint of a wireless channel of service class 1 computed from the discrete Markov channel model with the normal ized Doppler frequencies 10”2 and 5 • 10~3, where the average power is equal to 16 dB ....................................................................................................127 4.11 The transmission rate constraint of the wireless channel for sendee class 1 computed from the discrete Markov channel model with the av erage power equal to 16 dB and 12 dB, where the normalized Doppler frequency frequency is 10~2.................................................................................128 4.12 Traces of the transmission rate constraint with time-varying power and normalized Doppler frequency based on absolute priority schedul ing, where service classes 1 and 2 provide the buffer overflow proba bility guarantees at 10~4 and 10~ 3 , respectively ........................................129 4.13 The Y-PSNR comparison of three video transmission systems under a non-stationary wireless environment with the time-varying power and speed................................................................................................................. 133 5.1 A wireless video system with FEC-based rate adaptation via channel state feedback......................................................................................................... 138 5.2 Multiplexing and packetization of video packets into a group of chan nel packets............................................................................................................... 139 5.3 The plot of RPI value of each GOB packet for the ‘Silent Voice’ se quence and ’Glasgow sequence’.......................................................................... 141 5.4 RPI and the packet size distribution of the ’Glasgow’ sequence. . . . 143 5.5 The general wireless channel model consists of path loss, long term fading(shadowing) and flat fading model (short term fading) ................ 144 5.6 Adaptive filter structure for estimating and predicting channel state in fo rm a tio n ...........................................................................................................147 5.7 Absolute error in percent between long term fading param eter and predicted one by using two concatenated filter..............................................148 5.8 The concatenated code structure used in the proposed communication system....................................................................................................................... 149 5.9 The performance of RCPC with different code ratios in the Rayleigh flat fading channel................................................................................................. 150 xii Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 5.10 The packet error probability with different inner code ratios and total product code ratios with channel SNR equal to 7 dB and N rs equal to 50 bytes............................................................................................................... 151 5.11 Adjustable video packet based Dynamic programming for concate nated code rate allocation................................................................................... 157 5.12 Budget allocation for each window based on the aggregated RPI with bandwidth and packetization constraints........................................................ 158 5.13 The sample trace for selected RCPC (inner) and RS (outer) code ratios with the SNR operation point equal to lOdB and video trans mission rate 41 kbps............................................................................................. 162 5.14 Performance comparison in the averaged PSNR with different opera tional SNR values using a protection solution computed by dynamic programming under equal/unequal error protection and w ith/w ithout RPI with ‘Silent’ video sequence....................................................................... 164 5.15 Performance comparison in the averaged PSNR with different opera tional SNR values using a protection solution computed by dynamic programming under equal/unequal error protection and w ith/w ithout RPI with ‘Glasgow’ video sequence.................................................................. 165 5.16 Performance comparison in the averaged PSNR with the optimal se lection of product code or other choices of the product code....................166 5.17 Comparison of quality degradation in averaged PSNR with 2-channel- slot delay or no-delay in channel state information feedback...................... 168 xiii Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. A bstract The effort to deliver multimedia contents reliably over hostile time-varying channels has gained extensive attention in the last decade. Several novel components in a video transmission system over non-stationary wireless channels are examined in this thesis research, including efficient wireless channel modeling under non-stationary environments and the interaction and joint selection transmission parameters be tween the video source and the transmission module. The thesis work consists of three main contributions. First, a novel adaptive mapping from a set of instantaneous observed channel SNRs in a non-stationary wireless environment to a variable length Markov chain (VLMC) model is proposed. The proposed scheme consists of two main components: estimation of an unknown fading channel signal-to-noise ratio (SNR) distribution and discrete VLMC modeling. First, to obtain the fading SNR distribution, a ker nel density estimation algorithm is applied to the feedback channel SNRs. Then, with the estimated fading SNR distribution and feedback channel SNRs, an itera tive partitioning mechanism and a construction of the context tree are performed to xiv Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. obtain the VLMC model, which yields a much larger and structurally richer class of models than ordinary higher order Markov chains. The frequency of channel model adaptation is dictated by the dynamic of fading statistics. Two applications of this model are presented. The first one is the computation of im portant fading param eters including the fade duration, the level crossing rate, and the channel capacity. The second one is the design of a real-time optimized transmission protocol with an adaptive packet-size. The adaptive packet-size protocol improves the throughput by adaptively changing the packet-size along with varying channel parameters. The accuracy of the proposed VLMC scheme and the performance of its applications are demonstrated via simulation in a micro-cell non-stationary wireless environment. Second, we propose a dynamic QoS management scheme for adaptive priori tized video transmission under time-varying wireless channels. This framework em ploys QoS interaction between video coding and transmission modules, QoS mapping mechanism, video quality adaptation, and channel service rate estimation. The is sues under study include: (1) the rate constraint derivation of transm itting multiple priority classes with QoS (2) the development of a QoS mapping mechanism that maps the statistical QoS guarantees of multiple priority classes to its correspond ing expected video quality optimally, and (3) a QoS interaction protocol to provide trade-off between the video quality requirement and the transmission capability of xv Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. multiple priority classes under time-varying wireless channels. It was shown via sim ulation that the proposed architecture and algorithms offer better video services and enhanced end-to-end video quality. Last, we present a robust video transmission scheme over fading wireless channels that relies on a coordinated protection effort in handling channel and source varia tions dynamically. Given the priority of each source packet based on the Relative Priority Index (RPI) and the estimated channel condition, an adaptive protection scheme via proactive forward error correction (FEC) is developed. This scheme is derived by using joint source-channel criteria. A concatenated code is designed with both the Reed-Solomon (RS) code and Rate-Compatible Punctured Convolutional (RCPC) codes. We use the dynamic programming approach in matching the rel ative priority of source packets to instantaneous channel conditions. To obtain a realistic joint, source-channel adaptation scheme, special attention has been paid to the channel status feedback in terms of accuracy and delay, the product code trade off, and the involved packetization efficiency. The performance improvement due to adaptation via a dynamic programming solution is demonstrated by simulating the wireless transmission of error resilient ITU-T H.263+ video. Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. C hapter 1 Introduction 1.1 Significance of th e R esearch The dream of being able to communicate anywhere at any time with any type of data will soon become reality due to the synergy of telecommunication and infor mation technologies, the growth of the consumer electronics market, and prevalence of content providers. Fig. 1.1 illustrates typical connections between core networks and their air interfaces in the third and fourth generation (3G and 4G) wireless system (i.e. UMTS specification [40]). Backward compatibility from 3G and 4G to the second generation (2G) core technology such as GSM or IS-41 is required. Compared to 2G wireless systems, 3G and 4G wireless systems provide new and important features, especially new extensive bandwidth application and QoS provi sioning. Furthermore it allows negotiation of the transmission properties of appli cations with the radio bearer such as the bandwidth requirement. A ttributes that characterize wireless transmission include the throughput, the transfer delay and 1 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Table 1.1: Different service classes and applications of 3G systems. Traffic class Conversational Streaming Interactive Background Applications voice video games Streaming multimedia Web browsing network game E-mails the data/packet, error rate as well as subjective quality evaluation in the case of multimedia communications. To become commercially successful, the 3G and 4G systems support a wide range of applications that demand different QoS (quality of service) requirements as given in Table 1.1. The 3G wireless communication standard, called IMT-2000, has been developed for transm itting multimedia data. A higher data rate provided by IMT-2000 allows multimedia transmission over wireless channels with a range of bandwidths varying from 64 kbps in a highly moving environment to 2 Mbps in a slowly moving envi ronment [40, 4], Due to wireless channel characteristics, new methods for robust and effective wireless multimedia delivery have been intenswely studied. Among numerous media types, video is the most im portant and commonly used one. In this research, we focus on the development of efficient techniques for reliable video transmission under time-varying environment. As a consequence of Shannon’s separation principle [90], there have been tradi tionally two different approaches to address the wireless video transmission problem. The first one deals with the media source. The codec is designed in such a way that the compressed video stream can recover from errors induced by wireless channels smoothly. Graceful quality degradation to match with changing wireless channel 2 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. IS-41 core network Evolved GSM core network CDMA 2000 Multi-carrier GPRS IP core network Edge Inter-working functions WCDMA TDD Mode TD-CDMA Second Generation Third Generation Figure 1.1: Relationship between core-networks and their air interface alternatives in 3G wireless communication systems. condition is the main objective of this approach. Examples in this category include Multiple Descriptions (MD) [48] and D ata Partitioning (DP) [49] schemes. In the MD case, the encoder breaks a bit stream to several equally im portant descriptions, and sends them through different channels, where a diversity of wireless channels is assumed available. The quality of transm itted media depends on the number of de scriptions received correctly at the decoder. For DP, the compressed video bitstream is reorganized. For example, the header, the motion vector and DCT portions of each MB are grouped with the same type of data in other MBs and coded separately. Sometimes, the reversible variable length code (RVLC) [118] is applied so that the 3 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. decoder can decode a certain portion of corrupted data robustly. The robust source coding approach is suitable for delay-sensitive and/or bandwidth-stringent systems. The second approach lies in the transmission side. It uses additional channel resources for redundant data insertion to make video transmission reliable. In the physical and link layer. Forward Error Correcting Code (FEC) [119], Automatic Re peat Request (ARQ) and hybrid FEC/A RQ [117] are all classified to this approach. FEC introduces redundancy to the bitstream for error correction and detection while ARQ retransm its the portion of corrupted or dropped video bitstream to the decoder. Furthermore, in this approach, the transmission side may provide multiple service guarantees for transmission, where each service is differentiated by guarantees of packet loss/delay. These two approaches have their own shortcomings. Only dealing with the source side, video quality may be not acceptable when the channel is very noisy (e.g. only few packets arrive at the decoder correctly). W ithout considering video source char acteristics, efficient usage of channel resources can be severely hindered. For ex ample, different parts of a video bitstream are not equally im portant and should be treated differently. Hence, to gain both acceptable video quality and efficient resource utilization, coordinated source and channel schemes are desirable. The research carried out so far is centered around the following four issues. 1. Analyze each portion of the underlying video bitstream based on its contribu tion to end-to-end video quality for source prioritization. 4 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 2. Adaptively adjust the usage of wireless channel resources (with focus on chang ing the QoS to video bitstream) based on the changing of channel condition. • 3 . Seek the optimized solution by considering the priority of each video portion and the adaptive level of QoS provisioning to get the highest video quality under an adjustable video codec structure (e.g. scalable video). 4. Determine the proper tradeoff between the resolution of the received video packet,, the packet loss probability and propagation distortion. 1.2 Issues of W ireless V ideo Transm ission 1.2.1 W ireless C h a n n el M o d elin g The knowledge of channel characteristics plays an im portant role in the wireless com munication system design, since the system can allocate resources more effectively to combat the noisy channel effect [128, 129, 57], There are two main approaches to model the wireless channel, i.e. the mathematical approach, which is charac terized by the probability density distribution of received SNR, and the simplified model approach, which is characterized by the discrete number of states and the transitional probabilities among states. It is however a difficult task to character ize a fading channel with closed-form mathematical formulas, since it demands a joint probability density of fading channel characteristics and a complex integration 5 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. step in fading param eter computation [110]. Moreover, the characteristics of wire less channels are time-varying and non-stationary. Thus, a simple, adaptive, yet accurate wireless channel model is more desirable. Among several channel models, the discrete Markov model is the most popular one for wireless channel behavior description. 1.2.2 C h a n n el C od in g Channel Coding is often utilized as a tool applied to the information bearing digi tal signal and aimed at optimizing the performance of digital communications sys tems. These techniques include error control coding, interleaving, coded or uncoded modulation and demodulation, signal detection and equalization, transmission and reception diversity, etc. The performance optimization usually involves trade-off to be made between the power, the bandwidth and the complexity of the signal pro cessing required to guarantee transmission errors of the source data below a certain threshold. The most suitable channel coding and modulation techniques for a given application must also take into account the characteristics of the channel through which transmission is to going to take place. • Error Control Coding ARQ, FEC and hybrid ARQ/FEC [60] are among the most popular meth ods for reliable video transmission. However, different types of service impose different constraints on their applicability. In a delay-stringent, system such 6 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. as real-time video communication, ARQ is often not allowed so that FEC should play the dominant role. For some families of FEC codes which allow a dynamic change of coding strengths conveniently (e.g. BCH [11] and RCPC [12]), unequal error protection can be applied to bits of unequal importance to represent the multimedia content. However, in applications such as broadcast, multicast or point-to-point without feedback, the transm itter cannot change the FEC code ratio dynamically [79]. Embedded codes [61] provide an attrac tive alternative solution. A highly complex structure with a large amount of computations based on iterative decoding principle such as turbo code may also be emploj^ed [6]. In contrast, in a delay-tolerable system, ARQ appears to be a favorite choice since channel diversity can be obtained in the ARQ process. • Channel Modulation Unequal error protection based on channel modulation has been considered for European digital audio and video broadcast [79, 91]. Each unequally impor tant layer is mapped to different bit error rate requirements and achieved by different modulation schemes. The integration of channel coding and modula tion with unequal error protection can be achieved by using the trellis coded modulation (TCM) concept. In a system whose feedback channel status is available, adaptive modulation [5] can be applied. Together with adaptive modulation, many proposed adaptation techniques can be used. In the 3G 7 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. system, an adaptive CDMA system can adjust the spreading gain and perform power control [40]. The 2G TDMA system may use adaptive modulation and power control to assure received data quality [5]. All the above adaptation techniques have the same design objective, i.e. protecting video data more when the channel is in a bad state. 1.2.3 Source C od in g Source Coding is a process of removing redundant information from the raw source symbol stream. There are two major types of source coding: lossless and Lossy. Lossless source coding can identically reconstruct the source information symbols. However, due to the bandwidth shortage problem both in transmission and storage, trade-off between quality and bit rates should be considered. By exploiting human perception, lossy source coding permits a certain amount of errors or quality loss in the source coding procedure. In multimedia compression standards such as H.263+, JPEG, G.723 or MPEG-4, lossy coding is actually preferred over lossless coding because of its strong compression ability. Following sections will describe video source coding techniques. • Robust Entropy Coding In most existing compression standards, the variable length code (VLC) is em ployed to remove statistical redundancy. However, it may lead to disastrous decoded results due to the loss of synchronization and error propagation. To 8 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. resynchronize the decoding process after the occurrence of an error, synchro nization words are inserted in the coded bitstream. In the H.2634- video com pression standard, a synchronization word is put in the header of every GOB. Furthermore, reversible variable length codes (RVLC) [118] and error resilient entropy codes (EREC) [84] have been proposed to partially recover corrupted data and to reduce the overhead of synchronization words for graceful quality degradation. • Error Detection and Concealment To maintain robust video quality, error detection and concealment is needed to reduce the error propagation effect. Once an error is detected, both spatial and temporal concealment method can be utilized to reduce its impact. In temporal concealment methods, a motion vector is estim ated from adjacent MB to recover the loss effect. In spatial concealment, data in a corrupted MB are recovered based on pixels of its adjacent region in the same video frame [126, 80]. • Data Partitioning In association with unequal error protection, data partitioning provides an attractive option to generate a robust bit stream in today’ s video coding stan dards such as H.263+ and MPEG4. Motion vectors and DCT coefficients of several MBs are coded and put together with some synchronization word as the separator [49]. Thus, different types of data can be protected differently due 9 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. to their roles. RVLC can also be used to encode partitioned data for partially corrupted data recovery. • Multiple Description By using the multiple description (MD) scheme [48], we encode a video into two or more descriptions at the encoder. Each description is sent through a different channel to achieve channel diversity. Even there is only one descrip tion received correctly, the decoder is still able to decode received data with degraded visual quality. The quality of received video depends on the number and the quality of received descriptions at the decoder. 1.2.4 J o in t S ou rce C h an n el C od in g Shannon [90] identified the major components of signal processing for communi cations and storage as source coding, channel coding and secrecy. Furthermore, he proved that, in the ideal case, they could be solved independently and then combined for the optimal performance. Researchers sometimes forget th at this independence result holds only for restricted models of the source and the channel and at the price of infinite delay and infinite computing power. In practical situations, joint source-channel coding can yield a significantly better performance. Improvements have been demonstrated in modulation, but remain to be seen in the combination of compression and secrecy. 10 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Reliably delivering real-time multimedia data to a mix of users over possibly time varying wireless channels requires innovative solutions. Worst case scenarios must be improved if efficient multimedia communication is to fulfill its promise. A possible solution is to use scalable; adjustable or multi-resolution source coding methods combined with adaptive transmission schemes so that different degrees of services and protection (e.g. priorities) can be provided. Such approaches have not yet been studied and optimized extensively. A substantial amount of work on the joint design of real-time audio/video coders for packet services is still needed in order to achieve good quality and efficient use of networking resources. Of particular interest is the coding scenario for image and video transmission with a multi-resolution data representation property [86]. These schemes are useful in image and video browsing and web-based applications, and expected to be crucial for multi-media wireless network systems. Study of optimal strategies in the presence of feedback is im portant in leading to computationally feasible algorithms [77]. 1.2.5 Q u ality o f S ervice To offer differentiated QoS to different applications is one of the most im portant issues in 3G and 4G wireless systems. However, for the high-speed high-quality fixed communication network as the Internet, it is already difficult to guarante QoS in terms of end-to-end perceived quality and delay. Then, for the wireless commu nication network with low-capacity error-prone time-varying links, the mobility of 11 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. users, and the dynamic topology especially in wireless ad-hoc networks, QoS guaran tee is even more challenging. Generally speaking, a QoS provisioning system should have an adaptive capability to cope with changing network conditions. Furthermore, applications should have QoS-adaptive property with adjustable QoS parameters. 1.2.6 C ross-L ayer D e sig n To efficiently delivery video bitstreams with a limited amount of resources in wireless channels, the protocol stack must evolve as shown in Fig. 1.2. This figure indicates that the information must be exchanged across all layers in the protocol stack. The information exchange allows the protocol to be adaptive to the application require ments and underlying network conditions in a global manner. Moreover, all protocol layers must be jointly optimized with respect to constraints and characteristics to meet the requirements of both the application and the transmission capability of lower protocol layers. 1.3 C ontribution of th e R esearch Contributions of this research are detailed below. • We have proposed a novel adaptive fading channel model based on the variable length Markov chain (VLMC). The proposed VLMC channel model adaptively changes the structure and the number of channel param eters along with fading 12 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Operation Physical Layer Application Layer Networking Layer Link Layer Cross-layer adaptation System Constraint Figure 1.2: Adaptive cross-layer design and operation. statistics. Compared with the fixed order Markov chain, VLMC requires fewer parameters and thus a lower complexity in channel modeling. • We have derived the transmission rate constraint of multiple priority classes given the corresponding statistical QoS guarantees of multiple priority classes. The transmission rate constraint can be used as tools in allocating the trans mission resource to each priority classes for video transmission. • We have proposed a QoS mapping mechanism from the link layer to the ap plication layer under a time-varying environment. The proposed scheme is achieved based on the knowledge of the transmission rate constraint and the distortion property of the video sequence. The dynamic programming algo rithm is utilized to obtain the optimal solution of QoS mapping. 13 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. • We have developed an adaptive QoS provisioning framework, which employs the interaction between the application QoS requirements and the link layer capability to select the optimal QoS parameters for video transmission. • We have investigated a new solution in integrating the relative priority index (RPI) with unequal error protection. We assign each video packet an im portance index with RPI. This is a fine grain prioritization approach to the video source in comparison with the traditional class-based or session-based approach. • We have developed a feedback-based adaptive concatenated FEC scheme. The new scheme is compatible with the multimedia communication standard H.324 as well as H.223 Annex C (RCPC) and Annex D (Shorten RS code). In this scheme, the decoder has the responsibility to estimate and predict the channel status by using a concatenated adaptive filter algorithm. Given the channel state information and the video packet size, the proposed algorithm is able to select the best combination of concatenated code parameters. The degradation of end-to-end video quality due to feedback delay has been examined. This algorithm finds applications in dealing with a time-varying environment with a self-configuration system. • We have proposed an optimization solution that takes into account the impor tance of each video packet and the channel state information to determine the 14 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. proper protection level to minimize the error propagation effect under the bit budget and the packetization constraints. 1.4 O utline o f th e dissertation In this introductory chapter, we have described the various issues associated with multimedia transmission over time-varying and non-stationary wireless channel. The remainder of this dissertation is organized as follows. In Chapter 2, we briefly review some background of wireless video communica tion systems and standards and related previous work such as optimization methods for robust video communication, channel and source adaptation techniques and an alytical models for error propagation in video transmission. In chapter 3, a novel adaptive mapping from a set of instantaneous observed channel SNRs in a non- stationary wireless environment to a variable length Markov chain (VLMC) model is proposed. Chapter 4 is devoted to dynamic QoS management for adaptive pri oritized video transmission under time-varying wireless channels. This framework employs QoS interaction between video coding and transmission modules such as QoS mapping, video quality adaptation, and channel service rate estimation. In Chapter 5. we present a robust video transmission scheme over Rayleigh fading wireless channels that relies on a coordinated protection effort in handling channel and source variations dynamically. Finally, concluding remarks and future work are stated in Chapter 6. 15 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. C hapter 2 W ireless V id eo Transm ission: B ackground R eview 2.1 In trodu ction To transm it a compressed video bitstream robustly in wireless channels, joint con sideration of channel and video characteristics is one of the most hot research topics. Research work along this direction can be categorized into the following classes: • channel coding design to match source characteristics; • rate allocation between source and channel codes; • power allocation to transm itted symbols based on source and channel consid eration; • adaptive source coding based on informed channel state information; • source decoding based on residual source redundancy; • adaptive video playback deadline based on channel characteristics; 16 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. • QoS provisioning for transm itting video bitstream under time-varying channel characteristics: • optimized rate distortion scheduling for video packet transmission. The rest of this chapter is organized as follows. The wireless video communication system in described in Section 2.2. The existing adaptation algorithms in wireless communication channels are stated in Section 2.3. The optimization techniques used in image/video codec design are examined in Section 2.4. Finally, previous work on robust wireless image/video communication are reviewed in Section 2.5. 2.2 W ireless V ideo C om m unication System A wireless communication system can be divided into 4 major components: 1. The video encoder, which compresses video signals and uploads the video bit stream to the media server. The standardized video encoders include the H.26x family and the MPEG family. 2. The media server, which stores the compressed video bitstream and transm its them on demand. 3. The transport mechanism, which delivers video packets from server to the client for the best received video quality. 17 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 4. The video decoder, which decompresses and recovers some lost information during transmission. An example of the video communication system as standardized by ITU is de scribed in the following section. 2.2.1 V id eo T ran sm ission S ystem : IT U -H .3 2 4 The ITU-T Recommendation H.324 [55] is a low bit rate multimedia communi cation standard and serves as a good foundation in the design of modern wireless multimedia communication system. H.324 consists of the audio codec module, the video codec module and several system protocols. Typical applications include video conferencing or video telephony. The structure of H.324 is shown in Figure 2.1. The mandatory elements of H.324 include the modem (V.34), the multiplexer (H.223), and the control protocol (H.245) while media portions consisting of data (V.14,LAPM), audio (G.723) and video (H.261/H.263) are optional. Thus, H.324 has the flexibility of data communication between the most fundamental terminals or the most advanced terminals. It also supports a wide range of multimedia products and services. In this research, we are primarily focused on wireless video transmission. Thus, our scope is restricted to H.263+ and H.223. Some basic knowledge of H.263+ and H.223 will be reviewed in the following subsections. 18 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Multiplex/ Demultiplex H.223 Network Modem control V.25ter V.RLAPM Audio I/O Video I/O Receive Audio codec G723 User data Applications T.120 etc Control Protocol H.245 procedure Control Figure 2.1: The block diagram of an H.324 multimedia terminal. 2.2.1.1 IT U -H .2 6 3 + video standard The structure of the H.263+ [19] video encoder is depicted in Figure 2.2. It exploits motion predicted compensation to reduce temporal redundancy. It also uses the discrete cosine transform (DCT), quantization and entropy coding to reduce spatial redundancy. H.263+ supports five standardized picture formats, namely, sub-QCIF, QCIF, GIF, 4CIF and 16CIF. For the QCIF resolution, each picture is of size 176 x 144. It is composed by 9 Group Of Blocks (GOB). Each GOB have 11 Macro Blocks 19 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. (MB) where each MB contains a 16*16 pixel array for the luminance component and two 8*8 pixel arrays for the U and V chrominance elements respectively. Video Output Entropy Coding I Intra | Inter Video Multiplex Frame Memory DCT Inverse DCT Inverse Motion Compensation Coding Control Figure 2.2: The structure of an H.263 video encoder. There are two basic modes for encoding MB. i.e. the Intra and the Inter Modes. When coding with the Inter mode, the current MB is predicted from a shifted 16*16 block with the motion vector in the previously reconstructed video frame. The prediction error between the current and the predicted MB information is encoded by using DCT followed by quantization, run-length entropy coding, respectively. For the Intra mode, the encoder does not use temporal prediction in the encoding process. Instead, the picture is encoded directly as a still image with DCT, quantization and entropy coding. The popular video quality measurement is the Peak Signal to Noise 20 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Ratio (PSNR). PSNR is defined as 101og(2552/MS'.E) with the unit in dB, where MSE denotes the mean squared error between the original input video signal and the reconstructed video at the decoder. Normally, the overall performance is measured by averaged PSNR. It is worthwhile to point out th at PSNR does not correlate very well with perceived visual quality. However, 0.5-1 dB difference in the PSNR value normally results in visible difference of video in comparison. 2.2.1.2 IT U -H .223 m ultiplexing standard H.223 [55] is a multiplexer standard used to facilitate the combination of various multimedia sources such as control, data, audio and video information into a single packet. The structure of H.223 is shown in Figure 2.3. Conceptually, the multiplexer consists of two layers, i.e. the adaptation and the multiplex layers. In the adaptation layer, error detection and correction based on CRC (Cyclic Redundant Checking), FEC and the sequence numbering information is performed on the source bit stream. The multiplex layer combines inputs from different sources and form packets of a variable or fixed packet size for transmission. As illustrated in Figure 2.3, the input to the adaptation layer is called AL-SDUs (Adaptation Layer Service D ata Units) while the output from the adaptation layer is called AL-PDUs (Adaptation Layer Protocol D ata Units). The input and the output of the multiplexing layer are called MUX-SDUs and MUX-PDUs, respectively. To operate in an environment of a higher bit error rate such as the mobile channel, H.223 Annex B (CRC), Annex C (Rate Compatible Puncture Convolutional code) 21 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. AL-SDU A L-PDU M U X-SDU Audio 170 Video I/O Audio Coder LAPM H.245 Control D ata Protocol Video Coder Physical Layer A pplication Layer Adaptation Layer M ultiplex Layer Figure 2.3: Components of the H.223 standard. and Annex D (Shorten RS code) are utilized for error detection and correction in the adaptation layer of the underlying multimedia bitstream. In this work, we focus on the adaptation layer for video transmission by exploiting H.223 Annex B, Annex C and Annex D. 22 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 2.3 A d ap tation Techniques in W ireless C om m unication S ystem s Wireless channels have a higher bit error rate than wired one. The bit error rate for a typical multimedia mobile application is between 10~2 and 1Q~3 while that for a wired network application is in the order of 10~6. Due to the time-varying nature of wireless channels, timely estimation of the channel status and effective utilization of the information for data protection is critical to successful transmission of multimedia data. Spectral efficiency can be achieved via the change of coding and spreading schemes and/or increasing the con stellation density (i.e. the modulation scheme). Furthermore, adaptation techniques have been adopted in wireless communication standards such as CDMA (IS-95), TDMA(IS-136), GSM and wideband CDMA (cdma2000 and UMTS WCDMA) [40], They are briefly reviewed below. Rate adaptation in 2G CDMA systems is achieved via a combination of variable spreading, coding and code aggregation. The IS-95B system utilizes Walsh code aggregation for higher rates. In 3G CDMA systems such as WCDMA and cdma2000, higher rates are achieved via a combination of variable length of spreading code and different rates of coding. To match the channel state information in a CDMA system, pilot strength measurements are used to estimate the channel state information. For IS95B and cdma2000, the estimated channel state information is received at 23 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. the base station through the pilot strength measurement message (PSMM) or the supplemental channel request message (SCRM). Quantitative measures such as the block error rate, bit error rate (BER), received power, path loss can be used as channel quality metrics in CDMA. In 3G wireless systems, a self-configuration mobile terminal provide a good choice in dealing with the noisy characteristics of wireless channels. A mobile terminal progressively learns the channel status and adapts itself to the available channel resource as efficiently as possible. In TDMA systems [40], slot-by-slot adaptation is performed with adaptive coding and modulation [40], while the transm itted symbol rate remains the same. Time slot aggregation and incremental redundancy transmission are used to achieve a high throughput in GPRS-136 (General Packet Radio Service) and Enhanced GPRS (EGPRS). The estimated channel state information is not required in incremental redundancy transmission. Instead, incremental amounts of redundancy are trans m itted until the receiver is able to successfully decode the data frame. In addition, transmission of redundant information can get the advantage in time diversity dur ing decoding. However, this may not be suitable for d e^ -strin g e n t applications. Channel quality metrics in TDMA systems can possibly be the frame error rate, the bit error rate (BER), the long term fading parameter or path loss. They are estimated at the receiver and sent back to the base station through properly defined messages. In this research, we focus on burst by burst adaptation based on the TDMA system. 24 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. It is worthwhile to point out th at most techniques described above perform rate adaptation in a slow manner. For example, rate adaptation is usually performed in response to slow fading variation or path loss as a result of feedback delay and lacking of an adaptation mechanism on the symbol-by-symbol basis. 2.3.1 R eview of P rev io u s W ork Some researchers studied adaptive modulation techniques with an objective to achieve the maximum throughput or to maintain low bit error probability. Goldsmith [33] utilized the estimated short-term fading parameters to adjust the modulation type of each symbol without considering feedback delay. Region-based modulation selec tion was used to achieve a target bit error probability. However, the applicability of these adaptive algorithms have been questioned due to the time difference between the channel status estimation and the corresponding source adjustment. Goeckel [34] proposed an adaptive modulation system based on the estimated outdated channel information. The system was designed by assuming a static channel model, which would be valid only if the delay between channel estimation and data transmission is very small. As mention in Chapter 1, the symbol-by-symbol adaptation may not be feasible in the real world application so that the burst-by-burst adaptation concept was adopted in [5] and [41] for adaptive modulation, where time-scale adaptation 25 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. was carried out in channel slots or frames in CDMA and TDMA systems. Region- based adaptation th a t maintains low bit error probability was still in use in their modulation selection scheme. When a wireless system have more available bandwidth, the forward error cor recting code (FEC) provides an efficient way to maintain the visual quality of trans mitted video. Like adaptive modulation, the level of FEC protection can be adaptive due to the flexibility of FEC schemes such as the RCPC (Rate Compatible Puncture Convolution) code [42] or the shorten RS (Reed-Solomon) code. Other techniques using the power, the spreading code, etc. are described in [63]. In multimedia sendees, the quality of received video is more of our concern than the channel quality measurement such as BER. The characteristics of a multimedia source should be brought into consideration in system design. In [41], an adaptive system that applies different FEC classes to different sensitivity portions of video data with burst-by-burst adaptive modulation is shown in Fig. 2.4 for video com munication. The H.263 [19] video codec was adopted by this system. The video bitstream was classified into several classes and protected by BCH [97] or turbo codes [6]. There were two main classes in this system. The first class was protected by BCH(127,50,13) while the second class was protected by BCH(127,92,5). The header was also strongly protected by BCH (127,50,13). The symbol rate was kept constant in this scheme. 26 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Video Packet A N-class k — ► Encoder ► Assembly ► Mapper \ FEC encoder ►Bits/symbol &TDMA Video Packet Mapper N-class N Mapper QAM Decoder Disassmebly « \ FEC decoder / & TDMA DEM 4 Modulator Figure 2.4: The refigurable video transceiver system. To prevent error propagation in a corrupted video packet, a flexible adaptive packetizaton was described in [18]. In [15, 18], a system was proposed to drop corrupted video packets, rather than allowing the contaminated bit stream to effect the future video frame. A key feature in this adaptive packetization algorithm is the provision of a strongly protected binary packet acknowledgment flag, which instructs the remote decoder not to update the local and remote video reconstruction buffers in the event of a corrupted video packet error. The modulation type is selected based 27 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. on pseudo-SNR, which is the ratio of the desired signal power with the residual ISI power and the effective noise power. 2.4 O ptim ization Techniques W ith the resource constraint in a wireless system and characteristics of multimedia, an optimization technique seek the best way in choosing system parameters. Based on the traditional rate-distortion framework, analysis can be performed under the stationary source model, such as the Gaussian or the Laplacian model, to provide a performance bound. However, when dealing with a real-world system, the distri bution of the source does not match well the simple m athematic model to give a tractable analytic solution. Furthermore, discrete optimization has to be carried out due to the nature of the observed data set. For example, quantization levels in video coding take a set of discrete values. The cost function adopted in the optimization setting is often the mean square error (MSE), which is defined as m s e = J 2 ( U j ) - x t(j))2. where X t(j) is the reconstructed value of X t{j).In this work, we focus on MSE- based optimization with consideration of the coding process from the video encoder and the transmission process due to the channel effect. Existing tools in solving a 28 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. constrained optimization problem include the Lagrangian method and the dynamic programming search. They are briefly reviewed below. 2.4.1 L agrangian o p tim iza tio n approach The Lagrangian method can be written as aT(A) = arg min(D(xk) + A R(xk), (2-1) where x*(X) is the optimal solution with a given A , D(xic) denotes a distortion measure and R{xk) corresponds to the resource in terms of bit rates required by solution Xk- The search of the Lagrangian optimized solution can be interpreted geometrically as shown in Fig. 2.5. The optimal solution is the operating point which is first touched by the plane of an absolute slope A . By varying the A value, an optimal solution can be determined to meet the bit budget constraint. It is worthwhile to point out that the Lagrangian search can provide the optimal solution if the R-D curve is convex. When the R-D curve is not strictly convex, the obtained solution could be suboptimal. Also, the R-D curve has to be calculated at first before the optimization process begins. 2 .4 .2 D y n a m ic p rogram m in g app roach Another approach to solve the constrained optimization problem for the optimal solution is dynamic programming. The structure of dynamic programming is given 29 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. D (xj) Plane wave o f slope X Convex distortion curve R(X i) Figure 2.5: Illustration of the Lagrangian optimization principle. in Fig. 2.6. Simply speaking, dynamic programming finds the optimal solution by tracing all possible paths of the form of trees or trellis and selecting the optimal path that gives the minimum cost function. Each branch of a tree or a trellis has an associated cost while each node has the information of the demanded resource in accumulation in term s of bits. The node that demands a resource exceeding the bit budget will be pruned out to result in a trellis search. Compared with the Lagrangian optimization, dynamic programming can still provide the optimal solution even when the R-D curve is not convex. However, the complexity of dynamic programming is higher than that of the Lagrangian approach. 2.4.3 R e v iew o f P rev io u s W ork The optimization framework has been widely used in resource allocation to achieve the maximum quality of transm itted media [77]. In [99], Gersho and Shoham used 30 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Exceeded bit budget Node Path transition from Node to Node Aggregate distortion Used Resource Optimal path Figure 2.6: The structure of dynamic programming. the Lagrangian multiplier to search for the level of quantization for each block of an image for maximum quality. For rate control. Ortega et al. [43] utilized both the Lagrangian multiplier and the dynamic programming approaches to find the optimal solution for video transmission with delay and buffer constraints. Optimization solutions mentioned above was developed based on the assumption of independent problems (i.e. previously selected optimization parameters do not effect the current selection). For the dependent problem, the independent assumption is not valid any long due to frame dependency [87, 54] or video packetization in time varying channels [54]. Trellis-based dependency problem was solved in [87] for video coding. To reduce the complexity in the search of the optimal solution, video characteristics such as the monotonic principle were employed in [87]. 31 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 2.5 R eview o f R obust W ireless V isual C om m unications We review previous work on robust visual communications in this section. 2.5.1 M u lti-r eso lu tio n B a se d V id e o T ran sm ission The wavelet image coding algorithm encodes a still image by using the wavelet transform. The modern wavelet-based image coding algorithm was first proposed by Shapiro [92] and called the embedded zero tree wavelet (EZW). Further improvement was done by Said and Pearlman [93]. An advantage of the wavelet-based coder is the flexibility in adjusting the source rate. To transm it an image over wireless channels, most existing solutions exploit unequal protection which takes into account the sensitivity of each portion of the compressed bitstream. However, the assignment of a priority level to a portion is heuristic. Man [67] classified a bitstream into 3 classes, i.e. the node test sequence, the descendent test sequence and the refinement subsequence. Each class was pro tected by a different code ratio based on its importance. Also, Sherwood and Zeger [94] simply transm itted the EZW bitstream by using RCPC (Rate Compatible Punc ture Covolutional) code with the CRC parity check. Their work was also extended to the product code consisting of RS (Reed-Solomon code) and RCPC in [95] to deal 32 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. with both error and erasure channels, where two dimensional decoding (in the hori zontal direction by RCPC and the vertical direction by the RS code) were performed. Results obtained in [95] showed a better performance than their previous work. Far- vardin [13] used RCPC in progressive transmission. Image data are first protected with a light level of protection. The status of the communication channel is then fed back to the transm itter. When the information packet is corrupted and a negative acknowledgement is received by the transm itter, it will increase the protection level and transm it more parity bits to the decoder. The procedure will be performed iter atively until the delay time exceeds the constraint or the protection level reaches the maximum level. Then, some packet data will be dropped. Obviously, this scheme is not suitable for a delay-constraint system. Finally, multiple description provides another robust approach for image transmission. Ramchandran [96] proposed the use of concatenated codes with multiple description in transm itting an image over wireless channels. In the context of video coding, Cheung [14] applied 3D subband coding together with RCPC for video transmission over a noisy channel. Allocation of code rates of RCPC in each bit plane layer in video is optimized with the feedback of channel state information. 33 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 2.5.2 D C T -B a se d V id eo T ran sm ission DCT coding is widely used in image and video compression standards such as JPEG, MPEG and H.263+. Error resilient DCT coding schemes were proposed by re searchers. Modestino [68] proposed an unequal error protection scheme applied to the indices of transform coefficients by assigning different channel code ratios. On the other hand, Farvardin [115] designed channel optimized scalar quantizers to quantize DCT coefficients. Furthermore, TCQ [69] was applied to DCT coefficients for robust image encoding and transmission. Another approach is to use the idea of iterative decoding such as Turbo coding to receive better quality received image as described in [81]. However, as pointed out by [68], iterative decoding is not suitable for a delay-stringent system. For a DCT-based video transmission system, Girod [31] proposed a tool to model the error propagation characteristics of video. The channel state information can be fed back to the transm itter to decide the intra MB refresh rate. Cherriman and Hanzo [15, 16] examined various issues in both power control and OFDM multiplex ing for transm itted video. In [17], the MPEG2 video bitstream was protected by FEC. The bit stream was separated into 2 classes, i.e. the header and the payload. The header was protected with a strong FEC code while the payload was protected based on the priority and the statistical behavior of payload contents. 34 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 2 .5.3 A p p lica tio n A d a p ta tio n and C ross-L ayer D e sig n Due to the time-varying nature of wireless networks, video applications adaptively select QoS parameters offered by the transmission systems. For example, the trans mission system may offer an application a rate-delay trade-off curve derived from the capability of lower-layer protocols [125]. The application layer must decide a proper operating point on this curve (say, some applications may be able to tolerate higher delay while some may not). Furthermore, the energy constraint results in an other trade-off between the network performance and longevity. Generally speaking, the trade-off is a multi-dimensional function, including rates, delay, BER (bit error rates), longevity and so on. These trade-off curves will vary as network conditions change from time to time. The work by Alwan et al. [3], Mirhakak et al. [73], and Ram anathan et al. [88] considered the application adaptation for wireless networks. As shown in these papers, if the wireless application has a good adaptive capability, it is possible to achieve an overall good video transmission performance despite poor network conditions. Moreover, at the decoder end, the video application can adaptively adjust the speed of playback based on channel conditions. The objective is to reduce the delay introduced by the client buffer and to provide rate scalability within a small range. For example, the application of adaptive playout for voice was studied by Liang et al. [64] while adaptive playout video was investigated by Steinbach [100]. 35 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. C hapter 3 N on-station ary Fading C hannel M odeling w ith A daptive V ariable L ength M arkov C hains (VLMC) 3.1 Introduction The knowledge of fading channel characteristics plays an im portant role in wireless communication system design, since the system can allocate resources more effec tively to combat the noisy channel effect [128, 129, 57]. It is however a difficult task to characterize a fading channel environment with a closed form of a mathematical characterization, because this process requires a joint probability density of fading channel characteristics and a complex integration step in fading param eter compu tation [110]. To simplify the analysis and design of communication systems, the approximation of fading characteristics by a discrete Markov model has been widely 36 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. adopted due to its well developed theory [44] and simplicity in computing channel parameters (e.g., fade duration or throughput estim ation)[112]. Gilbert and Elliot [37, 29] first proposed the use of a finite-state Markov chain (FSMC) model to approximate a lossy channel with memory. Several extensions have been done by researchers for the modeling of fading channels with a Markov model, e.g. [112, 32. 120, 121, 111, 7, 101, 102, 58]. The mapping from a fading channel to the first-order discrete Markov model was proposed by Wang and Moajaeri [120] with the Rayleigh fading statistics. This first-order Markov assumption for fading channel modeling was verified by Wang and Chang [121] and Tan and Beaulieu [111], Sajadieh et al. [101] used a continuous-time Markov double-chain to model the Rayleigh fading channel and to compute the Markov state dwell time. The contingency table was adopted by Babich and Lombardi [7] to determine the order of the Markov chain and construct the Markov model for narrow-band Rayleigh and Rician fading modeling. The modeling of a burst error channel using the hidden Markov model (HMM) was proposed by Sivaprakasam and Shanmugan [102], where the HMM parameters were estimated by the Baum-Welch algorithm. The HMM was also adopted by Turin and van Nobelen [112] for fading channel modeling with measured data and for distribution analysis of fading param eters (e.g., the fade duration). The use of the Markov model to study the non-stationary error trace in the GSM system was performed by Konrad et al. [58]. The application of the Markov model to protocol design and analysis was studied in [111, 128, 129, 113], 37 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. including the protection level design of forward error correction (FEC) codes and the throughput efficiency analysis of ARQ (Automatic Repeated reQuest). However, a number of challenging issues remains to be addressed in order to make a discrete Markov model practically useful in representing a wireless fading channel. They can be stated as follows. • The exact fading envelope and the fading channel SNR distribution are un known in practice. Hence, the fading channel model based on a discrete Markov chain with a known channel SNR distribution (e.g. the Rayleigh, Rician, or Nakagami distribution [104]) may not match the actual wireless fading envi ronment well. • The fading envelope and the fading channel SNR distribution are statistically non-stationary [20] (i.e. the statistical distribution of the fading envelope or the fading channel SNR is time-varying) due to the change of mobile’ s speed [21] and environment [70]. Hence, the common assumption th at the fading channel is statistically stationary during communication periods, which has been made made in most previous work on discrete Markov models, may not be realistic. • The number of parameters characterizing the discrete Markov model (e.g. the order of discrete Markov model and the Markov state transitional probability) should be as small as possible to reduce the complexity in channel modeling but should be large enough to efficiently approximate the fading channel. 38 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. • The fading channel modeling algorithm based on the discrete Markov model should possess a real-time adaptive capability to capture the change of fading channel statistics, which is useful in applications such as real-time optimal protocol design for mobile communication systems [128, 129, 57]. To address these issues, an adaptive Markov method is proposed in this research to model a non-stationary wireless channel. The objective is to achieve real-time discrete Markov channel modeling, which represents the fading channel character istics from time to time based on the observed channel state information. In par ticular, a dynamic mapping scheme is developed for converting the instantaneous observed channel state information of time-varying fading statistics to a discrete variable length Markov chain (VLMC) model [9], which possesses a much larger and structurally richer class of models than ordinary higher order Markov chains. The proposed dynamic mapping scheme consists of two main components: (i) fading channel SNR distribution estimation and (ii) discrete Markov modeling. The fading SNR distribution estimation is explicitly used to capture local changes of unknown fading statistics based on feedback estimated fading channel SNRs. Then, an iterative partitioning mechanism and a context tree construction procedure are applied to the estimated fading SNR distribution and feedback channel SNRs to derive the VLMC model for approximating the wireless fading channel. Parameters characterizing VLMC model (e.g., order and transitional probability) is computed through the constructed context tree and feedback estimated fading SNRs. The 39 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. frequency of adaptation (i.e., the number of channel SNR samples used for channel model update) is determined based on the dynamic of fading channel statistics. Two applications of this VLMC model are also presented. The first one is the computation of fading parameters, including the fade duration, the level crossing rate, and the channel capacity. The second one is the design of a real-time trans mission protocol with an adaptive packet-size. The adaptive packet-size protocol improves the throughput by adaptively changing the packet-size along with the varying channel parameters estimated with the derived VLMC. The accuracy of the proposed adaptive channel modeling scheme and the performance of its applications are demonstrated via simulation in a micro-cell non-stationary wireless environment. The rest of this chapter is organized as follows. The overview of the proposed adaptive channel modeling is described in detail in Section 3.2. The dynamic map ping from the non-stationary observed channel state information to a discrete VLMC model, including the estimation of fading channel SNR distribution, fading parti tioning, and VLMC modeling, is discussed in Section 3.3. The channel model update scheme under time-varying fading channel statistics is discussed in Section 3.4. The application of the derived adaptive VLMC model to the fading param eter estimation and the adaptive transmission protocol design is detailed in Section 3.5. Experimen tal results are shown in Section 3.6. Finally, concluding remarks are given in Section 3.7. 40 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 3.2 O verview of th e P roposed A d ap tive C hannel M odeling In this section, we first describe the non-stationary characterization of a wireless fading channel. Then, the overview of the proposed adaptive channel modeling under non-stationary fading environments will be presented. Due to the time-varying speed of the mobile terminal and various environmental factors, it is inevitable for a wireless channel to have the non-stationary property. For example, during the communication process, the line-of-sight transmission path may exist for some period of time, and may disappear when the mobile terminal is obstructed by buildings (Fig.3.1). This implies the time-varying nature of the fading envelope distribution (i.e., from the Rician distribution to the Rayleigh dis tribution) . The speed of a mobile terminal may often change, which results in time- varying correlation between fades [21](i.e., the time-varying normalized Doppler fre quency). Furthermore, surrounding building patterns of a mobile terminal and its time-varying distance to the base-station also affect channel characteristics. The variability of long-term fading parameters was investigated by measuring the aver aged transm itted SNR. patterns in down-town Chicago in [70]. It was reported that the standard deviation of the log-normal long-term fading varied in the range of 6.5-10.5 dB for a large area and 4-7 dB for a small area. The time-varying distance from a mobile term inal to a base-station also introduced the non-stationary effect 41 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Base-station Building pattern 2 Figure 3.1: The non-stationary nature of a fading channel. to a wireless channel model due to the time-varying power fall-off of the path loss model. To model the wireless fading characteristics under a non-stationary environment, the adaptive channel modeling based on a discrete VLMC model as shown in Fig.3.2 is considered. First, the unknown statistics of the fading channel at time of interest will be tracked and estimated from an instantaneous set of observed fading SNRs (see the discussion in Section 3.3.1). It is assumed th at the fading SNR distribution is efficiently estimated at the receiver [30, 103] and feedback reliably to the transm itter. The kernel density estimation algorithm will be utilized to obtain the distribution of fading characteristics due to its smoothness and accuracy property comparing to other methods [122]. The number of estimated fading SNRs, L(t), used in estimating fading statistics is an adaptive param eter depending on how fast fading channel statistics change. If the fading channel statistics vary slowly, the size of L(t) tends to be large since there is enough time to collect channel state information before 42 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. fading channel statistics change. In contrast, if the fading channel statistics are highly dynamic, the size of L(t) cannot be large to be able to track the change of fading statistics. Then, with the estimated fading channel statistics, real-time fading channel modeling based on discrete VLMC will be derived. The VLMC is a potentially higher order Markov chain with a natural parsimonious structure of the Markov state transitional probability. The VLMC derivation procedure consists of two main steps: (i) fading partition ing (Section 3.3.2 .1) and (ii) discrete VLMC construction (Section 3.3.2.2). First, the whole range of the fading channel SNR associated with the estimated fading SNR distribution will be partitioned to several intervals. Given the transmission policy in terms of the modulation type, we will set up the optimization framework to obtain the optimal partitioning set of the fading channel SNR. The objective of the optimization framework is to maximize the closeness of the corresponding fad ing channel features (e.g. the probability of bit error) of the discrete VLMC model and those of the current fading channel statistics. The optimal fading partitioning solution is obtained through an iterative algorithm. The VLMC representing the fading channel is then constructed based on the optimal fading partitioning set and feedback fading channel SNRs. To be more specific, each feedback fading SNR sample is first quantized to its corresponding partitioned SNR interval depended on its SNR value. The quantized fading SNR sample is represented by the letter corresponding to each partitioned 43 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Speed o f A daptation L(t) D elay VLMC Channel Modeling Estimated Channel SNR Estimation of Fading SNR Fading Partition Mechanism Figure 3.2: The proposed VLMC channel modeling system for a non-stationary fading channel. SNR interval. Then, with the set of assigned letter corresponding to fading SNR samples, a context tree is constructed to obtain the VLMC model to represent the statistics of fading SNR samples. We determine the param eters characterizing the VLMC channel model, including the order of discrete Markov, the transitional prob ability, and the number of states, from the constructed context tree. The parameters of VLMC can be adaptively changed through the time-varying context tree struc ture to best represent current fading channel statistics. We will dem onstrate that the adaptive fading channel modeling based on VLMC will result in less complex and efficient channel modeling compared to traditional FSMC. 44 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 3.3 M apping from P h ysical M easurem ents to D ynam ic Variable L ength M arkov C hains (VLM C) 3.3.1 F ad in g C h an n el S N R D istr ib u tio n E stim a tio n The unknown time-varying fading channel SNR distribution could be estimated through a set of feedback fading channel state information measurements from the mobile terminal to the base station denoted by r = ( 7 ( 1) , 7 ( 2 ) , . . . , 7 The number L(t) of feedback channel SNR measurements is a time-varying pa rameter depending on how fast channel characteristics change. Based on these dis crete samples, we apply a kernel density estimation (KDE) method [122] for the fading channel SNR distribution estimation. The KDE method is adopted because of its smoothness and accuracy in comparison with other techniques such as the his togram method or the linear regularization method. The KDE-based channel SNR distribution can be w ritten as (3-D 4 5 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. where K(-) is an arbitrary probability density function (e.g., the Gaussian density function) and h is the bandwidth of the kernel function K(-). To determine the optimal h th at gives the best “closeness” between the estimated and the real channel SNR distributions, an iterative approach called the “solve the equation method” (STE) described in [122] can be used. To study the performance of fading channel SNR distribution estimation, first we compare the estimated fading channel SNR distribution obtained from kernel density estimation with the theoretical model of the Rayleigh fading channel with averaged power equaling 10 dB. The results are given in Fig.3.3 with different fading normalized Doppler frequencies. The sample size of SNR used for estimating fading SNR distribution is 5000. As seen from the given results, the estimated fading SNR distribution is closed to the theoretical one. Furthermore, we consider the effect of the sample size of SNR used in fading SNR distribution estimation. The Kullback- Liebler (K-L) distance that measures the distance between two distributions is used as a metric. The K-L distance [22] could be expressed as D(}(n), /(y )) = /'(7) In (3-2) where £ )(/(y ),/(y )) is the K-L distance between two distributions /(y ) and /(y ), which correspond to the estimated and real fading channel SNR distributions, re spectively. 46 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 0.08 — Theoretical model Normalized Doppler frequency = 0.005 — Normalized Doppler frequency = 0.01 0.07 0.06 0.05 0.04 0.03 0.02 0.01 6 7 8 9 10 11 12 13 14 15 SNR (dB) Figure 3.3: The fading SNR distribution estimate for 5000 simulated observations from Rayleigh fading channel with different Normalized Doppler frequency. The K-L distances of several estimation results are given in Fig.3.4, where the Rayleigh fading effect is considered. We try different sample sizes, ranging from 1000 up to 50000. The averaged power is set to 10 dB. As shown in the figure, when the fading channel is characterized by a higher normalized Doppler frequency (i.e., a fast varying channel), we have a better estimate for a fixed sample size. On the other hand, more samples lead to better estimation if the normalized Doppler frequency of the fading channel is kept the same. Furthermore, when the number of channel SNR samples is sufficiently large, the effect of normalized Doppler frequency becomes 47 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Normalized D oppler F requency = 0.01 Normalized D oppler F requency = 0.005 2.5' CD o c e g T o b .a CD 13 o CO 5 3 0.5 M emory (sam ples) Figure 3.4: The Kullback-Liebler measure between the actual and the estimated channel SNR distribution of the Rayleigh fading as a function of the number of samples used to estimate the channel SNR distribution parameterized by normalized Doppler frequency. negligible and the fading SNR distribution obtained from estimation is closed to the theoretical fading distribution. 48 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Estimated Channel SNR distribution Transmission policy Zero crossing points Figure 3.5: The partitioning of the channel SNR range. 3.3.2 D isc re te M arkov M o d elin g 3.3.2.1 Fading P artitioning To represent a non-stationary wireless channel with a dynamic Markov model, we first partition the whole SNR range (in the dB domain) of the estimated channel SNR distribution, -which is obtained from Section 3.3.1, to N intervals, i.e. 7 o < 7 i < 72 • ■ • < 7 j v - i < I n , where 70 and 7 ^ are the boundary points of the estimated channel SNR distribution as shown in Fig.3.5. Theoretically, 70 and 7 ^ can be —00 and 0 0 , respectively. In interval f 7 i_ i,7 i), a transmission policy denoted by 77 is used, which may represent a certain modulation type or some power control scheme. To determine the best partitioning set Y* = {7 1 , 7 2 , ■ • ■ , 7w -i) °f tlie whole SNR range, a cost function is introduced with the objective to enable the Markov model 49 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. to best capture the main features of an actual fading channel. Specifically, the cost function is chosen to minimize the difference of the probability of bit error properties between the Markov model and the real fading environment using the mean square error (MSE) criterion as error probability characterizing discrete Markov model when the fading channel condition is in the Markov states corresponding to the SNR range [ 7 i_ ii7 i) with transmission policy 77. For example, when the fading channel condition is in the discrete Markov state a as shown in Fig.3.9, where state a represents the SNR range [ 7 0, 71) (see Fig. 3.5), the bit error probability characterizing the fading channel condition at Markov state a is ei(7 T x). Let Pbi'j, 77) be the bit error probability of the real fading channel to be approximated by the discrete Markov model with the same transmission policy 77. For example, if the transmission policy is with the BPSK modulation, the bit error probability of the fading channel with symbol SNR equal to 7 can be described as (3.3) where /(y ) is the estimated fading channel SNR distribution and e^Wi) is the bit [82; P 6(7 ,i? P 5 R ) = Q ( ^ ) , (3.4) 50 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. where Q(-) is the Gaussian Q function. To minimize the cost function, we take the Similarly, by taking the partial derivative of the cost function with respect to and setting it to zero, we obtain To determine the best partitioning and the bit error probability of each parti- based on updated ^ ( 77), we compute 7 j with Eq,(3.6). The above process is re peated until the partitioning value 7 j and the bit error probability ^ ( 77) converge. The optimal set of both E* = {e^(7 Ti), e ^ ^ ) , . .. ,e*N(ipy)} and T* obtained from the iterative algorithm will be used as parameters to construct the VLMC chan nel model in Section 3.3.2.2. The performance of the proposed fading partitioning mechanism is shown in Figs. 3.6 and 3.7. The fading channel in the simulation is of the Rayleigh distribution with the normalized Doppler frequency equaling 0.01. Hence, with the same BPSK modulation transmission policy for all fading intervals, the bit error probability of fading SNR distribution given in Eq.(3.4) is utilized. partial derivative of the cost function with respect to e4(7 7 ) and set it to zero, which leads to f j h f W i JJU fli)/(7)<*7 (3.5) (3.6) tioning, we perform iterations between Eqs.(3.5) and (3.6). T hat is, we start with a set of initial values of 7 * and compute the corresponding e^zTj) with Eq.(3.5). Then, 51 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. As shown in Fig. 3.6, when the number of intervals in the fading partitioning mechanism increases, the set of bit error probabilities obtained from the partition ing mechanism is closer to that given by Eq. (3.4). This implies an improvement in approximating the fading channel parameters. The performance of the fading partitioning mechanism using the optimal 16-interval partitioning obtained via iter ative search is compared with that of 16-interval equal partitioning, where the size of each partitioned interval is the same, in Fig. 3.7. As shown in Fig. 3.7, the iterative optimal partitioning mechanism provides a much better approximation of the probability bit error than th at of equal partitioning, especially in the range with a higher probability. Sometimes, we may deal with constraints in communication system design, (e.g., to maximize the spectral efficiency based on an adaptive modulation with a power constraint [38]). Then, the cost function can be reformulated to obtain a set of the optimal fading partitioning which takes both spectral efficiency maximization and the target BER [38] into account: where M * is the modulation type used in the transmission policy for SNR interval % subject to the power constraint C2 = £ log2(AA)/(7)d7 (3.7) t f < a , i=i (3.8) 52 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. BPSK — 4 intervals — 8 intervals 16 intervals 10" o CD lo o > . J D CO J D O a! 10'4 SNR (dB) Figure 3.6: Comparison of bit error probabilities obtained by theory and the pro posed fading partitioning mechanism where A; (7 ) is the power policy for SNR interval i and A is the averaged power used in transmission. Then, cost functions in Eqs.(3.3) and (3.7) can be combined so that modulation type M * as well as power policy A*(7 ) are adopted as the transmission policy for SNR interval i. The combined cost function becomes N r / i mm(C'i + AC2) = m in(]T / (e^M*, Aj) - Pb(-y, Mh A*))2 • / ( y ) ^ (3.9) N rn +XJ2 log2(Mi)/(7)d7). i= 1 Jli-1 where parameter A can be properly chosen to obtain the best partitioning ( 7 i, 7 2 > ■ ■ ■ 17N-i} the possible SNR range. Iterative com putation is required 53 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. BPSK — Iterative partition — Equal partition o 0 ) o > . 0 1 2 3 4 6 5 7 8 SNR (dB) Figure 3.7: Comparison of bit error probabilities obtained by equal partitioning and the proposed iterative fading partitioning mechanism, with 16 intervals under the BPSK modulation and the Rayleigh fading environment to obtain a good choice of A , which determines the relative weight between the “closeness” of channel parameters between the Markov model and the estimated channel SNR distribution and the spectral efficiency of the adopted channel policy under the power constraint. In the following discussion, we focus on the cost function given by Eq.(3.3) for simplicity. 54 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 3.3.2.2 A daptive Tree Structure For V LM C In this section, an adaptive context tree structure is proposed to construct a variable length Markov chain (VLMC) [9], which changes its structure and parameters adap tively for the best modeling of a fading channel. VLMC is still a Markovian process whose structure has a sparse memory with some states lumped together. Therefore, VLMC possesses parsimonious structure for the transitional probability, which will reduce the complexity of parameter computation in comparison with the full Markov chain. VLMC can have a variable order in its structure. The example of VLMC can be illustrated in Fig. 3.9, where the maximum order of VLMC is 2. VLMC in Fig. 3.9 has a variable order, in which some Markov states are characterized by the first order Markov state (i.e. states a, c, and d) and some are characterized by the second order Markov state (i.e. states ba. bb, be ,bd). To describe the concept of VLMC, we first introduce a finite categorial space y, which is the set consisting of alphabets representing partitioned SNR ranges. For example, when the SNR range is partitioned to 4 intervals, the finite categorial space is assigned to be y = {a, b, c, d,} as shown in Fig. 3.5. The alphabet in y is assigned to each feedback SNR in F depending on which interval the feedback SNR belongs to. To give an example, if feedback SNR 7 (2) at time i is in the SNR range [ 7 0, 7 *), alphabet “a” is assigned. The alphabet set obtained from assigning alphabets in y to T is denoted by A. 55 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Based on the assignment procedure, we first let x) = x,fXj^\... xl be a sequel of alphabets in the reverse order, where 2^ £ x, i < f c < j and i, j are integers. Then, concepts of the “context” function and the variable length memory are introduced to serve as tools in VLMC construction [9]. D efinition 1 Let (X t)tez be a sequel with values X t £ x- We define a function denoted by c that maps an infinite sequence (an infinite past) to a possibly shorter string as c ' ■ Z -00 Z - l + i + t . I = /(arLoo) = min{k\ P (X t+1 = xt+1 {Xf^ = = P{Xt+i = xt+i\X (k+l+t = xfik+1+t), Mxt+i £ %}, where c(-) is called the context function and c(xt _0 O ) = x ( k+l+t is called the context of the process time t with length k. Let 0 < ( < 00 be the smallest integer such that i ( s ‘ _oo) < C e X °°- The number ( is called the order of the context function c(-), and i f f < 00, (Xfitez is called a variable length Markov chain of order (. Obviously, the VLMC of order ( is a Markov chain of order ( with an additional structure having variable order I inside its structure. As seen from the above defi nition, the context represents how previous alphabets contribute to the knowledge 56 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. of the incoming alphabet with the length of context corresponding to the number of the past alphabets. To give an example, consider a time series x i , ... ,xi{t) from an alphabet set A obtained from the fading assignment procedure described above, where Lit) is the number of fading sample used in channel modeling. Suppose th at the transi tional probability characterizing the VLMC based on assigned alphabet set A and categorial space y as in Fig. 3.5 can be described as P (X t — xt\Xt-\ — a), P(X t = XtlXt-! = c), P{Xt = x t\Xt^ = d), Xt- 1 = a x t-\ = C x t-i - d P (X t = xt\Xt_i = 6, X t^2 G { a , b, c, d}), xt^x = 6, x t~2 G {a, b, c, d} Hence, the contexts used to characterize VLMC in order to represent the fading channel in this example are a, x t-i = a, x ~ arbitrary c, x t-\ = c, x 1 '^2arbitrary d, xt-\ = d, x 1 ^ 1 arbitrary ba, bb, be, bd, xy„i = b, ay_ 2 G {a, b, c, d}, x ^ ^ a rb itr a r y . 57 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. W ith concepts of contexts and variable length memory introduced, we can de scribe the VLMC construction methodology. To construct VLMC from elements of A, a graphical tree structure can be effectively used. The param eters of VLMC, such as the order, the transitional probability, and the number of Markov states, can be determined from the structure of the tree as follows. D efinition 2 Let c(-) be a context function of VLMC and its context tree r be defined as t := {ug : ug = c{xt _00))ugu ^ r , V « € y I < (}, where u > i is the context of length I corresponding to nodes of the tree and ugu is the concatenation of context and alphabet u. The tree structure of VLMC does not have to be complete. It depends only on those contexts that have the influence to the occurrence of the incoming alphabet, thus reducing the size of parameters of a complete Markov model. The terminal nodes of the tree will represent the state of the VLMC, whereas the depth of the tree corresponds to the order of VLMC. Fig. 3.8 gives a tree structure representing the 2nd-order VLMC of the above example with 7 terminal nodes and 1 internal node. In contrast, the complete 2nd-order Markov model has 16 term inal nodes. The finite state representation of the fading channel characterized by the tree structure given by Fig. 3.8 is illustrated in Fig. 3.9. As seen from Fig. 3.9, each state in the VLMC corresponds to a terminal node of the tree and is equipped with the corresponding context, transitional probability, and probability of bit error from E* . 58 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 1 be bd • Terminal node O Internal node Context Figure 3.8: The tree structure of VLMC of order 2, where the set of contexts repre senting the state of fading channel SNR is {a, ba, bb, be, bd, c, d}. The parameters to characterize VLMC include the state probability and the state transitional probability of each node ('i.e. a specific context) in the associated tree structure. To compute the state probability and the state transitional probability, let us define where N(lu i) is the number of contexts obtained from A, PU l is the probability of context l l > i, and P(uq\lo i) is the conditional probability of uq, uq € x, given the past sequel of alphabets equal to context uy. Due to the quick growth of the number of nodes when the order of VLMC in creases, the order of VLMC is often limited to a maximum value, denoted by (max, to reduce the computational complexity during tree construction. The value of (max will be set to trade-off between the accuracy of VLMC in representing a fading channel and complexity in channel modeling. There are mainly two techniques to iV (^) P(u0\eJi) L(t) - l + l Njutuo) N (uji) ’ 59 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. e2(7i2) P(b|bd) >(b|ba) P(b|bc; P(b|bb! P (a|b J = {P(a|ba),P(a|bb),P(ajbc),P(albd)} P frlb J “ {P(c|ba),P(c|bb),P(c|bc)P(c|bd)} P (c|b J = {P(c|ba),P(c|bb),P(c|bc),P(cjbd)} P(b|a) P{b|c) P (c|b J P (d|bJ P(b|d) JW bJ P(a|a) e iC i) e3(jt3) e4(m 4) Figure 3.9: The finite-state representation of a fading channel. construct the tree and compute VLMC’s parameters from a given set of feedback channel status A. They are the forward and the backward approaches. The back ward approach builds up the tree with a given set of channel status until the depth of the tree reaches (max. Then, it prunes the nodes corresponding to the context ujuo, where u0 £ y, when P(-|cc) ~ P{-\ujuq). However, the backward approach requires some unnecessary calculations due to the VLMC param eter com putation of the pruned nodes. The extra calculation burden of the backward approach makes it less attractive in determining the adaptation VLMC structure during the communi cation process. In this work, the forward approach is adopted to construct the tree due to a lower complexity required to calculate channel parameters of the VLMC in comparison with the backward approach. The forward approach for constructing the VLMC tree structure can be described step by step as follows. 60 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Step 0: Initialize the root node of the tree. Let j — 0, where j is the length of the context. Set up the set of context having length 0 and define as an empty set, Q0 = {}. Step 1: For each element in the set of context having length j, ujj € Qj, compute P(u'\ujj), P(u'\iOju), PUj,and PW jU from A, where u ',u G %. Step 2: Compute the averaged Kullback Liebler (K-L) distance between P{-\ujj) and P{-\u!ju) from parameters obtained from Step 1 as u'€x P(u'\uijU) Select two thresholds K\ and K 2. If D > K\, add the children node cor responding to context ujjU to the parent node with context uij. If D > K 2l include context ujjU in the context set Clj+i of length j + 1. Note that K\ mea sures whether uiju improves the knowledge in predicting the incoming alphabet compared with uij while K 2 measures the tendency to improve the knowledge of incoming alphabets if one continues to grow the tree from node tuj. • Step 3: Set j = j + 1, and return to Step 1. The algorithm for the tree construction will continue until the length of the context reach (max or the context set flj of length j > 0 is an empty set. The time complexity of the VLMC channel modeling with the context tree results from the tree construction and the computation of transitional probabilities. The 61 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. averaged time complexity of tree construction is O (nlogn), where n is the terminal node number of the tree. The computation of transitional probabilities during tree construction is 0{m N node), where m is the number of feedback fading channel SNRs used in channel modeling and Nnode is the total node number of the tree. Thus, the total time complexity of VLMC channel modeling is 0(n log n + m Nnod e)• The complexity of channel modeling lies mainly on the transitional probability computation. Due to the variable order structure of VLMC, the number of tran sitional probabilities for characterizing VLMC of order Q max is less than or equal to |xl^m < “ (lxl 1); which is required by FSMC of order (max- For example, the number of transitional probabilities needs to characterize VLMC of order 2 with |x| = \{a,b,c,d}\ = 4 is less than or equal to 16. This shows the advantage of VLMC in reduction complexity when compared to traditional FSMC. 3.4 C hannel M odel U p d ate The number of feedback fading SNRs used to obtain the channel model highly af fects the performance of the proposed adaptive channel modeling. The performance degradation can be categorized into the following two cases. 1. If the number of feedback fading SNRs used for channel modeling is not enough, the current fading statistics may not be well represented with the proposed VLMC channel modeling. 62 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 2. If the channel is highly dynamic, the excessive number of feedback fading SNR results in the loss of sufficient adaptivity in tracking the change of fading statistics. Note that in practical applications we may apply the training mode periodically for channel model update, where the training sequence is used to obtain the fading statistics during a certain period of time and then the training mode is switched off. Although the training mode is switched off, the parameters of channel model may still continue to be improved from the feedback fading SNRs of the transm itted data. In this work, to determine the proper time interval of each channel model up date, (i.e. the value of L(t)), two sets of feedback channel SNRs are introduced. One contains the accumulated long-term feedback channel SNRs and the other the latest channel SNRs. The short-term feedback channel SNRs are used to update the chan nel only if the fading statistics computed from long-term and short-term feedback channel SNRs are different. The K-L distance of fading statistics computed from the long-term and the short-term channel SNRs is used as a metric for similarity comparison. We consider two criteria in making the decision whether the short-term or long-term feedback channel SNRs are used for channel model update as follows. 63 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 1. Suppose th at / s(7 ) and fi(j) are the estimated fading SNR distributions of the short-term and long-term fading, respectively. The KL distance between fsij) and //(p) can be computed. If it is greater than threshold e, i.e. D(fi(7)\\fs(7)) = [ .M7) l°g > £ > (3-10) J-o° Jsil) the short-term feedback channel SNRs are used for channel model update. 2. If the condition in (3.10) is not met, the following criterion is considered in stead: D(Pl \\Ps) = E E > e, ( 3 . 1 1 ) u > jeT'u0€x ”s('M o|W J) where r is the context set consisting of terminal nodes of the tree, Vc and Vs are sets of VLMC transitional probabilities obtained from the long-term and the short-term feedback channel SNRs, and P L (u 0\ujj) and P s ( u 0\iOj) are transitional probabilities obtained from the long-term and short-term channel characteristics, respectively. The short-term feedback channel SNRs are used for channel model update if the condition in (3.11) is met. Otherwise, the long term will be used for channel model update. In above, e is a small positive value that plays the im portant role as a threshold in switching between short term and long-term data. 64 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Fading statistics 1 Fading statistics 2 Fading statistics 1 T2 Time Figure 3.10: Adaptation of the tree structure along with the changing of fading statistics. Note that the first criterion is used to track the change of the fading SNR dis tribution. {e.g. from Rayleigh to Rician). The second criterion is used to track the change of fade correlations, which may happen when the mobile speed is changed along the time {i.e. the fading channel SNR distribution may not change even though the fading correlation is changed). Fig. 3.10 illustrates the adaptation of the tree structure along with the changing of fading statistics. 3.5 A pplications of D erived V LM C In this section, we present two applications of the derived VLMC model. The first one is the computation of fading parameters, including the fade duration, the level 65 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. crossing rate, and the channel capacity. The second one is the design of a transmis sion protocol with an adaptive packet-size. The adaptive packet-size protocol im proves the throughput by adaptively changing the packet-size along with the varying channel parameters estimated by using the derived VLMC. 3.5.1 F ading P a ra m eter E stim a tio n The fading parameters considered here are the fade duration, the level crossing rate, and channel capacity. They play an im portant role in understanding the fading characteristics for efficient communication system design. Since the fade duration and the level crossing rate involves the second-order statistics of the envelop of the fading effect (i.e., the joint probability density), it is in general difficult to compute and obtain a closed form solution for them. Here, we would like to provide approximations to these quantities based on the derived VLMC. The averaged fade duration is defined as the period for which the fading SNR stays below a specified SNR level declared as the fade period. The fade duration is a useful channel parameter in the design of efficient transmission protocols. For example, it can be used to determine channel transmission parameters such as the rate of forward error correction (FEC) codes, the retransmission number, or the 66 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. length of packet transmission. The averaged fade duration below 7 * based on the VLMC model can be expressed as (3.12) where 07 is the context corresponding to the SNR range below 7 * , P{ooj) is the context probability of 0 7 , P{u>j\ujj) is the transitional probability from state 07 to itself in the VLMC model, and Ts is the symbol time. The level crossing rate is defined as the frequency for which the fading SNR crosses a specified SNR level in either the positive or negative going direction [104]. The fading crossing rate of the SNR level 7 can be written as where 7 is the fading SNR, 7 is the fading SNR slope, and / ( 7 , 7 ) is the joint probability distribution between the fading SNR and the fading SNR slope. As shown in Eq.(3.13), the computation of the level crossing rate of fading SNR demands the knowledge of / ( 7 , 7 ), which is difficult to derive. On the other hand, with the derived VLMC model, the fading level crossing rate of the SNR level j j +i can be easily approximated bj^ (3.13) (3.14) 67 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. where cum is the context corresponding to the SNR ranges below jj+i, L V j is the context corresponding to the SNR ranges above 7 j+i, and R is the symbol rate for transmission. Note th at there may be more than one context having the same SNR range. For instance, the four contexts {ba, bb, be, bd} in Figs.3.5 and 3.9 are mapped to the same partitioning range [ 7 1, 7 2). Moreover, the averaged channel capacity of the fading channel can be simply computed from the VLMC. Suppose that the averaged SNR corresponding to VLMC state representing SNR interval [ 71- 171), 7 i can be computed as Hence the averaged channel capacity for the fading channel can be computed as C = J2 + 7j)-P(VV?), (3.16) VW 7 7 1 1 ' where C is the averaged channel capacity, is the context corresponding to the SNR ranges [ 71- 17,;), and P{u>mj) is the VLMC state probability corresponding to context ujmj. 3.5.2 A d a p tiv e P ack et-S ize T ran sm ission To improve the efficiency of transmission protocols over non-stationary wireless chan nels, transmission parameters such as the FEC protection level or the transmission Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. packet length should be adaptively selected based on the channel characteristics. In this section, an adaptive packet-size transmission protocol is proposed for packet transmission over a non-stationary wireless channel. The objective of adaptive pack etization is to obtain the maximum throughput over the transmission period. The throughput of packet transmission over the non-stationary fading channel can be w ritten as [46] where Id, is the transm itted packet size, Dah is the overhead in the transm itted packet, and Ps{Di, T y ^ ), T,(7 j)) is the probability of successful packet transmission. To derive the probability of successful packet transmission, we consider the rela tionship between the packet size and inter-fade/fade durations [105]. The packet size is assumed to be less than the summation of one inter-fade and one fade durations on the average. The whole packet may stay completely in the inter-fade or the fade duration. It is also possible that the packet may stay in both the inter-fade and the fade durations. Hence, the probability of successful packet with packet size Di can be expressed as F5 .(A ,T / (7 ,),T ,(7,)) = P ,(A ) • (1 - &i)D i - A) + Pf(Di) (3.18) (3.17) .( l-e j)B ‘/(T/ (7i) -A ) + (v) ~ ' 69 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. (1 — ej)D % Ti ('/•)+* ■ Ti(~h) + Tf ('yi)dt' The first and second expressions in Eq.(3.18) represent the probability of suc cessful packet transmission when the whole packet is in inter-fade and fade duration, respectively. The last expression corresponds to the probability of successful packet transmission when the packet stays in both inter-fade and fade durations, where the portions of the underlying packet stay in interfade and fade durations are equal to t and Dl —t. respectively. Since Pi(Di) is the probability th at the whole packet stays in the inter-fade duration, it can be expressed as Similarly, Pf(D{) is the probability that whole packet stays in the fade duration and can be expressed as is the fade duration defined in Eq.(3.12), Tt(ji) is the inter-fade duration where the summation is over all contexts with an SNR range above 7 j (or SNR levels correspond to the inter-fade), Ta is the symbol time, P(ujm) is the context Tihi) + T / ( 7 i ) ’ given by (3.19) 70 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. probability of ujm, and P(ujm\uim) is the transition probability from context ojm to itself, and /(•) is the indication function defined as /(a ) = < 1 if a > 0 0 if a < 0 . (3.20) Let e .j and e3 be the averaged bit error probability of states obtained in Section 3.3.2.1, which correspond to averaged bit error probabilities of inter-fade and fade duration. Then, e* and e, can be written as e, = ^ e (c o i)P(o;l), U .H ei = U Jj where u t and ojj are contexts corresponding to inter-fade and fade, respectively, and P{uJi) is the probability of state corresponding to context uy. W ith Eqs.(3.17), (3.18), and (3.19), we can determine the optimal packet-size, which provides the highest throughput, dynamically based on the channel characteristics. 3.6 E xperim ental R esults In this section, we present a sequence of experiments to dem onstrate the main con cepts and results discussed in previous sections. We begin with the performance of 71 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. VLMC to model a fading channel under the stationary environment and its appli cation in Section 3.6.1. Then, the adaptive channel modeling algorithm is evaluated under non-stationary fading environments in Section 3.6.2. 3.6.1 S ta tio n a ry P erform an ce We evaluate the performance of VLMC for fading channel modeling. The parameters of the desired VLMC model such as the order, the number of transitional probability, and the number of Markov state are studied. To begin with, let us consider a short term flat fading and log-normal shadowing environment. The short-term fading is characterized by the Rayleigh distribution, where the averaged SNR value is equal to 10 dB and the modulation scheme is BPSK, while the log-normal shadowing of both urban and suburban environments proposed by Gudmunson in [35] is considered with the average of shadowing samples equal to 10 dB. To build the VLMC model, 106 samples of the short-term effect and the log-normal shadowing are generated. The channel SNR distribution is estimated from the generated fading samples. The estimated fading SNR distribution is then partitioned based on the discussion given in Section 3.3.2.1 with 4 intervals. The Q max of VLMC is set to 4 for complexity reduction in channel modeling. 72 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Table 3.1: The param eters of the VLMC channel model for the Rayleigh fading channel. “Rayleigh fading” Nf VLMC Order Number of Parameters Number of States Average Maximum VLMC FSMC VLMC FSMC 1 ■10" 3 1 1 10 768 4 256 5 • HR3 1.5 2 14 768 8 256 1 • 1(R2 1.987 3 27 768 12 256 3.6.1.1 Fading Channel M odeling via V LM C First, let us consider the VLMC channel model for the Rayleigh fading. The nor malized Doppler frequencies in simulation are 10”2, 5 • 1CT3 and 10-3. The order, the number of states, and the number of channel parameters of VLMC required to represent the Rayleigh fading in different normalized Doppler frequency conditions are shown in Table 3.1. Since VLMC has a variable order in its structure, the aver aged order of the VLMC channel obtained by averaging the order corresponding to the VLMC state is used as the required order of VLMC. When the fading channel correlation changes slowly (for the case th a t the normal ized Doppler frequency is equal to 1CT3), the first order Markov model is sufficient. However, when the change of the fading correlation is faster (for cases that the nor malized Doppler frequency is equal to 5 -10” 3 and 1CT2), the first order VLMC is not sufficient to model the channel well. As shown in Table 3.1, the averaged numbers of the order required at the normalized Doppler frequency of 5 • ICC3 and 10~ 2 are 1.5 and 1.987, respectively. Furthermore, when the fading correlation changes slowly, 73 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Table 3.2: The param eters of the VLMC channel model for the log-normal shadowing channel. “Log-normal Shadowing” P, o2(dB) VLMC Order Number of Parameters Number of States Average Maximum VLMC FSMC VLMC FSMC 0.3,4.3 1 1 16 768 4 256 0.82,7.5 2.18 4 34 768 11 256 the numbers of VLMC states and transitional probabilities are lower than those of a faster fading channel. VLMC is also utilized to model the log-normal shadowing effects. According to results in [35], correlations (p) between shadowing samples are set to 0.3 and 0.82 and variances (a 2) are set to 4.3 and 7.5 dB for urban and sub-urban environments, respectively. As shown in Table 3.2, the first order VLMC is sufficient to characterize the log-normal shadowing in the urban environment while the higher order VLMC is required to describe the sub-urban environment with an averaged order of 2.18. The numbers of VLMC states and transitional probabilities needed to describe the log-normal shadowing in the sub-urban environment are higher than those of the urban environment. To see the advantage of VLMC channel modeling over FSMC [120], let us com pare the numbers of states and transitional probabilities in Tables 3.1 and 3.2. The number of states required by FSMC with order 4 is 256 and the corresponding number of transitional probabilities is 768. As given in Tables 3.1 and 3.2, VLMC demands a smaller number in states and transitional probabilities to describe the 74 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. fading channel than FSMC. This advantage can reduce the channel modeling com plexity significantly. 3.6.1.2 Stochastic P roperty The stochastic performance of VLMC to model a fading channel is evaluated by studying its autocorrelation properties. In particular, we would like to compare the autocorrelation values obtained from the VLMC model with derived theoretical values of the log-normal shadowing and the short-term fading. The autocorrelation function of VLMC with N states can be calculated by where 7 (m) is the SNR representing one VLMC state at time m, p(7 (m) = 7 , ( 7 (0 ) = 7j) is the transitional probability from a state at time 0 with SNR equal to py to another state at time m with SNR equal to 7 * , and 7 ; is the SNR representing the VLMC state with its interval equal to [ 7 ^ 17,). These intervals can be obtained from fading partitioning and computed via (3.15). The autocorrelation of log-normal shadowing can be expressed as [35] N N R [ m } = EE 7 i • = 7 i | 7 ( 0 ) = lj); ( 3 . 2 1 ) i = i y = i 4>(m) = p' 7 1 1 (3.22) 75 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. — Lognormal - + - VLMC w / 10 partitions * VLMC w/ 8 partitions VLMC w/ 4 partitions 0.9 0.7 ' ■* - X • - x ■ - X ■ - X - - X - - X - X - - X ’ - i X - — X — X — X — X — X - X — X — X — X — X — X — X — -X — X — - X - > 0.3 0.2 0 5 10 15 20 25 30 m Figure 3.11: The autocorrelation comparison between the closed form of the log normal shadowing and those obtained from VLMC channel modeling in urban en vironment. where p is the correlation between shadowing samples. It is equal to 0.3 for the urban environment and 0.82 for the sub-urban environment. The comparison between val ues calculated with (3.22) and obtained from VLMC is shown in Fig. 3.11 and 3.12. It is observed that, when the number of fading partitioning intervals increases (which means more VLMC states), the stochastic performance of VLMC becomes closer to the statistics of the log-normal shadowing in both urban and sub-urban environ ments. However, the autocorrelation value between shadowing samples obtained from VLMC does not completely decorrelate due to the existence of transitional probabilities. 76 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. — Lognormal —f~ VLMC w/1 0 partitions -* VLMC w/ 8 partitions VLMC w/ 4 partitions 0.9 0.7 c 0.6 o « CD fe 0.5 o o <o. 0.3 0.2 <— X — x ~ - x ~ X — X - X - - H - 0 10 5 15 20 25 30 m Figure 3.12: The autocorrelation comparison between the closed form of the log normal shadowing and those obtained from VLMC channel modeling in suburban environment. The autocorrelation of the Rayleigh fading can be expressed as (j)(m) = J0(2irfdm), (3.23) where fd is the normalized Doppler frequency. The autocorrelation values calculated via (3.23) and those obtained using the VLMC model at the normalized Doppler fre quency of 1CT3 and 1CT2 are compared in Fig. 3.13 and 3.14. As shown in the figure, when the number of fading partitioning intervals increases, the stochastic property of the approximating VLMC is closer to that of the theoretical one. W hen the 77 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 0.9 0.7 D 0.6 o o o S < 0.5 — Rayleigh fading — VLMC w/ 1 0 partitions ■ - VLMC w/ 8 partitions — VLMC w / 4 partitions 0.4 0.3 0.2 250 200 300 100 150 50 0 m Figure 3.13: The comparison of autocorrelations obtained form the closed-form for mula and the VLMC channel modeling with the normalized Doppler frequency equal to 1CT3 for Rayleigh fading. fading channel changes slowly (corresponding to the normalized Doppler frequency equaling 10~3) , the autocorrelation of VLMC is closer to the theoretical value given by (-3.23) in a wider range in comparison with the faster fading case. It is the same as in the log-normal shadowing case that the autocorrelation obtained from VLMC cannot decorrelate completely due to the existence of transitional probabilities. 3.6.1.3 Fading Param eter E stim ation To verify the fading param eters obtained from the proposed VLMC model, the level crossing rate and the fade duration of the Rayleigh fading are examined and 78 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. — Rayleigh fading VLMC w /10 partitions — VLMC w/ B partitions VLMC w/ 4 partitions 0.5 c o o o o 3 < -0 .5 0 10 20 30 40 50 60 70 80 90 100 m Figure 3.14: The comparison of autocorrelations obtained form the closed-form for mula and the VLMC channel modeling with the normalized Doppler frequency equal to 1CT2 for Rayleigh fading. compared to the closed form solution. The level crossing rate of the Rayleigh fading channel with SNR equal to 7j +1 can be expressed as [38] / 27Tw,; , 1 u + l K „ T , = — , (3.24) where 7 is the averaged power of the Rayleigh fading, Ts is a symbol time, f d = FdTs is the normalized Doppler frequency, and Fd is the Doppler frequency. The fading level crossing rate at an SNR level equal to 73 as shown in Fig. 3.5 is examined. The 79 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. CD e p e g CO or cn _ c c o w o 6 CD > 0 ) Rayleigh Fading * VLMC Fading Model Normalized Doppler Frequency Figure 3.15: Comparison of level crossing rates obtained from the VLMC model and computed from the closed form solution for the Rayleigh fading channel. theoretical curve computed using (3.24) is plotted and compared with th at obtained from VLMC in Fig. 3.15. W ith the same parameters used in computing the level crossing rate, the closed form solution of averaged fade duration of the Rayleigh fading channel can be ex pressed as [104] fe (i )2 _ i ) T (M 5) where is the averaged fade duration with its SNR below level 7 . In the experiment, the channel is assumed in the fade duration if the SNR level is below 73 80 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 450: \ Rayleigh Fading -* VLMC Fading Model \ \ 400 \ \ \ Normalized Doppler Frequency Figure 3.16: Comparison of the averaged fade duration computed from the VLMC model and the closed form solution for the Rayleigh fading channel. as shown in Fig. 3.5. The averaged fade duration obtained from the VLMC channel model is compared to that of the closed form solution given by (3.25) in Fig. 3.16. The comparison of the averaged channel capacity between the Rayleigh fading and the VLMC channel model is shown in Fig. 3.17. The normalized Doppler frequency of consideration is 10“2. The averaged channel capacity of the Rayleigh fading can be computed via (3.26) 81 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 5.5 3.5 VLMC Fading Model — Rayleigh Fading 2.5 SNR (dB) Figure 3.17: Comparison of the averaged channel capacity computed from the VLMC model and the closed form solution for the Rayleigh fading channel with the nor malized Doppler frequency of 1CT2. where 7 is the averaged SNR value. As given in the figure, the averaged channel capacity estimated using (3.26) and obtained from VLMC is close to each other. 3.6.2 A d a p tiv e C h an n el M o d elin g The performance of adaptive channel modeling is considered in this section. The Rayleigh fading channel is used in our simulation. The non-stationary environment is simulated by randomly selecting the averaged power and the normalized Doppler frequency in a range. The averaged power is varied from 10 to 20 dB while three normalized Doppler frequencies (10~2, 5 • 10~3, and 1 0 ~ ~ 3) are chosen. The fading 82 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. channel characteristics are assumed to change from one interval to another along time. The performance of adajjtive channel modeling is compared w ith that of non- adaptive channel modeling. The estimated channel SNR is partitioned to 4 intervals, and the maximum order (max of VLMC is set to 4. 3.6.2.1 A dap tation Perform ance First, we consider the adaptation to an enviroment with a time-varying SNR channel condition. The speed of adaptation (short-term feedback) is set to each interval containing 10,000 fading samples. The non-adaptive channel model is chosen to represent the Rayleigh fading channel at SNR equal to 12.8dB, which is actually the channel model obtained from the first interval as shown in Fig.3.18. The normalized Doppler frequency is fixed at 10~2. The absolute difference between the averaged channel capacity obtained from the VLMC channel model and the Rayleigh fading can be written as A C = | CR ayleig h — C | , where C-uayieig} x is obtained via (3.26) and C is the averaged channel capacity obtained from VLMC via (3.16) with the averaged SNR of the changing environment. It is used as the performance metric in the top of Fig. 3.18. Furthermore, the KL distances between the theoretical Rayleigh fading distribution and adaptive/non- adaptive VLMC channel models are also shown in the middle of Fig.3.18. It is clear that the adaptive channel model under a time-varying power environment provides 83 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 30 < 1 0.4 9 > 0.3 .2 0.2 o * 0.1 0 Adaptive C hannel Modeling - x - N on-adaptive C hannel Modeling ■ x ? * * * > V % \ a * . A ' A Ant /At 1 A. * x. * if 1 M ' \ ' \ l] 1 ' \ ' t y ' / ' ; 1 ' ^ x j v U Ns * 5 0 10 15 20 25 30 35 40 45 50 A daptive C hannel Modeling N on-adaptive Channel Modeling l\ I \ ' 1 t I ' M I t| ' ' \ i * M , S * ' / V t 0 10 15 20 25 30 35 40 45 50 20 C Q " O cc 15 z C O 10 0 5 10 15 20 25 30 35 40 45 50 interval Figure 3.18: Performance comparison between adaptive and non-adaptive channel model in terms of the difference of the channel capacity (top) and the KL distance (middle) under a time-varying power environment as indicated in the bottom of the figure. a far better approximation in terms of a lower KL distance and a lower channel capacity difference than the non-adaptive channel model. N ext; we consider the adaptation to an environment with a time-varying mobile speed, which is represented by the time-varying normalized Doppler frequency. We use the level crossing rate at the SNR level of 73 as shown in Fig.3.5 as the parameter to measure the change of the fading correlation. We compare the difference of the level crossing rate obtained from adaptive/non-adaptive channels given by (3.13) with th at of the Rayleigh fading channel given by (3.24). The non-adaptive channel 84 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Adaptive Channel Modeling - x - N on-adaptive Channel Modeling £ 0.01 x 0.005 40 50 20 25 Adaptive Channel Modeling N on-adaptive Channel Modeling 2.5 20 0.01 Q- 0.005 25 Interval 40 45 20 Figure 3.19: Performance comparison between adaptive and non-adaptive channel models under a time-varying mobile speed environment. model is fixed to the one obtained using data in the first-time interval. As shown in Fig. 3.19, the adaptive channel model is superior to the non-adaptive one in tracking channel characteristics. The level crossing rate obtained from the adaptive channel model is more accurate. Furthermore, the parameters of VLMC such as the averaged order and/or the state number of an adaptive channel model can vary dynamically and represent the Rayleigh fading in a time-varying mobile speed environment better. In Table 3.3, the performance of adaptive channel modeling with different adap tation speeds is considered. The memory represents the number of short-term feed back channel samples used in the channel model update. We observe that, when 85 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Table 3.3: Performance for the Rayleigh fading channel approximation with various channel adaptation speeds. “Rayleigh fading” Memory (symbols) AC (bps/Hz) A Level Crossing Rate • Symbol Time Normalized Doppler Frequency 1 ■ 1 0 ~ 3 5 • 10” 3 1 • 1 0 -3 1000 0.4130 0.0025 0.002 0.00067 5000 0.1486 0.0013 0.0015 0.000576 10000 0.0272 0.0005 0.001 0.00027 the number of channel samples is smaller, errors in estimated channel parameters such as the averaged channel capacity and the level crossing rate become larger. This implies th at the number of channel samples is not enough to model the channel accurately. These errors can be reduced by choosing a proper adaptation speed. When the number of channel samples used for channel update becomes larger, the VLMC channel model provides a better approximation of channel statistics through adaptive channel param eter estimation. However, the choice of the adaptation speed is highly dependent upon the dynamics of fading statistics. W hen the statistics of a fading channel vary too fast, a large memory may result in inefficiency in tracking the changing fading channel statistics. 3.6.2.2 A daptive Packet-Size Transm ission The adaptive packet-size protocol is examined for packet transmission over a non- stationary Rayleigh fading channel. The BPSK modulation is used. The packet- size is chosen to match the time-varying channel characteristics to maximize the 86 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. throughput. The fade and inter-fade duration definitions are set to the level SNR below 73 and above 7 3 , respectively, as shown in Fig. 3.5. The averaged SNR of the Rayleigh fading channel is set at 10 dB. The overhead of the transm itted packet is 16 symbols. For the adaptive packet size protocol, the packet size is chosen in the range of 50-500 symbols to maximize the throughput. The throughput of packet transmission over a non-stationary Rayleigh fading channel is simulated and reported in Fig. 3.20. The short-term feedback channel SNR is set to 10,000 fading samples for channel adaptation. The normalized Doppler frequency in each segment is randomly generated from three values [i.e. 10~2, 5 • ICC3 and 10~3). Compared to the fixed packet-size transmission protocol, the adaptive packet-size protocol provides a higher throughput over the non-stationary channel environment. 3.7 C onclusion The adaptive channel modeling for a non-stationary fading channel was proposed in this research based on the VLMC structure. It consists of two main components: the channel SNR distribution estimation and VLMC channel modeling. The channel SNR distribution was carried out with a kernel density estimation algorithm and used to estimate the unknown distribution of a non-stationary wireless channel. Then, the fading partitioning mechanism and VLMC channel modeling were performed. The performance of the proposed adaptive channel modeling m ethod and its applications were verified by simulations in non-stationary wireless channel environments. 87 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 0.7 Adaptive Packet Size — Packet Size = 50 sym bols — Packet Size = 80 sym bols 0.6 0.5 0.3 0.2 0.1 0 10 15 20 50 5 25 30 35 40 45 Interval 0.01 g. 0.008 jii 0.006 0.004 0.002 0 5 10 15 20 25 30 35 40 45 50 Interval Figure 3.20: The throughput comparison using adaptive and fixed packet sizes over a non-stationary wireless channel. Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. C hapter 4 D ynam ic QoS M anagem ent for P rioritized W ireless V ideo D elivery 4.1 Introduction W ith the development of 3G [40, 24, 71] and 4G [1] wireless standards, new broad band video applications can be offered to mobile users. In addition to delivering high bit rate video applications, 3G and 4G systems are also expected to provide multi ple quality of service (QoS) guarantees to different types of user applications. For example, the packet-switched connection in the UMTS system provides 4 different services differentiated by delay sensitivity: conversational, streaming, interactive, and background classes [40]. An im portant issue in providing multiple QoS guar antees to video applications in 3G and 4G wireless systems is the dynamic QoS management for services with mobility support [1]. A dynamic QoS management 89 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. system allows both video applications and the priority network that supports mul tiple priority classes to interact with each other to cope with service degradation as well as overall resource management according to the changing characteristics of wireless channels [1, 39, 6 6]. Being different from wired networks, the wireless network possesses time-varying and non-stationary links due to the following factors: (1) fading effects coming from path loss, large-scale fading, and small scale fading [72], (2) roaming between het erogeneous mobile networks (e.g. from wireless LAN to wireless WAN), and (3) the change of the mobile speed, average received power, and surrounding environments [51, 72]. The variation of wireless link quality, measured by the level of signal-to-noise ratio (SNR) or the bit error rate (BER), results in time-varying available transmis sion bandwidth at the link layer (also called the channel service-rate [124, 10]). It also leads to time-varying delay of arrival video packets at the application layer, especially when the retransmission scheme is performed at the link layer. W ith a finite buffer size at the link layer, the time-varying channel service-rate can induce buffer overflow (and, therefore, video packet loss) due to the bit rate mismatch be tween the transm itting video packet and the channel service-rate. At the application layer, due to time-varying delay of arrival video packets, some video packets may be useless for playback if its arrival time exceeds its scheduled playback time. W ith time-varying link quality in a wireless environment, providing multiple QoS guarantees for video applications with absolute QoS differentiation [106, 116] 90 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. may not be effective. The transmission service may not be able to maintain the absolute end-to-end video quality as it commits, which may cause interruption or even term ination to an existing video service. Thus, it is more reasonable to provide QoS in a form of soft QoS guarantee, which allows the QoS param eters in the priority network to be adjusted along with changing channel conditions. The relative QoS differentiation [1, 28, 116] is one of the possible solutions to be used in the next generation adaptive QoS system. Along with the adaptive QoS provisioning of the prioritjf network, a video bitstream should also have the QoS-adaptive property, where video quality can be adjusted to maintain graceful transition due to changing channel conditions. Among several possible approaches for video quality adaptation e.g. video transcoding [89] and scalable video [123], scalable video is utilized in this chapter due to its low complexity and high flexibility. To efficiently adjust QoS param eters of both video applications and wireless com munication systems, interaction between these two modules should be established. Ideally, the QoS provisioning mechanism should find a good compromise between the video quality requirement and the available transmission resource. However, different QoS domains of video applications and wireless communication systems make it difficult to choose optimal QoS param eters during communication. On one hand, the transmission of the priority netwrork offer the QoS guarantee in terms of the probability of buffer overflow and/or the probability of delay violation. On the other hand, the QoS of video application is measured objectively by the mean 91 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. squared error (MSE) and/or the peak-signal-to-noise ratio (PSNR). Therefore, the mapping between these two QoS domains is required to make efficient selection of QoS parameters possible. In spite of the research work carried out in the past (see the review in Section 4.2, there are still some challenging problems especially under a wireless channel with a time-varying service rate. They are stated below. 1. An adaptive QoS service model, which allows the QoS param eters of both priority networks and video applications to be adaptively adjusted according to time-varying wireless channel conditions. 2 . An cooperative communication between priority networks and video applica tions to provide the QoS selection to meet the video quality requirement and the transm itting capability of the priority network. 3. Resource allocation within the priority network based on its QoS guarantee by considering time-varying wireless channel characteristics. To address these issues, a novel dynamic QoS management is proposed in this research. The QoS parameters of video applications and priority networks are adap tively adjusted based on their interaction. The available transmission resource is allocated to each service class based on its QoS requirement. Then, the expected video quality corresponding to the admissible QoS provisioning set of the priority network is derived under the QoS mapping mechanism. The main contributions of this work include the following. 92 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 1. The derivation of the rate constraint of a priority network with proper trans mission resource allocation. 2. The development of a QoS mapping mechanism th at optimally maps statisti cal QoS guarantees of a priority network to video classes based on allocated transmission resource. 3. The establishment of an QoS interaction procedure between video applications and the priority network to provides the trade-off between the video quality re quirement and the transm itting capability of the network based on the current channel condition. The rest of the chapter is organized as follows. Some previous work is reviewed in Section 4.2. The proposed dynamic QoS management framework is described in Section 4.3. The derivation of the transmission rate constraint of a priority network, is presented in Section 4.4. The QoS mapping between video applications and the priority network for wireless time-varying service rate channel is discussed in Section 4.5. In Section 4.6, the interaction between video applications and the priority network is studied. Simulation results are given in Section 4.7. Some concluding remarks are given in Section 4.8. 93 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 4.2 R eview o f P revious W ork There have been some efforts in the study of priority networks th a t provide mul tiple priority classes for multimedia delivery [107, 108, 83, 109, 114, 74, 75]. Most previous work focused on the partitioning of multimedia data into smaller units and transm itting these units with different priority classes. The partitioned multime dia units are prioritized based on its contribution to the expected quality at the end-user while the priority network provides different QoS guarantees depending on its corresponding service priority. We give a brief review of previous work in this section. Servetto et al. [107] proposed an optimization framework to segment a variable bit rate source to several substreams. Then, the resulting substreams are transm it ted in multiple priority classes with ATM connections. The objective of this scheme was to minimize the expected distortion of the variable bit rate source due to trans mission. Shin et al. [108], Tan et al. [114], Sehgal et al. [109], Padm annabhan et al [83], M artin [74], and Masala et al. [75] used the priority network to deliver multimedia data. Shin et al. [108] prioritized each video packet based on its error propagation effect if it is loss. Video packets were mapped differently to transmission service classes with the objective to maximize end-to-end video quality under the cost and/or price constraint. Tan et al [114] and Masala et al. [75] examined the same problem as that, formulated in [108] with different approaches for video prioriti zation. The transmission of prioritized speech packets over a network with multiple 94 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. priority classes was considered by M artin [74], Sehgal et al. [109] provided an op timization framework for audio transmission over a network with multiple priority classes employing the feedback-on-demand setting. 4.3 D ynam ic QoS M anagem ent Fram ework for T im e-V arying W ireless C hannels The proposed dynamic QoS management framework is shown in Fig. 4.1. The proposed scheme considers video packet transmission through a wireless link-layer model, which has a time-varying and non-stationary channel service-rate (or, equiva lently, the available transmission bandwidth). Here, we choose the wireless link-layer model since it is more suitable for QoS simulation and analysis in a packet-switched network than the physical channel model. (Note that it is difficult to compute QoS metrics such as the packet loss probability and/or delay violation using the phys ical channel model [124].) The time-varying non-stationary chan n el-service rate is modeled by the discrete Markov model (see Section 4.4.1), where the states of the Markov model characterize the channel service-rate under different channel condi tions. The Markov model is fully described by its transitional probability matrix. Due to the non-stationary property of the 'wireless channel, the param eters of the Markov model (e.g. elements in the transitional probability matrix) have to be 95 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. periodically updated to track the changing channel characteristics as described in Chapter 3. The transm itted video packets are characterized by their loss and delay proper ties, which contribute to end-to-end video quality. The priority network provides differentiated services based on the statistical QoS guarantee {i.e. QoS is guaranteed in terms of some probabilities) of buffer overflow and delay violation probabilities. The scheduling algorithm for video packet transmission in the priority network is chosen to be the absolute priority scheduling with the following properties. 1. Video packets in a higher priority transmission service are transm itted first until its buffer is empty. Then, video packets in a lower priority transmission service will have a chance to be transm itted. 2. Video packets in the same buffer is transm itted sequentially. 3. The video packet th at misses its scheduled playback time is not sent and flushed out from the buffer. Based on effective bandwidth and effective capacity theory as described in Section 4.4.2 [25, 124], statistical QoS guarantees of transm itting multiple priority classes can be translated to be the transmission rate constraints (see the discussion in Sec tion 4.4.3). The transmission rate constraints will specify the maximum data rate, which can be transm itted reliably with statistical QoS guarantee in each transmis sion service. They can be used as guidelines in allocating video bitstream s and transmission resources for video transmission. 96 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Transmission Video QoS QoS requirement provisioning Video substream 1 Q oS(l) Video substream 2 QoS M apping & QoS Adaptation QoS(2) Transmission Control Video Encoder Video substream K Video Input QoS(K) Transmission Module Video M odule Adaptive Channel Modeling Channel Feedback Time-Varying and Non-stationary Wireless Channel Video Decoder Video Output Figure 4.1: The system architecture of QoS adaptation for a wireless video trans mission system. To select proper QoS parameters of video applications and the priority network under dynamic channel conditions, their interaction should be implemented as shown in Fig. 4.1. The objective of this interaction is to find a trade-off to simultaneously provide a desired video service with the available transmission resource. Due to different QoS domains of video applications and priority networks, a QoS mapping mechanism is needed to translate QoS parameters from one domain to the other. W ith the derived transmission rate constraint corresponding to the QoS parameters of the priority network, the optimized QoS mapping from video bitstreams to various priority classes in the priority network can be obtained via dynamic programming. 97 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 4.4 R ate C onstraint D erivation for P riority N etw orks In this section, the transmission rate constraint of a priority network under a wire less time-varying service-rate channel is derived. The transmission rate constraint specifies the maximum data rate transm itted reliably with statistical QoS guaran tee. It serves as an effective tool to allocate the resource for data transmission with varying link quality. Time-varying service-rate wireless channel is first characterized in Section 4.4.1. Next, the background of effective bandwidth and effective capacity theory is reviewed in Section 4.4.2. Then, the derivation of the transmission rate constraint for a priority network that transm its multiple priority classes is presented in Section 4.4.3. 4.4.1 T im e-V aryin g S e r v ic e-R a te W ireless C h an n el The time-varying service-rate wireless channel is modeled by the first-order L-state Markov model in this work as shown in Fig. 4.2 [120]. The channel state at time u is expressed as X c(u), where X c{u) € {1,..., L), in this model. Each possible channel state X c(u) = i represents a channel link condition characterized by channel service-rate ri, which is the available transmission rate. 98 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Figure 4.2: The discrete Markov model of a wireless time-varying service-rate chan nel. In a wireless time-varying service-rate channel, the channel service rate of Markov state i at the link layer (in the unit of bps) can be computed as rt = R - log2( 1 + 7i)> (4.1) where R is the bandwidth in Hz and 7 , is the SNR value used to measure the wireless channel at Markov state i [51]. The channel model can be characterized by the transitional m atrix ( \ Pll, ■■■ PlL Ptransition \ P l i ■ ■ ■ P l l (4.2) / where pi7 = P (X c(u) = j\X c(u — 1) = i) and is the transitional probability from Markov state i at time u — 1 to state j at time u. Note that, under a non-stationary 99 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. environment with the time-varying mobile speed and the averaged power [51], the parameters of the Markov model are time-varying and should be updated periodically to keep track with the changing wireless channel conditions [51]. Next, let us consider the general framework in computing the expected channel service-rate from the Markov model. Assume that the channel model is updated at time u — b due to the non-stationary nature of the wireless channel. Suppose that, from time u — b to time u, the underlying Markov model is nearly stationary so that the transitional probability m atrix of the Markov model does not change much. Assume that the transitional probability m atrix can be expressed as in Eq. (4.2). Let the state probability of the Markov model at time u — b be P state{u - 6) = [Pi(u - 6),... ,pL(u - b)}, (4.3) where pj(u — b) is the probability of being in Markov state j at time u — b. Hence, the state probability of the Markov model at time u can be expressed as P state{ u ) — Pstate(u P)Ptransiti0n- ( 4 - 4 ) 100 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Hence, with the derived Markov state probability at time u, the expected channel service-rate at time u, denoted by rC hannei{u), given the updated channel model at time u — b can be written as L channeli'U) ^ , d ' Pi(^) j (45) i— 1 where p% (u) is the probability of state i at time u, and r* is the channel service-rate of state i obtained from Eq. (4.1). 4 .4 .2 E ffective B a n d w id th and C ap acity First, with K priority classes, it is assumed that service class i has a higher priority than service class j, if i < j. Service class i provides a statistical QoS guarantee for transm itting data in terms of the packet loss probability. It can be derived from theory of large deviation as [124] P(Bi(t) > Bmax?) « Q • , (4.6) where B;(i) is the buffer occupancy of service class i at time t, BmaXi is the maximum buffer size of service class i, 9u is the QoS exponent corresponding to the guaranteed packet loss probability provided by service class i, £ * is the probability that the buffer of service class i is not empty, and ^ • e~ei’iBmaxi is the guaranteed packet loss probability of service class i. 101 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. The accumulated data amount of the source stream generated for transmission under service class i from time 0 to t is a random variable of the form: a,i(t) = / rs i(Xs i(u))du, Jo where rs i(XS ii(u)) is the source data rate generated for service class i at time u under state X S'i(u). The data of amount ai(t) will be stored in the buffer of size Bm ax> i before transmission over service class i. Next, the time-varying service-rate channel effect for substream transmission should be considered. Let the random variable of data to be transm itted over service class i from time 0 to t be of the form Si(t) = [ Ci(Xc(u))du, Jo where at(Xc(u)) is the channel service rate of class i at time u with the random channel state X c(u). The stochastic behavior of the source and channel random processes can be described using the concept of effective bandwidth [25, 59] and effective capacity [124]. Given the QoS exponent 0iti, we have {a \_ __ h-aj(t){Ql,i) i n , (ft 1 V l,i 102 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. where eO i(t)(0i,i) and pSi{t){6i,i) are the effective bandwidth and the effective capacity, respectively. Given the source characteristic, the effective bandw idth provides a guideline about how much bandwidth a service channel should allocate to a data substream to meet the statistical QoS requirement. In contrast, the effective capacity imposes the constraint on the amount of information substream to be transm itted over the time-varying service rate channel while maintaining the statistical QoS guarantee provided by service class [124] under given channel characteristics. Furthermore, we let Aai(t)(Si,i) and A$(<)(#/,*) denote log-moment generating func tions defined as [25, 59, 124] A , _ r l o § E { e e^ ( t ) ) } £ — > c c t log E {e-e« s*W} t~*oa t To derive the transmission rate constraint of transm itting multiple priority classes, the theorem of the effective bandwidth and effective capacity [25] is utilized. It is summarized below. T h eo re m 1 Suppose the time-varying channel service rate is a stationary ergodic process satisfying the Gartner-Ellis condition with time varying channel service rate 103 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. c{t) and Of is the QoS exponent provided by transmission service class, which corre sponds to the effective capacity Then, we have n(Qr) »s(t){et), 0< er <6f, (4.8) + e - ^ e s n ( 9 r - 6 1 ) , 6 r > 6 1 , where 6r > 0 is the QoS exponent corresponding to the packet loss probability required by transmitting information and p(6r) is the effective capacity with QoS exponent 6r. Note that e * „ ( f c - V) = in above can be viewed as the effective bandwidth of S(t) with the QoS exponent 6r - Of. 4.4.3 D e riv a tio n o f T ran sm ission R a te C on strain t The transmission rate constraint of a priority network under a time-varying service rate channel is derived in this section. We start the derivation by first assuming that there are only 2 priority classes with QoS exponents 6iti and 6^2 corresponding to their guaranteed packet loss probabilities. The generated data substream for 104 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. transm itting over the first and second service classes from time 0 to t are on (f) and 02 {t) axrd stored in different buffers of sizes Bmaxi and B maX2, respectively. The statistical QoS guarantee of each service class is provided in form of the packet loss probability as shown in Eq. (4.6), which is computed based on the QoS exponent and the buffer size. W ith the absolute priority scheduling, the second service class has a lower priority than the first service class, and will be served only after all data in the buffer of the first service class is served. For the first substream in the high priority buffer, it is easy to see that the rate constraint of substream 1 with QoS exponent 6V ,i transm itted over the service class 1 with QoS exponent and buffer size BmaXl can be shown as min(/ri(#r. i < rchannd(t), (4.9) or min(Akn(t)(^,l)> a lOO) < rchannel (t), 0 < 1 < dti! (4.10) ■ ( to \ ' V , 6 V ,1 - 0 2 ,1 “r, 1 cy,! ^S i(i)( 0r,l 02, l ) ; 2^1 (t)) < ^"channel (t) j 0r,l ^ @ 1,1> where /ii(6 V,i) is the transmission rate constraint of substream 1 and rchannei{t) is the expected channel service rate computed from Eq.4.5. S\{t) is the random variable 105 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. of information that can be transm itted over service class 1 under the time-varying service rate c\(t) = c(t) from time 0 to t. For the low priority substream, the existence of substream 1 affects the trans mission rate constraint of substream 2 due to the absolute priority scheduling al gorithm. The derivation of the transmission rate constraint of substream 2 can be simply viewed as trying to transm it substream 2 alone with time-varying channel service rate c2(t) = c(t) - min ( / i , 1 ( 0 1 ) , a^t)), (4.11) where c2(t) is the time-varying channel service rate, which is seen b)^ substream 2 with the existence of substream 1. Suppose that substream 2 has its own QoS exponent requirement equaling 6bi2. Hence, from Theorem 1, the transmission rate constraint of substream 2 can be computed based on c2(t) as /As2( t) ( 0z,2) 0 < @ r,2 < @ 1,2 ■ < 4'12) + T T 2 * e S ,( t ) ( ^ .2 — ^ , 2 ) 0 „ 2 > $i,2 where e s 2{t){@r,2 ~ @ 1,2) is the effective bandwidth of 5 2(t) with QoS exponent provi sioning @ r,2 @ 1,2• S 2(t) is the random variable of information that can be transm itted over service class 2 under the time-varying service rate c2(t) from time 0 to t. 106 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Together with Eq. (4.9), the transmission rate constraint on both substreams 1 and 2 can be expressed as min(/r1(6l rjl), a\ (t)) < rchannei(t), (4.13) and m in(/ii(0T .ii), aq(f)) + mm(fi2{6rt2), a2{t)) < rchannel(t). (4.14) Eqs. (4.13) and (4.14) show that the transmission rates of substreams 1 and 2 are limited by fJL\{dr,\) and ^ 2(^ ,2), respectively. Moreover, the summation of the constraint on the rate of both substreams 1 and 2 should not exceed the expected channel service rate rchannei. Therefore, when the substream demands to send more data than the transmission rate constraint, in which the service class can allow with the statistical QoS guarantee, the rate shaper algorithm has to be applied to shape the information rate to meet with the transmission rate constraints. The procedure for deriving of the transmission rate constraint for 2 data sub streams can be easily extended to K substreams via k 'y ( , O i j { t)) < ^’ch a n n el (^), k l,2,...,iQ (4.15) 4 = 1 107 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. where /ii(#r,j) is the transmission rate constraint of substream i computed by assum ing th at the channel service rate seen by substream i. The effective channel service rate can be written as i —1 C i( t) = c(t) - ^ m i n ( / i 7(a,.J ), ay(t)), (4.16) i=i where 6T ti is the QoS exponent corresponding to the guaranteed packet loss prob ability required by source substream i and cq(f) is the random variable of data generation by the source. As seen from derivation, the transmission rate constraints of priority classes are dependent on each other. An example of this scenario is illustrated in Fig. 4.3 with K different classes. The choices of transmission rate at higher priority classes affect the lower priority classes. For example 2 choices (/xi(l) and /Ji(2), where /ii(2 ) > /ri(l)) of transm itted rate at the service class 1 (highest priority) provides different service-rate/packet loss probability curves at service class 2. When the choice of transm itted rate is /i1(l), the transm itting rate constraint of service class 2 is higher than that of the choice /ii( 2) given the statistical QoS guarantee. The same consideration can be applied for other service classes. 108 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Transmission Rate Constraint Transmission Rate Constraint Transmission Rate Constraint Hi(U) ( A ^ ' )h{2’v > M 2.0 Packet Loss Rate Service Class 1 Packet Loss Rate Service Class 2 Packet Loss Rate Service Class K Figure 4.3: Dependency of transmission rate constraint of transm itting multiple priority classes with QoS. 4.5 QoS M apping M echanism In this section, with the derived transmission rate constraint, the QoS mapping between statistical QoS guarantees of the priority network and the corresponding video applications is studied. The mapping scheme requires the information of the prioritized image sequence such as its bitstream size, distortion property, and delay requirement. Note that the proposed QoS mapping algorithm developed in this section can be applied to any kind of prioritized video coding (e.g., Relative Priority Index [108] or scalable video [123]). Here, we focus on the case of scalable video coding. Let us first consider the group of picture (GOP) structure of MPEG-4 PFGS [123] in Fig. 4.4. Suppose th at there are N frames in one GOP. Each frame consists of one base layer and several enhancement layers. Assume th at the video playback frame rate is fixed at F frames/second. If the mobile term inal starts to play back the first frame at time tp, frame n in the GOP should be received and ready to be 109 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. V ideo fram e 1 V ideo fram e 2 V ideo fram e 3 V id eo fram e N d isto rtio n B ase lay e r 1st en h an cem en t la y er 2nd en h an cem en t lay er 3rd en h an cem en t la y er V+l/F tp + 2/F P lay b ack d ead lin e t^+ (N -l)/F Figure 4.4: The GOP structure of MPEG-4 PFGS scalable video, displayed before time td(n) = tp + for uninterrupted playback. The base layer and enhancement layers in each scalable frame have the same playback time. The loss of each video portion contributes differently to the end-to-end video quality due to the dependent video encoding structure and the contents inside the frame. For example, a P frame depends on its preceding I/P frame. Also, the enhancement layers are useless if the base layer cannot be correctly received. The scalable video bit stream of the GOP is packetized to several fixed sizes before transmission over the priority network. Next, let us examine the priority network that can deliver multiple priority classes. Suppose that there are K priority classes as discussed in Section 4.4. Service 110 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. class i has a buffer size of BmaXi and a guaranteed buffer overflow probability of Based on large deviation theory, the QoS exponent 9i:i corresponding to the buffer overflow probability guaranteed by service class i can be computed from Eq. (4.6). T hat is, we have To summarize, with a given wireless channel model and absolute priority scheduling, service class i has the transmission rate constraint equal to jii{9r.i) as derived in Sec tion 4.4, where 8,ni is the QoS exponent of the guaranteed buffer overflow probability corresponding to the QoS requirement of video substream i. eter of service guarantees. The QoS exponent of the guaranteed delay bound can be derived from that of the guaranteed buffer overflow probability as [130, 124] where is the QoS exponent of the guaranteed delay bound of service class i. Given the QoS provisioning parameters of the priority network such as the buffer size and the transmission rate constraint, the optimal QoS mapping between the pri oritized video bitstream and the priority network can be established. The mapping scheme will decide an allocation policy, where portions of scalable bit stream will be assigned to a certain priority class. The objective is to maximize the expected e. m ax■ (4.17) The probability of playback time violation of each class provides another param- (4.18) 111 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. end-to-end video quality. The overall distortion will be derived using the dependent structure of scalable video [27] as shown in Fig. 4.4. It can be expressed as D e(N, K) = Do - E A D ,II^ , ( 1 - e(eW ) J I ,td.,)), (4.19) j where D e{N) K) is the total expected distortion from mapping N scalable frames to K different priority classes, and D 0 is the distortion if no video data are received, ADj is the distortion reduction if data unit j is correctly received. Note th at the term in above is the probability that data unit j and all data units j', which data unit j depends on, are correctly received while data unit j' is transm itted over the priority class i(j') with the QoS exponent corresponding to its guaranteed buffer overflow probability and playback deadline td , ■ On other hand, is the probability that video data unit j' is lost due to either buffer overflow or playback deadline violation when transm itted over service class i(j'). It can be computed as ),j' + (1 ~ (4.20) where £;,i(?'),y(0z,;(j')) and Q.gy) j'(#ci,i(y)> td-,) denote the probabilities that data unit j 1 is lost due to buffer overflow and playback deadline violation, respectively. Based 112 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. on Eq. (4.18) and the theory of large deviation, Qy) can be expressed as [124] where P(df > td.,) is the probability that video data unit j' arrives at the receiver later than td , • W ith the characteristics of scalable video and the QoS provisioning of the priority network, the optimization framework to search for the optimal mapping policy can be set up as follows. P roblem Form ulation I: Given the set of transmission rate constraints of the priority network derived in Section 4.4 and the expected channel service rate ^channel (t) at time t, determine the optimal mapping policy tt* from one GOP with N scalable frames to K priority classes such that (4.21) 7r* = argm inD e(fV, K), (4.22) subject to the following constraints (4.23) Vj'ei and k 'y [ m i n (fij(6rd), a.i(t)) < Tcjiannei(t), k — 1, 2,..., K , (4.24) 113 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. where is the transmission rate constraint of service class and by ^ is the size of video data unit j[ to be transm itted through service class i. There are two sets of constraints in the above problem formulation. First, the transmission rate of video bitstreams over each service class must not exceed the transmission rate constraint of the underlying service class. Second, the summation of transmission rate constraints of all service classes has to be bounded by the expected channel service rate. Note that, in the optimal allocation of the network resources to video applications, the maximum rate of video bit streams transm itted over service class i can be achieved when 0r< i = 6itl. Furthermore, the expected PSNR obtained from the optimal mapping algorithm can be viewed as a result of the mapping from the QoS provisioning network domain to the PSNR domain. 4.5.1 S o lu tio n to th e O p tim iza tio n P ro b lem The solution of the QoS mapping algorithm is studied in this section. It is derived based on the dynamic programming approach as shown in Fig. 4.5. In this case, the tree representing all possible solutions is created. Each stage of the tree corresponds to one of the video coding units j. To give an example, the video coding unit may correspond to video packet or the layer of scalable video. Each node of the tree at a given stage represents a possible cumulative rate occupancy in each service class. For example, as shown in Fig. 4.5, to get the accumulated data occupancy of video unit j — I, we add the rate to the buffer of each possible service class. Each branch has a 114 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. cost to indicate the expected distortion reduction corresponding to the transmission of video data unit j — 1 through the particular service class. The expected distortion reduction of each branch can be directly computed from Eq. (4.20). Therefore, as we traverse the tree from the root to leaves, we can compute the the accumulated expected distortion reduction for each of allocated solutions. Clearly, this is a way of representing all possible solutions since when the tree is traversed we get successive allocation of each video unit to all possible service classes. Due to the dependency structure among video information as shown in Fig. 4.4, the com putation of the expected distortion of data unit j demands the knowledge of how video units j ' , on which video unit j depends, is transm itted. Hence, when the tree is constructed, video data unit j is always in the stage th at follows the stages corresponding to data units j '. Let us consider a more specific example in the parameter com putation during the tree construction. Branch i from stage j to stage j + 1 has the associated expected distortion reduction resulting from the transmission of video unit j via service class i. The level of occupied resource is the size of video data unit j, which is equal to bj. The expected distortion reduction of video data unit j in association with branch i can be found via dij = — i (4-25) 115 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Prune node due to transmission rate constraint derivation Service class *kj+l Service class k+1 Service class k lk+lj+l Service class k ^ k j + 1 . Trace back Service class k+1 Service class k- lk+l j+1 Optimal node corresponding to lowest expected distortion Video unit j Video unit j+1 Video unit Figure 4.5: Illustration of the deterministic dynamic programming approach to de rive the optimal mapping algorithm. where (1 — e(6i!ij,tdj )) and (1 — £(Qi,i(j>),j', td ,)) are the guaranteed probabilities that video data unit j and j' are correctly received by transm itting through service classes i and i(j'), respectively. Their calculation can be achieved using Eq. (4.20). The accumulated rate and distortion reduction of nodes at stage j + 1 can be computed from stage j as ARij+i = Rj + AR\j, (4.26) and ADij+i — - Dj + ADij, (4.27) 116 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Distortion Distortion Distortion Packet Loss Rate Packet Loss Rate Packet Loss Rate Service Class 1 Service Class 2 Service Class K Figure 4.6: Dependency of the expected distortion of the priority network, where A R \j is the accumulated rate from state 1 to stage j , A D \j is the accumulated distortion reduction from stage j to stage j + 1 , Rj is the rate occupancy from stage j to stage j + 1, and Dj is the expected distortion reduction from stage j to stage j + 1. Based on this example, R j = bj and D j = d(i,j). Due to the transmission rate constraint of each service class as described in Eq. (4.23), it suffices to prune the branch th a t exceeds its corresponding transmission rate constraint. In other words, the tree cannot grow from the branch that vio lates the transmission rate constraint. The optimal mapping solution can be finally obtained by traversing back from the leaves, which has the maximum expected dis tortion reduction, to the root of the tree (see Fig. 4.5.) Note that, regarding the dependency of the transmission rate constraint described in Section 4.4, the expected distortion of each service class obtained from the QoS mapping mechanism is also dependent on each other. As seen from Fig. 4.6, the choices of QoS parameters from service class 1 influence the expected distortion of lower priority service classes. 117 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 4.6 Interaction betw een P rioritized V id eo and P riority N etw orks 4.6.1 Q oS B ou n d By the QoS bound, we mean a range of video quality requirements and transmission service capabilities, in which QoS param eters can be adjusted to cope with the time-varying wireless link quality. For example, the QoS bound of the video quality requirement at time unit t is defined in terms of video PSNR as T (t) = [PSNRL{t),PSNRu(t)], where P S N R ^ t ) and PSNRu(t) are the lower and upper bounds of PSNR at time t, respectively. The lower bound is determined by the minimum video quality that a mobile user can tolerate and the resolution of user’s display device. Given K service classes, service class i with buffer size BrnaX i provides the range of statistical QoS guarantee in terms of buffer overflow probability as 0 (f) = [eL{i),£u(i)], where £l(Q and £u(i) are the lower and upper bounds of the guaranteed buffer overflow probability by service class i, respectively. As discussed in Section 4.4.3, 118 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. the range of transmission rate constraint corresponding to the range of statistical QoS guarantee can be expressed as »,[/)]) where /q(#zyz,) and /q^ggc/) are the transmission rate constraints of service class i corresponding to £l {i) and £u{i), respectively. The range of guaranteed buffer overflow probabilities can be translated to its corresponding guaranteed delay bound as presented in Section 4.5. 4.6.2 Q oS A d a p ta tio n v ia V id e o -N e tw o r k In ter a ctio n Based on the notion of QoS bounds, we propose an adaptive algorithm th at employs the interaction between video applications and priority networks through the QoS mapping interface to select proper QoS parameters for video transmission. The algorithm is described step-by-step below and illustrated in Fig. 4.7. • Step 1: The video coding module sets up the QoS bound Y(i) in terms of the expected PSNR (or the expected distortion). Then, it sends the request for transmission (TxReq ) to the transmission module to set up the transmission process. • Step 2: Based on the knowledge of the available transmission resource and the current channel condition, the transmission module decides to accept or reject 119 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Q oS_Select TxJStatus Status TsReq Q o S T x QoS M apping M echanism T ransm ission Module Video C oding M odule Figure 4.7: Interaction between video applications and priority networks for QoS adaptation. the request from the video coding module. The decision of the transmission module (Status) is sent back to the video coding module. • Step 3: If the transmission module accepts the request from the video cod ing module, it offers a set of statical QoS guarantees for each priority class. Then, the QoS mapping interface translates the QoS provisioning of each ser vice class to the expected PSNR value as described in Sections 4.4 and 4.5. The QoS parameters of the network that provide the best PSNR and satisfy the video quality requirement T (t) will be chosen as the QoS parameters for video transmission. If there are no multiple QoS provisioning satisfying all constraints simultaneously. It implies th at the available transmission capabil ities cannot meet the video quality requirement of the video coding module under the current channel condition. Then, we go to Step 4. If all constraints are met, we go to Step 5 directly. 120 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. • Step 4: The transmission module requests the video coding module to adjust the video quality requirement. If the video coding module declines, then the transmission service cannot meet the application need and the whole process terminates. Otherwise, the video coding module adjusts the QoS bound {i.e. the PSNR range) and repeat the process in Step 3. • Step 5: The video coding module sends selected QoS parameters (QoS select) to the transmission module to set up QoS provisioning of each service class in the priority network. • Step 6 : The transmission module sends the acknowledgement to the video coding module after its QoS param eters are set up. • Step 7 : The prioritized video bitstream is upload to the priority network based on agreed QoS parameters. • Step 8 : When the change of the transm itting channel service-rate is detected during transmission, adaptation of QoS parameters of both the video applica tion and the priority network may be needed. Thus, we go to Step 3. Step 3 in the above algorithm contains some detailed procedures as described below. Let us define the set of statistical QoS guarantees of the priority network as T = {vi, .. ., }, where ry = {£*(1),.... £i{K)}, £i(j) € and Ng is the number of the set of statistical QoS guarantee. Furthermore, define £ = {rUl, ... , } as the 121 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. set of the total transmission rate constraint corresponding to the set of statistical QoS guarantees. We have (4 -28) »=i where fj,{ej{i)) is the transmission rate constraint corresponding to the case with packet loss equal to £j{i). Then, we have solve the following problem in Step 3. Problem Form ulation II: Given the current channel service-rate at time t equal to rchannei(t), we want to find a multiple QoS provisioning scheme for video transmission, v* from such that v* = arg max PSNR(t, vA, (4.29) V uj-esv subject to arg max PSNR(t, Vj) € T(f), (4.30) and r v * < r channel {t), (4 -8 f) where rv is the total transmission rate constraint corresponding to v* as derived in Eq. (4.28) and PSNR(t, Vj) is the expected PSNR obtained using the set of multiple QoS provisioning Vj. 122 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 4.7 E xperim ental R esu lts In this section, we present the simulation results of the proposed dynamic QoS management for prioritized video transmission over time-varying and non-stationary wireless channels. The non-stationary behavior of wireless channels is simulated by randomly changing the normalized Doppler frequency and average power. The normalized Doppler frequency is chosen from the set of {10-3 , 5T0- 3,1(D2} reflecting the time-varying mobile speed while the average SNR of the received signal varies from 10 to 20 dB. 4.7.1 T ran sm ission R a te C on strain t We conducted experiments to study the derived transmission rate constraint of the priority network in this section. In particular, we adopt a time-varying service-rate channel modeled by the Markov process from the work in Chapter 3. Consider time-varying service rate channel c(t) that can be well modeled by the Markov process with N states. The closed form of effective capacity and effective bandwidth for transmission data rate S(t) can be written as [25, 59] and In (tt{ee‘APstate}) es(t){0i) 9, 123 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. where Hs(t)(&i) and es(t){Oi) are the effective capacity and the effective bandwidth of S(t), respectively, Pstate is the transitional probability m atrix of the discrete Markov model and is the spectral radius of m atrix U. Let T* = {rq,..., rjy} be the vector of service rates at various discrete Markov states, and A = diag{~r*) be the diagonal m atrix with its diagonal components corresponding to those in T*. Furthermore, the QoS exponent of the guaranteed packet loss probability can be expressed as In {P(B>B max ) where £ is the probability for the buffer to be empty. First, let us consider two service classes with absolute priority scheduling for packet transmission. The first class has a higher priority than the second class. The packet size is 200 b3ff.es. The expected service rate of the wireless channel is 380 kbps. As seen in Fig. 4.8, the transmission rate constraint of the first service class (i.e. the high priority class) computed from the closed form of effective bandwidth and effective capacity and those obtained from the simulation are close to each other over a wide range of packet loss probabilities. The lower the packet loss probability requirement, the less reliable the transm itted data rate. Simulation results given in Fig. 4.8 also dem onstrate the buffer size effect on the transmission rate constraint. The larger the buffer size, the more data rate we can transm it under the same packet loss probability requirement. 124 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Simulation: Buffer size 250 packets - x - Model: Buffer size 250 packets Simulation: Buffer size 500 packets - 0 - Model: Buffer size 500 packets 1 0 - * l-----------' ----------- 1 ----------- 1 -----------1 ----------- 1 ----------- 1 ----------- 1 ---- 0 50 100 150 200 250 300 350 Transm ission R ate Constraint (kbps/s) Figure 4.8: The transmission rate constraint of a wireless channel with two service classes, which is computed from the discrete Markov channel model corresponding to the normalized Doppler frequency = 1CT2 and average power = 16 dB and the buffer sizes are chosen to be 250 and 500 packets. Based on the transmission rate constraint derived in Section 4.4.3, Fig. 4.9 shows the transmission rate constraint of service class 2 (i.e. the low priority class) over a wide range of guaranteed packet loss probability requirement. As shown in these simulation results, the transmission rate constraint of service class 2 with a buffer size equal to 250 packets is dependent on how much service class 1 occupies the wireless link. The transmission rate constraint of service class 2 with a lower QoS requirement of service class 1 (with guaranteed packet loss probability = 10" 2 and transmission rate constraint = 109 kbps) can provide a higher transmission rate than 125 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. that with a higher QoS requirement of service class 1 (with guaranteed packet loss probability = 1CT4 and transmission rate constraint = 54.5 kbps). Transm ission R ate Constraint of C lass 1 = 54.5 kbps - x - Transm ission R ate Constraint of C lass 1 = 1 0 9 kbps 100 150 Transm ission R ate Constraint of C lass 2 (kbps/s) 200 250 Figure 4.9: The transmission rate constraint of a wireless channel with two service classes and a buffer size of 250 packets based on absolute priority scheduling when the packet loss rate requirement of class 1 is equal to 10~~2 and 1 0~4. The transmission rate constraint of service class 1 in different wireless channel environments is compared below. The maximum buffer size is set at 500 packets. First, the effect of different normalized Doppler frequencies to the transmission rate constraint is shown in Fig. 4.10. If the channel changes slowly (with a low normalized Doppler frequency), the transmission rate constraint is lower than that with a higher normalized Doppler frequency given the same guaranteed buffer overflow probability. 126 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Normalized Doppler Frequency = 10 2 Normalized Doppler Frequency = 5* 10~3 250 300 200 (kbps/s) 100 Transm ission R ate Constraint (kbps/s) 150 50 Figure 4.10: The transmission rate constraint of a wireless channel of service class 1 computed from the discrete Markov channel model with the normalized Doppler frequencies 10" 2 and 5 • 1(D3, where the average power is equal to 16 dB. The effect of the average power on the transmission rate constraint is shown in Fig. 4.11. As shown in the figure, given the statistical QoS guarantee, the transmis sion rate constraint can be increased by enhancing the average power transmission. Traces of the transmission rate constraint of two service classes under an environ ment with time-varying power and normalized Doppler frequencies are shown in Fig. 4.12. 127 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Average pow er = 16 dB Average pow er - 12 dB 250 300 200 Constraint(kbps/s) 100 Transm ission Rate Constraint(kbps/s) 150 50 Figure 4.11: The transmission rate constraint of the wireless channel for service class 1 computed from the discrete Markov channel model with the average power equal to 16 dB and 12 dB. where the normalized Doppler frequency frequency is 10~2. 4.7.2 A d a p tiv e V id e o T ran sm ission w ith D y n a m ic Q oS A d ju stm e n t To evaluate the performance of dynamic QoS adjustment for aptive scalable video transmission, three priority classes with absolute priority scheduling are consid ered. The guaranteed buffer overflow probabilities are [10“ 5.10~3], [10~4,10” 2], and [10~2,2 ■ 10"1] for priority classes 1, 2 , and 3. respectively. The guaranteed proba bility of delay violation corresponding to each priority class can be derived from the 128 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 1 0 0 Service C la ss 1 © © to 1 0 0 30 T Service cla ss 2 i/i C L £ } © 25 CO 20 10 40 50 60 70 80 90 1 0 0 0 20 30 100 x 1 0 ' a 1 0 5 100 Interval Figure 4.12: Traces of the transmission rate constraint with time-varying power and normalized Doppler frequency based on absolute priority scheduling, where service classes 1 and 2 provide the buffer overflow probability guarantees at 1 0 ~ 4 and 10~3, respectively guaranteed buffer overflow probability as described in Section 4.5. The transmis sion rate constraint of each priority class can be translated from the corresponding set of QoS guarantees and used in the QoS mapping mechanism. Finally, the QoS mapping mechanism proposed in Section 4.5 is used as a QoS interface between the video and the transmission modules. The GIF Foreman sequence is encoded by the PFGS scalable video codec with the target bit rate of the base layer equal to 100 kbps. The frame rate is 10 frames/s and each GOP contains 10 frames. The time-varying and non-stationary wireless channel is simulated by varying the average power and the normalized Doppler frequency and modeled by the discrete Mavkov chain as described before. The channel is assumed 129 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. to be stationary during one GOP interval. The buffer size is set to be 1000 packets with the packet size equal to 200 bytes. We compare three video transmission systems with different characteristics: • System No. 1: no QoS interaction and adaptation with the expected PSNR equal to 29 dB. • System No. 2 : no QoS interaction but with QoS adaptation system with the expected PSNR equal to 32 dB. • System No. 3: with both QoS interaction and QoS adaptation with the PSNR value in the range of [26, 37] dB. The guaranteed multiple QoS provisioning of the first transmission system is fixed at ICR4, 10”3, and 0.1 for the three priority services, respectively, while the QoS provisioning of the second and third systems are adaptively changed based on channel conditions. However, QoS interaction for adjusting the video requirement in System no. 2 is not applied. The simulation results are given in Fig. First, let us compare Systems 1 and 2. Due to the low video quality requirement of System 1, System 1 provides more consistent video service than System 2, where the interrupted video service durations of Systems 1 and 2 are in the range of video frame 70-80 and 70-100, respectively (especially when the average received SNR is low). However, with QoS adaptation for video transmission, System 2 provides better end-to-end video quality outside the 130 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. interval of interrupted video service. W ith both QoS adaptation and interaction, the adaptive video transmission system (i.e. System 3) provides more consistent video service and enhanced video quality than the other two systems. 4.8 C onclusion In this chapter, we proposed a dynamic QoS management framework for adaptive prioritized video transmission over a priority network with multiple service classes. The main contributions of this work include the following: • The proposal of an adaptive QoS service model that allows QoS parameters to be adaptively adjusted according to the time-varying wireless channel con dition, • The establishment of an interaction mechanism between between the priority network and video applications to provide proper QoS section, • The development of resource allocation schemes to assign resources based on the QoS guarantee for each service class under the time-varying wireless chan nel. The proposed QoS mapping algorithm th at maps the statistical QoS guaran tees of the priority network to the corresponding service class with different video quality requirements, is utilized to achieve adaptive QoS. Our simulation results 131 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. demonstrated th at the proposed dynamic QoS management system can provide con sistent video service and enhanced end-to-end video quality over time-varying and non-stationary wireless channels. 132 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. interactive and adaptive Q oS system -it- N on-interactive an d n o n -a d a p ta tio n Q oS sy stem + N on-interactive an d adaptation Q oS system > 30 Interrupted V ideo S e 1 of S y s t e r t t ^ Interrupted V ideo S eivice of S y s te f tK l^ ___ 100 70 F ram e N um ber 20 T fc 16 100 0.01 = 0.008 fe 0.006 0.004 0.002 90 100 Fram e Number Figure 4.13: The Y-PSNR comparison of three video transmission systems under a non-stationary wireless environment with the time-varying power and speed. 133 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. C hapter 5 Prioritized V ideo T ransm ission w ith A d ap tive FEC P rotection 5.1 Introduction In spite of the challenge from time-varying and highly noisy characteristics of wire less channels, multimedia transmission over wireless channels has become reality with the advent of 3G wireless communication systems. Existing solutions address this challenge from both the source and the channel sides. From the source side, error resilience to achieve graceful degradation is considered by schemes such as data par titioning (DP) or multiple description (MD). From the channel side, wireless media can be delivered more effectively by utilizing available bandwidth/power resources and exploiting the estimated status of channels. Several adaptive feedback systems with rate adaptation, modulation, power control, handoff management, and antenna diversity have been studied [34, 5]. It is obvious that techniques from both source 134 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. and channel sides should be integrated and dynamically coordinated to achieve the best performance in response to network dynamics when we deal with the most severe wireless channel conditions. Each compressed video packet is assumed to convey its im portance information in terms of loss/ delay with respect to others, which is called the Relative Priority Index (RPI), in this work. The packet-based priority assignment can be regarded as a fine-grain prioritization scheme as compared to conventional layer and class- based approaches. The priority assigned to a video packet would be best if it can accurately represent its loss effect on visual quality. For example, one may assign the priority based on an error corruption model [53], which estimates the packet loss effect by tracking error propagation. Given the source priority in RPI, the estimated channel condition, and allocated channel resource, we develop a proactive forward error correction (FEC) method with code adaptation in this research. An optimized concatenated-FEC, composed by RS (Reed-Solomon) and RCPC (rate- compatible punctured convolutional) codes, is utilized to protect video packets of varying priorities and sizes. W ith the RS/RCPC-concatenated FEC, we develop an algorithm that uses dynamic programming to find the best m atch between the source priority and instantaneous channel conditions. Generally speaking, the channel adaptation mechanism depends on the accuracy of the estimated channel status as well as the underlying time-scale of the estimate [41]. In this work, adaptation is performed based on long-term fading estimation 135 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. that reflects an a-verage of the short-term fading effects over a reasonable window size. The proposed scheme depends 011 reliability and timely availability (i.e. delay) of feedbacks. A packetized transmission scenario is set up by following the ITU-T H.324M wire less video framework. In the experiment, error resilient H.263+ video multiplexed, by a variant of H.223 Annex B is adaptively protected. The overall gain is demon strated by simulated end-to-end video quality with ‘GOB (group of block)’ packets. To achieve realistic joint source-channel adaptation, special attention is paid to the channel status feedback in terms of accuracy and delay, the concatenated code rate selection, and involved packetization efficiency. This chapter is organized as follows. The system configuration, including the video codec and wireless channel param eters in use, is described in Section 5.2. Sec tion 5.3 presents priority assignment for each video packet based on the RPI concept. Section 5.4 illustrates the wireless mobile channel model used in simulations as well as two concatenated adaptive filter utilized to estimate channel state information. Section 5.5 shows the optimal selection procedure of concatenated code parameters, leading to the formulation of an optimization problem. The resulting dynamic pro gramming solution is given in Section 5.6. Simulation results are shown in Section 5.7. Conclusion and future work are given in Section 5.8. 136 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 5.2 System C onfiguration and V ideo P acketization The ITU-T H.324M wireless video delivery framework is depicted in Fig. 5.1. At the sender, bit stream of H.263+ video, ’G OB’ (Group of block)~based with error resilience and compression efficiency options (Annexes D, F, I, J, T plus random intra refresh with Fj = 5) is adopted [19]. Each ’GOB’ contains fixed number of MB (Macro Block) in the packet. Each packet is assigned a source priority index denoted by RPI (relative priority index). It is then multiplexed by a variant of H.223 Annex B. The header, the synchronous field, and the control field are strongly protected with the BCH (Bose-Chaudhuri-Hocquenghem) code. The cyclic redundancy code (CRC) is calculated to check its payload and appended for error detection. Our scheme is aligned with the H.223 MUX scheme except that the channel packet size is fixed in the latter. Based on RPI and the size of each video packet along with the feedback channel state information, unequal error protection (UEP) is conducted with the RS/RCPC concatenated code. Finally, the resulting channel packet is modulated and transm itted to the underlying wireless channel. At the receiver, the received signal is decoded by the ML (maximum likelihood) - based scheme and the long-term fading param eter is estimated. The calculated fad ing param eter is then feed back via a reliable channel to the sender. Decoded packets after the de-multiplexing stage are classified into three types: clean, corrupted (CRC check failure) and fatal (unrecoverable error in the header, the synchronization or 137 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Video Input Available Bit rate Video output Product Code Channel status FEC inner code FEC outer code Decoder Buffer FEC inner code FEC outer code Channel' Wireless Channel Encoder buffer H.263+ Decoder Adjustable H.263+ Encoder Corruption Model & Priority Assignment Optimized Adaptation Code Ratio With Uequal Error Protection Side status Information (optional) Figure 5.1: A wireless video system with FEC-based rate adaptation via channel state feedback. the control fields of multiplexing packet). Finally, to provide the end-to-end perfor mance in both subjective and objective measures, a decoder capable of handling the corrupted video stream is used for video decoding. As depicted in Fig. 5.2, source video packets of a variable size from a group of video frames are re-organized into a group of channel packets of a fixed size, which is called the channel slot. Since each channel slot consists of fixed-size channel packets, the payload and the redundancy should be negotiated to m aintain a fixed total 138 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. amount. To be more specific, the level of protection is chosen to give the maximum protection to the video packet stream with both the bandwidth and the packetization constraints. In order to provide the best adaptation under these constraints, a solution based on dynamic programming is devised and its gain is explored in the following sections. Priority Assignment RPI calculation ♦ Different size GOB 3 ' Joint soutcs channel design for packetinng video packet for wireless transmission_ _ _ GOB packet Channel packet overhead Header Syn CF Packetization scheme W ireless channel variati RS code ratio T to RS/RCPC RCPC code ratio Variation of size relying on EPI value and size of GOB, CRC as m il as RS parity GOB packet and CRC w i t RS parity RCPC parity GOB packet and CRC with RS parity RCPC parity GOB packet and CRC with RS parity RCPC parity Changeable RCPC parity to fit with feed size of channel packet Figure 5.2: Multiplexing and packetization of video packets into a group of channel packets. 5.3 C orruption M odel for P riority A ssignm ent To implement unequal error protection (UEP), we would like to assign RPI to each video packet under the assumption th at it can reflect the loss effect to the received 139 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. video well. The corruption model is the tool used to estimate the impact of the video packet loss to the overall received video quality [53]. This impact can be measured by the induced distortion (e.g., in terms of mean square error, MSE). For a motion-compensated video coder, MB-based corruption can be modeled by taking into account error concealment, tem poral dependency, and loop filtering. By assuming th at the loss impact of each MB is independent, the total impact of one MB loss is expressed as the sum of the initial error and the propagation error energy. In addition, the propagation error of MBs in subsequent frames should be weighted. If Intra-MB refresh is used, error propagation will be confined and a pre-defined number of frames, denoted by M , is often sufficient to estimate the total impact of a MB loss. As a result, the total energy of errors due to a MB loss in frame n can be expressed as M N 0-2 = a l + D (5.1) m — 1 j —1 where a\ is the initial error due to a MB loss (which depends on error concealment), wntj(m ,j) and stand for the propagation weight and effect to the jth MB (among N MB’ s) of the m th frame, respectively. In addition, cr^(m, j) = cr^/(l + 7 mj). where 7 mj is a param eter governed by loop filtering. The fundamental notion behind RPI is the decoupling of compression details from the network adaptation task. By assigning RPI to each packet in an appropriate manner, we attem pt to accommodate the demand of each packet to achieve best end-to-end video experience via network adaptation. Since the video codec has 140 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. several options to trade the compression efficiency for flexible delay manipulation, error resilience, and network friendliness, the proposed coordination framework has to provide a simplified interaction process between the video encoder and the target adaptive protection. Thus, the corruption model is utilized for the R PI assignment to each packet. As an example, the resulting RPI for the ‘Silent’ sequence and ’Glasgow’ sequence are depicted in Fig. 5.3. The example of R PI statistics can be shown in Table 5.1. The mean and standard deviation of RPI illustrate the correspondence between the video characteristics and its error propagation effect. 1000 £ 500 I 24S0 2500 G 0 8 number GOB number '’Silent’ Figure 5.3: The plot of RPI value of each GOB packet for the ‘Silent Voice’ sequence and ’Glasgow sequence’. 5.4 W ireless C hannel M odel The wireless channel model used to simulate the real-world situation includes the short-term fading, the long-term fading (shadowing), and the path loss as in Fig. 5.5. 141 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Table 5.1: The mean and the standard deviation of RPI for different sequences. Video Sequence Mean Standard Deviation Foreman 149.3723 266.5484 Glasgow 377.057 1509 Container 37.9261 146.2463 Hall 21.1941 35.831 News 61.7778 203.1266 Silent 43.5501 67.6116 The m ulti-path phenomenon generates an amplitude variation of the transm itted channel signal which leads to the short-term fading. The path loss reflects the degradation of the signal strength over a distance. Fluctuation of the expected power level based on the path loss model is called the long-term fading effect. For simplicity, it is assumed that power degradation due to the path loss is perfectly compensated by the power control method, and the SNR operational point is maintained throughout the wireless communication process. In this chapter, all simulations have been conducted in an urban micro-cell wire less environment. The carrier frequency f c is set to 2 GHz and the mobile unit moves at a moderate velocity. Based on these conditions, we can determine the wavelength A to be 0.15m since vc = f c/ A, where vc is the light speed. The normalized Doppler frequency used to construct the short-term fading is set to 1 0 -3 while the short-term fading is modelled as the Rayleigh flat fading channel. Thus, for BPSK-modulated transmission, the short-term fading model can be written as rk = ak -sk + nk, (5.2) 142 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 1500 o 1 0 0 0 10000 *? 8000 350 6000 300 250 4000 200 150 2000 100 RPI Packet size(byte) Figure 5.4: RPI and the packet size distribution of the ’Glasgow’ sequence. where r*, is the received signal symbol k at the decoder, a& is the variation of the Rayleigh flat fading amplitude (i.e., the short-term fading param eter), s* , is the BPSK-modulated signal, and Uk is the additive white Gaussian noise (AWGN), respectively. Also, the BPSK-modulated signal is assumed to have a unit power for simplicity. The long-term fading is obtained by averaging the short-term fading SNR (i.e., Ek/No derived from (5.2), where Ek is the received energy per symbol at time k) with a reasonable window size such as 10 ~ 20A [35, 127]. The long-term channel SNR 143 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Received Signal Power path loss Rayleigh fading log (distance) Figure 5.5: The general wireless channel model consists of path loss, long term fading(shadowing) and flat fading model(short term fading) variation SdB(i) in dB at the ith window is of the normal distribution characterized by the first-order Markov process SdB{i + 1) = P • SdB(i) + cr • jV(0,1), (5.3) where p is a spatial correlation between shadowing samples, a is the standard devia tion of the log normal, and iV(0,1) is the Gaussian random variable with zero mean and unit variance, respectively. In general, the long-term fading in a micro-cell has low to moderate deviations. Typically, we choose the standard deviation to be 4.3 dB as in [35, 127]. W ith the assumed velocity of a mobile unit and the time-scale 144 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. of the long-term fading, which corresponds to 20A case, we get a spatial correlation of 0.8348 to simulate the shadowing model. Also, in this work, we set the averaging window length according to one channel slot that consists of 5 channel packets. 5.4.1 E stim a tin g an d p red ictin g ch an n el sta tu s To obtain the required channel state information for feedback, we utilize concate nated structure of two filters. One is used to estimate short-term fading while another predicts long-term fading parameter. To capture the rapidly changing short-term fading parameter, Kalman filter is chosen in our approach to take advantage of its fast convergence and lower estimation error as in [45]. For simplicity, the BPSK-modulated signal is assumed to have perfect noise power estimation. Then the output of Kalman filter is described as dfc+i Q'k+ijk T A fc+i ' ^k+ i\ki (^'^) where ak+i is the estimated channel param eter at time fc + 1, a,k+i\k is the predicted one at time k + 1 from past fc, Kk+\ is the Kalman gain at time fc + 1, and Zk+\\k stands for the innovation term (i.e., residue between received and predicted signal for BPSK m odulation), respectively. 145 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. The variation process of the averaged power of the short-term fading, which gives the long-term fading S(i), can be tracked by the adaptive LMS filtering [45]. In LMS. weights of FIR filter will be adapted to minimize the mean square error by W (i + 1) = W(i) -2/i-e(«) • S(i), (5.5) where p is the convergence control param eter, W (i + 1) = {po(*); , Phi})} is the adaptive filter coefficient set, S (i) = {S(i), S(i — 1), • • ■ , S(i — L)} is the set of previous channel estimates, and e(i) = S (i + fc) — W T(i + 1) • S (i) is the error from estimated value, respectively. For now, we use the first-order adaptive filter to predict the future variation of long-term fading. Hence, in the case of f c — 1 units feedback delay, the estimated SNR is S(i + fc) = po(i) ■ S(i), where S(i + fc) is the prediction for long-term fading at (i + k)ih interval. The accuracy in estimation and prediction in the different range of channel con ditions can be shown in Fig. 5.7. 5.5 C oncatenated C ode U tilization The concatenated code is a specific method to constructing long codes from shorter codes and the code th at have capability in dealing with both random and burst error at the same time. A simple concatenated code is formed from two codes (ni,fci) binary code C\ and (n^, £2) non binary code C2 with symbol from GF{2k) as shown 146 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Average estimated short term to get long term fading Predict long term fading Received Signal Figure 5.6: Adaptive filter structure for estimating and predicting channel state information in Fig. 5.8. The symbol from non binary code is formed by their corresponding bytes of h binary symbols. The components codes C\ and C2 are called inner codes and outer codes, respectively. The concatenated code has a capability in correcting both random and burst errors. The concatenated code used in this research consists of RCPC and shorten RS codes for the adaptation purpose. In the following section, definitions of RCPC and RS codes are given. 5.5.1 R ate-C o m p atib le P u n c tu re d C onvolutional (R C P C ) C od e RCPC is used as inner code component of concatenated code. Based on puncture table in H.223 Annex C., RCPC can flexible change the coding rate from mother code 147 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 3.5 5 2.5 0 ) 3 o in n < 0.5 24 O peration SNR (dB) Figure 5.7: Absolute error in percent between long term fading param eter and pre dicted one by using two concatenated filter. rate. Also with maximum likelihood decoder, soft reliability (e.g. metric or other information) from inner decoder may be optionally send to outer code in overall improvement of decoding data. The bit error rate of RCPC in the Rayleigh flat fading channel can be shown as in Fig. 5.9. In the case of the Rayleigh fading with the AWGN environment, the performance of Viterbi decoding with the maximum likelihood sequence estimation (MLSE) for 148 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Encoder Decoder Channel Inner encoder Figure 5.8: The concatenated code structure used in the proposed communication system. RCPC can be bounded [42]. W ith estimated long-term fading param eter < S , the bit error probability P r c p c bounded as 1 C O P r c p c ^ R c p c , S) < — • YL cdPd(rRcpc, S), (5.6) d— dfree where Pd(rRCp c , S ) < ( ■ ------- )d, (5.7) * + * 5 • r RCp c and where Pd is the probability that a weighted path has a higher metric than all zero-path (assumed to be the correct one). 5.5.2 R eed -S o lo m o n (R S ) C o d e and C o n ca te n a te d C od e The Reed-Solomon (RS) code is a non-binary block code with codewords of a fixed length [119]. It is a maximum-distance separable (MDS) code. Let us use RS ( N r s , K r s ) to denote a RS code of code-ratio rRs = K r s / N r s , which consists of K r s 149 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 10 uncoded 10'1 -i 10' 8/9 -3 ,4 10' 8/32 - s 10' 25 30 15 5 10 20 0 SN R (dB) Figure 5.9: The performance of RCPC w ith different code ratios in the Rayleigh flat fading channel. information symbols and N RS — K RS parity symbols. This RS code can correct up - ) ;0 |^ b sR -irs) j S ymbol errors. In our work, a new version of the RS code called the shortened RS code is utilized since it supports a shortened length while keeping the same MDS property. By combining the inner code (RCPC) and the outer code (RS), we are led to serially concatenated code. The concatenated code has the potential to serve as a real-world implementation of the long block-code mentioned by Shannon. In addi tion, the concatenated code is suitable for joint source-channel coding design as in our case. By assuming th a t the BSC channel effect is obtained after RCPC decoding 150 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. (with perfect interleaving for fading compensation), the upper bound to codeword (i.e., packet) error probability Prs+rcpc f°r given N rs constraint can be expressed as N r s PRs\RCpc(rRS:r«crc,S)< £ NRsi - P'M - (1 - (5.8) i= ^ -B . g _ hzra£) j + 1 where Pm = 1 — (1 — PrcpcY is the probability of the RS symbol error and b is the number of bits per RS symbol. The resulting probability of the codeword error based on the concatenated code combination is depicted in Fig. 5.10. _ 1 0 t i 10' Product C ode R ate 8/14 - e - Product C ode R ate 8/16 - 0 - Product C ode R ate 8/18 Product C ode R ate 8/20 h — Product C ode R ate 8/24 0.3 0.4 0.2 0.5 0.7 0.9 Inner code ratio Figure 5.10: The packet error probability with different inner code ratios and total product code ratios with channel SNR equal to 7 dB and N rs equal to 50 bytes. 151 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 5.5.3 C o n ca te n a te d C o d e R a te S e lec tio n To figure out the best concatenated code in terms of different code ratio combination ( r *RCPC i r *R.s)i the best ratio combination is searched from the concatenated set of the RCPC code-ratio set T rcp c and the RS code-ratio set r Rs- The best ratio com bination, which minimizes the expected packet error probability, can be expressed as m in. Prs+rcpc(rRS, r RCp c,S). (5.9) ' R C P C ’r R S A practical algorithm to find the solution of (5.9) is given below. • Initialization. Initialize the minimum probability P ^ n of the packet error at a large value. Also, initialize r * RCPC and r * RS. • Step 1. According to the given overall product code-ratio and the channel state information, set rRc p c and calculate the upper bound PRc p c °f the bit error probability. • Step 2. Apply the size constraint to compute the RS code-ratio rRs from selected r Rcpc- Then, calculate the upper bound of packet error probability denoted by PRs+RCPcirR S ^ Rcpc,<S)- • Step 3. If minimum packet error probability P^fn is larger than PRs+rcpc(tRS, r RCp c ,S ) , update P £ g , r * RCPC and r * RS by Prs+rcpc^ R S ^ rcpc^ ) , rRCPC, and rRs, respectively. 152 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Step 4. Go back to Step 1 until all possible RCPC code-ratios in set T r c pc are searched. 5.6 O ptim ized A d ap tation M echanism 5.6.1 P ro b le m F orm u lation The averaged distortion of received video can be expressed as encoding transm ission 3 (5.10) where Denc0(ang is the averaged distortion from the encoding process and Dtransmission is the averaged transmission distortion, respectively. Since we are interested only in the loss impact due to the transmission error, the fixed encoding distortion is not considered further. The assumption which discards the encoding distortion in framework comes from that fact th at video encoder tries to keep compressed bit rate to be constant. Hence by average, the encoding distortion becomes approximately constant. The variable Dtransmission can be derived from the corruption model with the independent loss effect of each packet as N t o ta l transmission E R P I { ‘) - P R i+R C p c M > ) , r Rc p c (i).S { ,)) (5.11) ■ i= 1 153 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. where N totai is the whole number of packets and R P I (f) is the priority assignment reflecting error propagation due to loss. Given the size of the video packet and the predicted channel state, several combinations of RCPC and RS code ratios are possible. W ith the expected video distortion of (5.11), the optimal product code selection {(i-rcpc: rRs) | (r^CPC(i), r^g(i)), i = 1,..., N totai} can be formulated as follows. Find the optimal product code selection, {(*r c p c > r Rs)}’ such that ^ t o t a l mill ^tra n sm issio n mill ^ ^ FLPI{%) ^ R S + R C P C ; ^R C P C (^) • i rRCPCifRS rR.Cpc,rRs —f i —I (5.12) subject to the total bit budget constraint expressed by N t o ta l y i [5(f) ■ 1 / t r s {%) ■ l / r R C P c ( i ) \ < B total, (5.13) t=i where 6 (f) is the size of packet i in bits and Btotai is the total bit budget 1. From set up formulation above,total cost function for whole video packets based on Lagrange multiplier can be written up as ^ t o t a l J{A) = D transm ission + A * ^ [ 6(f) • 1 / rRs{i) ■ l / r RCpc{i)\ (5-14) i = 1 JThe source packet size b ( i ) is slightly different from the video packet size since it incorporates other fields for multiplexing. Also, the CRC field size is excluded in this formulation, which is however included in the actual simulation. 154 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. where J(A) is the total cost function and A is the Lagrange multiplier. Even though independent transmission loss is assumed in (5.12), dependency among packets is implicitly included through RPI assignment, which can reflect the multiple-packet losses as well as its own loss. However, the solution to (5.12) is not straightforward, since the product code selection {(rR cpc, Trs) | (rR cpc (i), i~Rs(i))-, i = Ntotai} is affecting each other. Under the given size constraint, the overall optimization may be attem pted if the channel state is known in advance. Since the knowledge on the future channel state is impossible, the above problem has to be converted to optimized adaptation over some predefined interval (either fixed or variable). Under the circuit-oriented wireless transmission (e.g., H.324M), multiplexed source packets of a variable size have to be accommodated by fixed-size channel packets. The straightforward way to enforce this channel-packet size constraint is to chop the concatenated source packets into fixed channel packets, which however renders the coordinated protection impractical. T hat is, since there is no established alignment between the source packet size and the channel packet size, it is difficult to assign RPI to channel packets. Note however th at this scheme is the most efficient one in terms of packetization efficiency. The other extreme is to align each source packet to one or more channel packets and to protect the packet based on its RPI. Although network adaptation can be best coordinated in the sense of UEP efficiency, we suffer 155 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. from the loss in packetization efficiency. Thus, an interval should be predefined to control the tradeoff of packetizatoin efficiency and UEP efficiency . In our previous work [50], the source packet has not been packed across channel slots to enable flexible packetization while keeping it fixed and simple. The resulting fixed interval (i.e., channel slot interval) corresponds to the time-scale of the channel state change and hence the time-scale of adaptive protection. However, although it is helpful in coordinating adaptive error protection, the channel slot scheme still suffers some level inefficiency in terms of packetization (especially when the number of channel packets is small). Thus, in this work, we use Ngroup source packets as the interval for packet alignment, which will occupy a variable number of channel packets. Note that, N group may or may not correspond to the number of source packets in a frame of video. This slightly improves packetization efficiency while maintaining the proposed channel-based adaptation framework. This consideration imposes another constraint expressed as 0 + 1 ) ' N g TQUp [6(f) • l / r R S {i) • l l r Rc p c { z)] Tlj ■ -B p a c fc e Z ) (5.15) i— j ' Ng roup 1 where Y^2°{a^ N 3rouv nj ■ Bpacket = Btotai and Bpacket is the bit for a channel packet and rij denotes the used channel packet number. One implicitly im portant point in optimization is the dependency problem in packetization of protected video source packet to channel packet in time channel variation environment e.g. previous concatenated code assignment will effect the 156 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. following one. To solve set up optimization cost function in (5.6.1), dynamic pro gramming adaptation based will be introduced in the next section. 5.6.2 D y n a m ic P ro g ra m m in g S o lu tio n to A d a p ta tio n and P a ck etiza tio n Optimal Trade off between size and propagation Distortion inside the group of Video packet propagation,! propagation^ — j . — P Size Size Optimal path corresponding ' to minimum distortion R1 R2 m R4 R1 R2 R3 R4 Figure 5.11: Adjustable video packet based Dynamic programming for concatenated code rate allocation. Dynamic programming provides an optimal solution to the constrained opti mization problem given in (5.13) and (5.15). However, considering the complexity 157 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. of dynamic programming and the lack of precise knowledge of future channel state information, we divide time into intervals (called the window) and seek to achieve a feasible sub-optimal solution. T hat is, N window packets are divided into groups of N group packets and optimized via dynamic programming as shown in Fig. 5.12. This sub-optimal solution requires the determination of the bit budget for each win dow from the total budget B totai. In our approach, the bit rate allocation for these Nwindow packets is based on their ‘Aggregated R P I’ (i.e. summation of all R P I’s in each optimization window). The budget B window for each window w is bounded by both the bandwidth and the packetization constraints, which can be calculated by the required transmission time for each window and the size of all packets included. ARPI(w) ® w iadow (^) ARPI(w+l) ® w in d o w (^ "^ l) ARPI(w+2) B, window (w+2) Window wt h Window (w+l)t h Window (w+2)t h ARPI(w): Aggregate RPI at window wth B w in d o w (w) budget allocate at window wt h Figure 5.12: Budget allocation for each window based on the aggregated RPI with bandwidth and packetization constraints. 158 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. The structure of dynamic programming search is given in Fig. 5.11. In the following, each stage of dynamic programming, denoted by index j, represents each group of video packets to be optimized in each independent window (not a sliding one yet). Each branch connecting nodes represents the determined best choice of the concatenated code ratio for all packets in each group, which is searched as described in Section 3.5. Note th at if this search is not determined independently, the overall dynamic programming scheme will become nested and too complicated. Then, each node in dynamic programming contains used bits and the corresponding distortion. The used bits and distortion for the entire transmission period can be obtained through summation of these windows along the path. At the jth stage corresponding to the jth group, the amount of used bits for the jth group can be written as N g r o u p UB(j) = ^ b(j, k) • 1 / r RCPC(j, k) • 1 /rRS(j, k), (5.16) fc=l where 6(j, k),rjzcpc(j, k), and rRs{j, k) are the number of source bits, the inner code ratio, and the outer code-ratio for the kth packet of the jt h group, respectively. The corresponding transmission distortion of the jth group is D transmission(j) = E f r r , 7 > RPI{j, k) ■ Pr s+rcpc^ R s U, k),rRCpc{j, k),S{j, k)), where S{j, k) is the es tim ated channel state for the kth packet of the jth group. At each transition stage, the accumulated used bit till the jth group should not violate the total bit budget 159 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. constraint of this optimization window, i.e. Yjl=oUB(i) < B w in(iow The accumu lated distortion at the jth node is also calculated in the same way. W hen these paths are merged, the path providing the smallest accumulated distortion will be chosen. Algorithm in optimization can be stated step by step as following. • Initialization. Initialize the tree for the optimization procedure. Set the overall used bit U B and the overall distortion to 0. Divide N W in tiow packets into groups of N group packets. Estim ate channel states for the optimization window. Proceed to Step 1-1 with the first group of packets. • Step 1-1. Choose the combined concatenated code-ratio from the available set. If all code-ratio combinations are searched, go to Step 1-5. • Step 1-2 . Calculate the average RPI and the average size of packets in current group. Also calculate the number of channel packets available for packing in current channel slot. • Step 1-3. Use the averaged RPI and size to find the best concatenated code- ratio combination (r * RCPC) r * RS). • Step 1-4. Calculate the required bit budget from ‘aggregated R P I’ for transm is sion and check whether the resulting packet can be accommodated in current channel slot. If not, divide packets across channel slot and repeat from Step 1-2 . If all packets in current group are accommodated, go to Step 1-1 and proceed with next combined concatenated code-ratio. 160 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. • Step 1-5. For each combined concatenated code-ratio, create a node in dy namic programming tree. Allocate redundancy to each source packet based on its RPI and size as well as channel state. Calculate the expected distor tion Dtransrmssion(j) and the overall used bit UB(j) for each node. If UB(j) of a concatenated code-ratio combination exceeds the bit budget, prune the corresponding node. • Step 2 . Repeat from Step 1-1 to Step 1-5 iteratively for all groups in the optimization window. If done, go to Step 3. • Step 3. Search the final node for the minimum distortion and trace back the tree path to get the set of the optimal concatenated code-ratio combination. 5.7 Sim ulation R esu lts The optimized FEC allocation scheme proposed in this work is simulated by employ ing variable-size H.263+ video packets with error resilience options. To reduce the implementational complexity of dynamic programming, the combined concatenated code-ratios are confined to the following possible set {8/10,8/12,8/14,8/16, 8/20, 8/24} with RCPC codes in the code-ratio set {8/9 ,8 /1 0 ,8 / 1 2, 8/14, 8/16,8/18, 8/20, 8/22,8/24,8/26,8/28,8/32} which is obtained by using the RCPC puncture table based on H.223 Annex C with mother code rate 8/32. To evaluate the performance 161 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 200 200 J 200 200 Figure 5.13: The sample trace for selected RCPC (inner) and RS (outer) code ratios with the SNR operation point equal to lOdB and video transmission rate 41 kbps. of the proposed adaptive FEC video transmission, 2000 frames of the low activ ity QCIF ‘Silent Voice’ sequence and high activity QCIF ’Glasgow’ sequence are used. The overall transmission rate of wireless data is set at 128 kbps. For the ‘GOB’ option, N W ind0 W consists of 36 video packets with N group equal to 9 video packets corresponding to 1 video frame at video data transmission rate 41 kbps. We first compare two adaptation systems with RPI-awareness and R PI blind modes. Fig. -5.13 shows the adaptation of concatenated code rate of ’Glasgow’ for each video packet in time varying channel with RPI awareness mode. Distribution of video 80 100 120 GOB num ber 162 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. RPI Range Percentage of Video Packet Drop RPI Aware RPI blind R P I < 49.1098 3.6667 4.0167 49.1098 < R P I < 98.2195 1.8778 1.6889 98.2195 < R P I < 147.3293 1.2611 1.1444 147.3293 < R P I < 196.439 0.7 0.7111 196.439 < R P I < 245.548 0.5944 0.5889 245.548 < R P I < 294.66 0.4667 0.5333 294.66 < R P I < 343.77 0.3389 0.3611 343.77 < R P I < 392.88 0.2778 0.2556 392.88 < R P I < 441.98 0.2889 0.2167 R P I > 441.98 1.3778 2.1723 Total 10.85 11.6899 Table 5.2: The video packet drop percentage in different RPI intervals for the Glas gow sequence with SNR equal to 16dB under the RPI-aware and the RPI-blind modes. packet drop in the ranges of RPI value due to transmission error can be illustrated in Table 5.2. The averaged PSNR.Y results both ’Glasgow’ sequence and ’Silent Voice’ se quence are depicted in Fig. 5.14 and 5.15. Results indicate th at when the adaptation system utilizes the R PI knowledge in optimization, the quality of received video in terms of both averaged PSNR_Y and of the video packet drop percentage is superior than that of the RPI-blind mode. However, the gain from the R PI mode depends on activities of underlying video. The performance comparison of adaptive optimal concatenated code param eter selection with other choices (the same total concatenated code rate) is shown in Fig. 5.16 of the Glasgow sequence. Table 5.3 gives the video packet drop percentage 163 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 29.5 28.5 £ 27.5 -<r- GOB R PI-A w are - 9 - GOB RPI Blind 26.5. 10 12 11 13 14 15 16 17 18 19 20 Figure 5.14: Performance comparison in the averaged PSNR with different opera tional SNR values using a protection solution computed by dynamic programming under equal/unequal error protection and w ith/w ithout RPI with ‘Silent’ video se quence. of different adaptive systems. All these simulations results verify the advantage of using the optimized param eter of the concatenated code over other systems. Next, the difference between instantaneous and delayed feedback channel state information in terms of PSNR values is depicted in Fig. 5.17. This figure illustrates performance degradation due to feedback delay, where we compare cases with 2 channel slot delay and no delay. Prom these results, we see that the feedback delay does deteriorate the end-to-end video performance. 164 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 24 C C 21 2 20 - k - GOB R P I-A w are - 6 - GOB RPI Blind 10 11 12 13 14 15 16 17 18 19 20 C hannel SNR (dB) Figure 5.15: Performance comparison in the averaged PSNR with different opera tional SNR values using a protection solution computed by dynamic programming under equal/unequal error protection and w ith/w ithout RPI with ‘Glasgow’ video sequence. It is worthwhile to point out th at the performance results given above are subject to inefficiency arising from the packetization constraint, which may be magnified due to the variable size GOB packets of the H.263+ video stream. To address this issue, sizes of source packets have to be regulated to match the size of the channel packet. For example, with embedded coding, it can be performed without any overhead. In the case of H.263+, the most promising choice is the sliced data partitioning scheme (Annex V) that allows the regulation of each packet based on the preset bits. The sliced data partitioning scheme also allows internal three-level layering 165 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 23 22 tc 21 5 20 Optimal -© - Inner code 0.72 -A - Inner code 0.57 Inner code 0.8 10 11 12 13 14 15 16 17 18 19 20 C hannel SNR (dB) Figure 5.16: Performance comparison in the averaged PSNR with the optimal selec tion of product code or other choices of the product code. (i.e., the header, the motion vector, and DCT coefficients), which may be utilized for size adjustment. 5.8 C onclusion A new adaptation scheme by considering video source characteristics and time vary ing wireless channel conditions was proposed. The advantage of this joint design is the achievement of both reliable quality of transm itted video and efficient re source usage. Based on the error propagation effect of video characterized by the Relative Priority Index (RPI) and channel state information, the optimization of 166 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Adaptation Scheme Video Packet Drop Rate Optimized adaptive concatenated code 10.85 Adaptive concatenated code with inner code 0.571429 11.72 Adaptive concatenated code with inner code 0.7273 11.83 Adaptive concatenated code with inner code 0.8 14.01 Table 5.3: The video packet drop percentage of the Glasgow sequence in comparing the RPI aware system among the optimized adaptive concatenated code with other adaptive schemes with SNR equal to 16dB. concatenated code allocation for each video packet was investigated. Dealing with both the budget constraint and the packetization constraint (i.e. the dependency problem), the optimal solution can be obtained via dynamic programming. The best parameters of the concatenated code, including the RCPC and RS code ratio, can be determined after the overall concatenated code rate was obtained from the optimization process. Compared to RPI-ignorance non-optimal code rate selection and adaptive RCPC, the improvement in video quality of the proposed method is significant in both objective and subjective measures. However, solutions in this chapter are only related to transmission distortion alone. Design of adaptive wireless video codecs th at take both encoding distortion and transmission distortion into account is a challenging topic. One possible solu tion is to the scalable video codec. Hence, the optimized adaptation for a scalable video codec over the wireless channel should be considered. Some preliminary work along this direction has been presented in this chapter. More complete work will be continued in our future research. 167 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. -0 .0 4 C D -0 .0 6 -g -0 .0 8 - 0.1 < -0 .1 4 -0.16 -0.18 - 0.2 C h an n e l S N R (dB) - 0.02 -0 .0 4 o n -0.06 -g -0.08 - 0.1 -D - 0 . 1 2 < - 0.1 -0.16 -0.18 - 0.2 10 11 12 16 17 18 13 14 15 C h an n e l S N R (dB) Silent Voice Glasgow Figure 5.17: Comparison of quality degradation in averaged PSNR with 2-channel- slot delay or no-delay in channel state information feedback. 168 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. C hapter 6 C onclusion and Future W ork 6.1 C onclusion In this research, we investigated efficient video transmission over non-stationary wireless channels. After a brief background review in Chapter 2, the adaptive chan nel modeling of a time-varying and non-stationary fading channel was proposed in Chapter 3 based on the VLMC structure. It consists of two main components: the channel SNR distribution estimation and VLMC channel modeling. The chan nel SNR distribution was carried out with a kernel density estimation algorithm and used to estimate the unknown distribution of a non-stationary wireless chan nel. Then, the fading partitioning mechanism and VLMC channel modeling were performed. The performance of the proposed adaptive channel modeling method and its applications were verified by simulations in non-stationary wireless channel environments. As shown in simulations, the adaptive channel modeling based on 169 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. the VLMC model provided a much lower complexity and thus higher efficiency in wireless channel representation than the finite state Markov chain. Next, in Chapter 4, we presented a novel dynamic QoS management architecture for adaptive wireless prioritized video transmission with multiple priority classes. This framework employed QoS interaction between video coding and transmission modules, the QoS mapping mechanism, and video quality adaptation. The main contributions of Chapter 4 include the following: (1) the rate constraint derivation of transm itting multiple priority classes with QoS, (2) the proposal of a QoS map ping mechanism that maps statistical QoS guarantees of multiple priority classes to their corresponding expected video quality optimally, and (3) the development of an adaptive QoS algorithm that provides QoS trade-off between the video qual ity requirement and the transmission capability of multiple priority classes under time-varying wireless channels. It was shown via simulation th at the proposed ar chitecture and algorithms offer better video services and enhanced end-to-end video quality. Newr rate adaptation techniques and error control schemes under an optimization framework for reliable video transmission through hostile channels were investigated in Chapter -5 . We presented an adaptive code rate adjustment for reliable wireless video transmission counted on feed back channel status. To estimate the status of time-varying channels, we adopted a concatenated adaptive filter approach consist ing of the Kalman and the LMS filters to handle both long-term and short-term 170 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. fading effects. A flexible channel code rate via concatenated code (RCPC/shortern RS) and adaptive RCPC is selected based on the informed channel state. We also utilized a new measure called the relative priority index (RPI), which quantify the relative importance of a packet in terms of its contribution to end-to-end visual quality to characterize the video source. Both R PI and the estim ated channel status were utilized in the joint source-channel codec design, especially in channel code rate selection. An optimization problem was formulated by considering the bit budget constraint and the packetization constraint, and solved by using dynamic program ming. It was shown that the RPI-aware mode provides a gain in both objective and subjective senses over the RPI-ignorance mode. Furthermore, the adaptive concate nated code together with the best concatenated code param eter selection gives a better performance than adaptive RCPC alone or other choices of the concatenated code’ s parameter. Finally, the uncertainty effect of the channel state information, e.g. delay feedback, on video quality degradation was studied. However, there are still challenging problems to be considered as extensions of our current research. Some examples are given in the next section. 6.2 Future W ork Possible improvements and new research directions of current work are described in the following. • Channel modeling of the networking layer behavior 171 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. A novel channel modeling algorithm based on the variable length Markov chain (VLMC) was studied in Chapter 3 for the wireless physical layer. In principle, the proposed VLMC algorithm can also be applied to the networking layer to model the TC P channel or other types of channels. However, details along this direction need to be developed. • Derivation of the transmission rate constraint of multiple priority classes under a more generic framework As mentioned in Chapter 4, the transmission rate constraint is one of the most im portant parameters in providing dynamic QoS in the next generation wireless transmission system. However, the derivation shown in Chapter 4 is based on the absolute priority scheduling. One possible extension is to derive the transmission rate constraint based on a more generic scheduling scheme. • Joint selection of error control and data rates for video transmission over mul tiple priority classes The proposed dynamic QoS management of transm itting multiple priority classes in Chapter chap4 can be enhanced by applying the error control scheme such as forward error correction (FEC) codes or the Automatic Repeat Request (ARQ) protocol. However, both FEC and ARQ require addition channel re sources. The optimized joint selection of data rates and error control schemes should be considered together with the transmission rate constraint of each service class. 172 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. • Dynamic QoS provisioning for wireless ad-hoc networks The QoS provisioning in wireless ad-hoc networks is an challenging issue due to time-varying wireless channel conditions and the change of the network structure. The dynamic QoS management scheme provided in Chapter 4 can be extended to deal with the time-varying ad-hoc wireless network to provide the best video service. • Complexity reduction of dynamic programming solution The complexity of obtaining the optimal solution via dynamic programming is too high for most practical applications. The high complexity of the developed algorithms is evidenced by the QoS mapping mechanism in Chapter 4 or the search of the optimal FEC code rate in Chapter 5. It is worthwhile to inves tigate an efficient way to reduce the complexity of the dynamic programming search while maintaining a similar visual performance. • Adaptive error concealment based on local texture adaptation Packet loss is inevitable during wireless transmission. Therefore, the error concealment scheme at the decoder side is needed to enhance received video quality. As described in [52], an error concealment scheme can be obtained using texture patterns in image frames. More research work along this direction is expected. 173 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. R eference List [1] B. Arroyo, J. Dasilva, and J. Fernandes, “Life after third generation mobile communications,” IEEE Comrnun. Mag., vol. 39, no. 8 , 2001. [2] A. Anastasopoulos and K. Chugg, “An efficient method for simulation of fre quency selective isotropic rayleigh fading,” Proc. IEEE 47th Vehicular Tech nology Conference, vol. 3, 1997. [3] A. Alwin et, al., “Adaptive mobile multimedia networks,” IEEE Personal Communications, vol. 3, no. 2 , pp. 34-51, 1996. [4] E. Berruto, M. Gudmonson, R. Menolascino, and M. Pizaroso, “Research activities on umts radio interface, network architectures, and planing,” IEEE Communications Magazine, vol. 36, no. 2 , pp. 82-95, 1998. [5] K. Balachandran, K. Kadaba, and S. R. Nanda, “Channel quality estimation and rate adaptation for cellular mobile radio,” IEEE Trans, on Selected Areas in Communications, vol. 17, no. 7, pp. 1244-1256, 1999. [ 6] C. Berrou and A. Glavieux, “Near optimal error correcting coding and de coding: Turbo codes,” IEEE Trans, on Communications, vol. 44, no. 10, pp. 1261-1271, 1996. [7] F. Babich and G. Lombardi, “A markov model for the mobile propagation channel,” IEEE Trans. Veh. Technol., vol. 49, no. 1, pp. 63-73, 2000. [ 8] P. Buhlmann and A.J. Wyner, “Variable length markov chains,” Annals of Statistics, vol. 27, pp. 480-513, 1999. [9] P. Buhlmann, “Model selection for variable length markov chains and tuning the context algorithm,” Annals of the Institute of Statistical Mathematics, vol. 52, pp. 287-315, 2000. [10] Y. Le Boudec and P. Thiran, “A short tutorial on network calculus i: fun damental bounds in communication networks,” IEEE Proceedings of ISC AS 2000, pp. 94-96, 2000. [11] A.R. Calderbank, “Multilevel codes and multistage decoding,” IEEE Trans. Commun., vol. 1 1, no. 2 , pp. 222-229, 1989. 174 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. [12] R.V. Cox, J. Hagenauer, N. Sesgadri, and C. Sundberg, “Variable rate sub band speech coding and matched convolutional channel coding for mobile radio channels,” IEEE Trans. C o m m u n vol. 39, no. 8 , pp. 1717-1731, 1991. [13] V. Chande, H. Jafarkhani, and N. Farvardin, “Joint source-shannel coding of images for channels with feedback,” Proc. IEEE Information Theory Work shop, 1997. [14] G. Cheung and A. Zakhor, “Optimal bit allocation strategy for joint source/channel coding of scalable video,” IEEE Transactions on Image Pro cessing, vol. 9, no. 3, pp. 340-357, 2000. [15] P. Cherriman, T. Keller, and L. Hanzo, “Orthogonal frequency division mul tiplex tranmission of h.263 encoded video over highly frequency-selective wire less networks,” IEEE Trans. Circuits Syst. Video Technol., vol. 9, no. 5, pp. 701-712, 1999. [16] P. Cherriman and L. Hanzo, “Error-rate based power-controlled multimode h.263-assisted video telephony,” IEEE Trans. Circuits Syst. Video Technol, vol. 48, no. 5, pp. 1726-1738, 1999. [17] P.-R. Chang and C.-F. Lin, “Wireless atm-based multimode cdma transport achitecture for mpeg2 video transmission,” Proceeding of the IEEE, vol. 87, no. 10, pp. 1807-1824, 1999. [18] P. Cherriman and L. Hanzo, “Programmable h.263-based wireless video transceivers for interference-limited environment,” IEEE Trans. Circuits Syst. Video Technol, vol. 8 , no. 3, pp. 275-286, 1997. [19] G. Cote, B. Erol, M. Gallant, and F. Kossentini, “H.263+: video coding at low bit rates,” IEEE Trans. Circuits Syst. Video Technol, vol. 8 , no. 7, pp. 849-866, 1998. [20] C. D. Charalambous and N. Menemenlis, “General non-stationary models for short-term and long-term fading channels,” EURACOM M 2000, pp. 142-149, 2000 . [21] M. Chu and W. Stark, “Effect of mobile velocity on communications in fading channels,” IEEE Trans, Veh. Technol., vol. 49, no. 1, pp. 202-210, 2000. [22] T. M. Cover and J.A. Thomas, “Elements of information theory,” John Wiley and Sons, Inc., 1991. [23] L. F. Chang, “Throughput estimation of arq protocols for rayleigh fading chan nel using fade- and interfade-duration statistics,” IEEE Trans, Veh. Technol, vol. 40, no. 1, pp. 223-229, 1991. 175 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 24] P. Chaudhury, W. Mohr, and S. Onoe, “The 3gpp proposal for imt-2000,” IEEE Comm,un. Mag., vol. 37, no. 12, pp. 72-81, 1999. 25] C.-S. Chang and J. Thomas, “Effective bandwidth in high speed digital net works,” IEEE J. Select. Areas Commun., vol. 13, no. 6 , pp. 1091-1100, 1995. 26] C.-S. Chang and T. Zajic, “Effective bandwidths of departure processes from queues with time varying capacities,” IEEE INFOCOM ’95, vol. 3, pp. 1001— 1009, 1995. 27] P.A. Chou and Z. Miao, “Rate-distortion optimized streaming of packetized media,” submitted to IEEE Trans, on Multimedia. 28] C. Dovrolis, D. Stiliadis, and P. Ram anathan, “Proportional differentiated sendees: delay differentiation and packet scheduling,” IE E E /A C M Trans, on Networking, vol. 10, no. 1, pp. 12-26, 2002. 29] E.O. Elliott, “Estimates of error rates for codes on burst-noise channels,” Bell-Syst. Tech. J., vol. 42, pp. 1977-1997, 1963. 30] T. Eyceoz, A. Duel-Hallen, and H. Hallen, “Prediction of fast fading parame ters by resolving the interference pattern,” IE E E /A C M Trans, on Networking, vol. 1, pp. 167-171, 1997. 31] N. Farber, K. Stuhlmuller, and B. Girod, “Analysis of error propagation in hybrid video coding with application to error resilience,” ICIP 99. Proceedings International Conference on Image Processing, vol. 2, pp. 550-554, 1999. 32] B.D. Fritchman, “A binary channel characterization using partitioned markov chains,” IEEE Trans. Veh. Technol, vol. VT-13, no. 2, pp. 221-227, 1967. 33] A. Goldsmith and S.-G. Chua, “Variable-rate variable-power mqam for fading channels,” IEEE Trans, on Communications, vol. 45, no. 10, pp. 1218-1230, 1997. 34] D. Goeckel, “Adaptive coding for time-varying channels using outdated fading estimates,” IEEE Trans, on Communications, vol. 47, no. 6 , pp. 844-855, 1999. 35] M. Gudmundson, “Correlation model for shadowing fading in mobile radio systems,” Electronics Letters, vol. 27, no. 23, pp. 2145-2146, 1999. 36] N. Frber B. Girod, “Wireless video,” Compressed Video over Networks, 2000. 37] E.N. Gilbert, “Capacity of a burst-noise channel,” Bell-Syst. Tech. J.,, vol. 39, pp. 1253-1266, 1960. 176 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. [38] A. Goldsmith and S. Chua, “Variable-rate variable power mqam for fading channels,” IEEE Trans. Commun., vol. 45, no. 10, pp. 1218-1230, 1997. [39] A. Goldsmith and S. Wicker, “Design challenges for energy-constrained ad hoc wireless networks,” IEEE Wireless Communications, vol. 9, no. 4, pp. 8-27, 2002. [40] H. Holma and A. Toskala, “W cdma for umts: Radio access for third generation mobile communications,” Wiley, 2000. [41] L. Hanzo, C. H. Wong, and P. Cherriman, “Channel-adaptive wideband wire less video telephony,” IEEE Signal Processing Magazine, vol. 17, no. 4, pp. 10-30, 2000. [42] J. Hagenauer, “Rate-compatible punctured convolution codes (rcpc codes) and their applications,” IEEE Trans, on Communications, vol. 36, no. 4, pp. 389-400, 1988. [43] C.-Y. Hsu, A. Ortega, and A. R. Reibman, “Joint selection of source and channel rate for vbr video transmission under atm policing constraints,” IEEE Journal on Sel. Areas in Communications, vol. 15, no. 6 , pp. 1016-1028, 1997. [44] R. A. Howard, “Dynamic probabilistic systems,” John Wiley Sons INC., 1971. [45] S. Haykin, “Adaptive filter theory,” Prentice Hall, 1996. [46] S. Hara, A. Ogino, M. Araki, M. Okada, and N. Morinaga, “Throughput performance of saw-arq protocol with adaptive packet length in mobile packet data transmission,” IEEE Trans. Veh. Technol., vol. 45, no. 3, pp. 561-569, 1996. [47] C.Y. Hsu, A. Ortega, and M. Khansari, “Rate control for robust video trans mission over burst-error wireless channels,” IEEE J. Select. Areas Commun., vol. 17, no. 5, pp. 756-773, 1999. [48] J. Kovacevic and V. K. Goyal, “Multiple descriptions: source-channel coding methods for communications,” Proc. 10th Tyrrhenian Int. Workshop on Dig. Comm.: Multimedia Comm., 1998. [49] W. Kumwilaisak J. Kim and C.-C. Jay Kuo, “A performance evaluation of data partitioning annex for wireless h.263+ video,” Proc. International Sym posium. on Multimedia Information Processing (ISMIP) ‘ 99, 1999. [50] J. Kim A Y . Kumwilaisak and C.-C. J. Kuo, “Reliable wireless video transmis sion via fading channel estimation and adaptation,” in Proc. Second IEEE Wireless Communication and Networking Conference, 2000. 177 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. [51] W. Kumwilaisak and C.-C. Jay Kuo, “Adaptive variable length markov chain for non-stationary fading channel modeling,” submitted to IEEE Trans, on Vehicular Technology, 2002. [52] W. Kumwilaisak and F. Hartung, “Intra-frame error concealment: non-linear pattern alignment and directional interpolation,” submitted to IEEE Trans, on Circuit and System for Video Technology, 2002. [53] J.-G. Kim, J. Kim, and C.-C. J. Kuo, “Corruption model of loss propagation for relative prioritized packet video,” SPIE Proc. Applications of Digital Image Processing XXIII, 2000. [54] W. Kumwilaisak, J. Kim, and C.-C. J. Kuo, “Reliable video transmission over fading wireless channel with joint source-channel adaptation,” IE E E /A C M Trans, on Networking, 2001. [55] G. Kamosa, “Video coding and robust transmission in error prone environ ment,” Master thesis .Dept, of Electrical Engineering, University of Califor nia, Los Angeles, 1998. [56] P. Karn, “Forward error correcting codes,” http://people.qualcomm.com/karn/code/fec. [57] S. Kellel, “Analysis of memory and incremental redundancy arq schemes over a nonstationary channel,” IEEE Trans. Commun., vol. 40, no. 9, pp. 1474- 1480, 1992. [58] A. Konrad, B. Y. Zhao, A. Joseph, and R. Ludwig, “A markov-based channel model algorithm for wireless networks,” m Proceedings of Fourth AC M In ternational Workshop on Modeling, Analysis and Simulation of Wireless and Mobile Systems, ACM MSWiM, 2001. [59] G. Kesidis, J. Walrand, and C.-S. Chang, “Effective bandwidths for multiclass markov fluids and other atm sources,” IE E E /A C M Trans, on Networking, vol. 1, no. 4, pp. 424-428, 1993. [60] S. Lin and D. Costello, “Error control coding,” Prentice Hall, 1983. [61] M.-C. Lin, C.-C. Lin, and S. Lin, “Computer search for binary cyclic vep codes of odd length up to 65,” IEEE Trans. Inform. Theory, vol. 36, no. 4, p p . 924-935, 1990. [62] J. Lu, A. Nosratinia, and B. Aazhang, “Progressive source-channel coding of images for feedback channels,” Proc. Data Compression Conference, 1999. [63] V. K. N. Lau and S. V. Marie, “Varible rate adaptive modulation for ds-cdma,” IEEE Trans, on Communications, vol. 47, no. 4, pp. 557-589, 1999. 178 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. [64] Y. J. Liang, N. Farber, and B. Girod, “Adaptive playout scheduling using time-scale modification in packet voice communication,” Proc. IEEE Inter national Conference on Acoustics Speech and Signal Processing, IC A SSP ’01, 2001 . [65] C.H.C.Leung, Y.Kikumoto, and S.A. Sorensen, “The throughput efficiency of the go-back-n arq scheme under markov and related error structures,” IEEE Trans. Commun., vol. 36, no. 2, pp. 881-887, 1998. [ 66] S. Lu. K. Lee, and V. Bharghavan, “Adaptive service in mobile computing environments,” Building QoS Into Distributed Systems, A. Campbell and K. Nharstedt, Eds. London, U.K.: Chapman Hall, 1997. [67] H. Man, F. Kossentini, and M .J.T. Smith, “Robust ezw image coding for noisy channel,” IEEE Signal Processing Letter, vol. 4, no. 8 , pp. 227-229, 1997. [ 6 8] J.W. Modestino, D.G. Daut, and A.L. Vickers, “Combined source-channel coding of images using the block cosine transform,” Proc. IEEE Trans. Com mun., vol. 29, pp. 1261-1274, 1981. [69] M. W. Marcellin and T.R. Fischer, “Trellis coded quantization of memory less and gauss-markov sources,” IEEE Trans, on Communications, vol. 38, no. 1, pp. 82-93, 1990. [70] M. J. Mason, G. C. Hess, and S. S. Gilbert, “Shadowing variability in an urban land mobile environment at 900 mhz,” Electronics Letters, vol. 26, no. 10, 1990. [71] S. Manistis, E. Nikolouzou, and I. Venieris, “Qos issues in the converged 3g wireless and wired networks,” IEEE Commun. Mag., vol. 40, no. 8 , pp. 44-53, 2002 . [72] M. J. Mason, G. C. Hess, and S. S. Gilbert, “Shadowing variability in an urban land mobile environment at 900 mhz,” Electronics Letters, vol. 26, no. 10, 1990. [73] M. Mirhakkak, N. Schult, and D. Thomson, “Dynamic bandwidth manage ment and adaptive applications for a variable bandwidth wireless environ ment,” IEEE J. Select. Areas Comm,un., vol. 19, no. 10, pp. 1984-1997, 2001. [74] J.C. de Martin, “Source-driven packet marking for speech transmission over differentiated-service networks,” IEEE Proceedings of International Confer ence on Acoustics, Speech, and Signal Processing, 2001. [75] E. Masala, D. Quaglia, and J.C. de Martin, “Adaptive picture slicing for distortion-based classification of video packets,” IEEE Proceeding Workshop on Multimedia Signal Processing, 2001. 179 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. [76] S. Nanda, K. Balachandran, and S. Kumar, “Adaptation techniques in wireless packet data services,” IEEE Commun. Mag., vol. 38, no. 1, pp. 54-64, 2000. [77] A. Ortega and K. Ramchandran, “Rate-distortion techniques in image and video compression,” IEEE Signal Processing Magazine, vol. 15, no. 6 , pp. 23-50,1998. [78] A. Ortega, “Optimization techniques for adaptive quantization of image and video under delay constraints,” Ph.D. Thesis, Dept, of Electrical Engineering, Columbia University, 1994. [79] H. V. Poor and G. W. Wornell, “Wireless communications: signal processing perspectives,” Prentice Hall, 1998. [80] J. W. Park, J. W. Kim, and S. U. Lee, “Dct, coefficients recovery based error concealment technique and application to the mpeg-2 bit stream error,” IEEE Trans. Circuits and Systems for Video Technology, vol. 7, no. 6 , pp. 845-854, 1997. [81] Z. Huang Peng and Jr. Y.-F. Costello, D.J., “Turbo codes for image transmission-a joint channel and source decoding approach,” Selected Areas in Communications. IEEE Journal on, vol. 18, no. 6 , pp. 868-879, 2000. [82] J. Proakis, “Digital communications,” Mcgraw-Hill, Third Ed., 1995. [83] V. Padmanabhan, “Using differentiated services mechanisms to improve net work protocol and application performance,” R TA S Workshop on QoS Support for Real-Time Internet Applications, 1999. [84] D.W. Redmill and N.G. Kingsbury, “The erec: An error resilient techniques for coding variable length blocks of data,” IEEE Trans, on Image Processing, vol. 5, no. 4, pp. 556-574, 1996. [85] R. E. Van Dyck and D. J. Miller, “Transport of wireless video using separate, concatenated, and joint source -channel coding,” Proceeding of the IEEE, vol. 87, no. 10, pp. 1734-1750, 1993. [ 8 6] K. Ramchandran, A. Ortega, K.M. Uz, and M. Vetterli, “Multiresolution broadcast for digital hdtv using joint source-channel coding,” IEEE JSAC, Special issue on High Definition Television and Digital Video Communications, vol. 11, no. 1, pp. 6-23, 1993. [87] K. Ramchandran, A. Ortega, and M. Vetterli, “Bit allocation for dependent quantization with applications to multiresolution and mpeg video coders,” IEEE Transactions on Im.age Processing, vol. 3, no. 5, pp. 533-545, 1994, 180 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. [ 88] R. Ram anathan and R. Hain, “An ad hoc wireless testbed for scalable, adap tive qos support,” IEEE WCNC, vol. 3, pp. 998-1002, 2000. [89] D. L. Reyes, A.R. Reibman, S.-F. Chang, and J.I.-I. Chuang, “Error-resilient transcoding for video over wireless channels,” IEEE J. Select. Areas Commun., vol. 18, no. 6 , pp. 1063-1074, 2000. [90] C.E. Shannon, “A mathematical theory of communication,” Bell Syst. Tech. Journal vol. 27, pp. 379-423, 1948. [91] R. Schafer, “Terestrial transmission of dtvb signals-the european specifica tion,” Int. Broadcasting Conversion, , no. 413, 1995. [92] J. M. Shapiro, “Embedded image coding using zerotrees of wavelet coeffi cients,” IEEE Trans. Signal Processing, vol. 41, no. 12, pp. 3445-3462, 1993. [93] A. Said and W. A. Pearlman, “A new fast and efficient image codec based on set partitioning in hierachical trees,” IEEE Trans. Circuits Syst. Video Technol., vol. 6 , no. 3, pp. 243-250, 1996. [94] P. G. Sherwood and K. Zeger, “Progressive image coding for noisy channel,” IEEE Signal Processing Lett., vol. 4, no. 7, pp. 189-191, 1997. [95] P. G. Sherwood and K. Zeger, “Error protection for progressive image trans mission over memoryless and fading channels,” IEEE Trans, on Communica tions, vol. 46, no. 12, pp. 1555-1559, 1998. [96] D.G. Sachs, R. Anand, and K. Ramchandran, “Wireless image transmission using multiple-description based concatenated codes,” Proceedings DCC 2000, 2000 . [97] R. Steele and L. Hanzo, “Proportional differentiated services: delay differen tiation and packet scheduling,” Mobile Radio Communications : Second and third Generation Cellular and WATM Systems, 2002. [98] J. Kim J. Shin and C.-C. J. Kuo, “Content-based packet video forwarding mechanism in differentiated service networks,” Proc. Packet Video Workshop ’ 2000 , 2000 . [99] Y. Shoham and A. Gersho, “Efficient bit allocation for an arbitrary set of quantizers,” IEEE Trans. ASSP, vol. 36, no. 9, pp. 1445-1453, 1988. [100] E. G. Steinbach, N. Farber, and B. Girod, “Adaptive playout for low latency video streaming,” IEEE Proc. International Symposium on Circuits and Sys tems, IS C A S’ 02, 2002. 181 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. [ 1 0 1 ] F.R. Kschischang M. Sajadieh and A. Leon-Garcia, “A block memory model for correlated rayleigh fading channels,” in Proc: IEEE Int. Conf. Commun., p p . 2 8 2 - 2 8 6 , 1 9 9 6 . [102] S. Sivaprakasam and K.S. Shanmugan, “An equivalent markov model for burst errors in digital channels,” IEEE Trans. Commun., vol. 43, no. 2, pp. 1347-1355, 1995. [103] C. Sgraja and J. Lindner, “Pilot-/data-aided estimation of rayleigh flat-fading channels,” Proc. 7th International OFDM-Workshop, 2002. [104] G.L. Stuber, “Principles of mobile communication,” Kluwer, 1996. [105] C. K. Siew and D. J. Goodman, “Packet data transmission over mobile radio channels,” IEEE Trans. Veh. Technol, vol. 38, no. 2, pp. 95-101, 1989. [106] I. Stoica, S. Shenkar, and H. Zhang, “Core-stateless fair queueing: Achieving approximately fair bandwidth allocations in high speed networks,” Proc. of ACM SIG CO M M ’ 98, p p . 118-130, 1998. [107] S.D. Servetto, K. Ramchandran, K. Nahrstedt, and A. Ortega, “Optimal segmentation of a vbr source for its parallel transmission over multiple atm connections,” Proc. In t’ l Conf. Image Processing, volume = 2, pages = 5-8, year = 1997,. [108] J. Shin, J. Kim, and C.-C. Jay Kuo, “Quality-of-service mapping mechanism for packet video in differentiated services network,” IEEE Trans, on Multime dia, vol. 3, no. 2, pp. 219-231, 2001. [109] A. Sehgal and P. A. Chou, “Cost-distortion optimized stream ing media over diffserv networks,” IEEE International Conference on Multimedia and Expo, vol. 2, pp. 857-860, 2002. [110] W. Turin, “Digital transmission systems: Performance analysis and model ing,” McGraw-Hill, 1998. [111] C. Tan and N. Beaulieu, “On first-order markov modeling for the rayleigh fading channel,” IEEE Trans. Commun., vol. 48, no. 12, pp. 2032-2040, 2000. [112] W. Turin and R. van Nobelen, “Hidden markov modeling of flat fading chan nels,” IEEE Trans, on Selected Areas in Comunn., vol. 16, no. 9, pp. 1809- 1817, 1998. [113] W. Turin, “Througput analysis of the go-back-n protocol in fading radio channels,” IEEE Trans, on Selected Areas in Comunn., vol. 17, no. 5, pp. 881-887, 1999. 182 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. [114] W. Tan and A. Zhakor, “Packet classification schemes for streaming mpeg video over delay and loss differentiated networks,” Proceedings of Packet Video Workshop 2001, 2001. [115] V.A. Vaishampayan and N. Farvardin, “Optimal block cosine transform image coding for noisy channel,” IEEE Trans, on Communications, vol. 38, pp. 327- 336, 1990. [116] B. Vandaiore, R. Jain, S. Fahiny, and S. Dixit, “Aquafwin : Adaptive qos framework for multimedia in wireless networks and its comparison with other qos frameworks,” Local Computer Networks, pp. 88-97, 1999. [117] Y. Wang and Q.-F. Zhu, “Error control and concealment for video commu nication: A review,” Proceedings of the IEEE, vol. 86, no. 5, pp. 974-997, 1998. [118] J. Wen and J.D. Villasenor, “A class of reversible variable length codes for robust image and video coding,” Proc. of IEEE International Conference on Image Processing, vol. 2, pp. 65-68, 1997. [119] S. B. Wicker, “Error control systems for digital communication and storage,” Prentice Hall, 1995. [120] H.S. Wang and N. Moayeri, “Finite-state markov channel-a useful model for radio communication channels,” IEEE Trans. Veh. Technol, vol. 45, no. 1, pp. 163-171, 1995. [121] H.S Wang and P.-C. Chang, “On verifying the first-order markovian assump tion for a rayleigh fading channel model,” IEEE Trans. Veh. Technol., vol. 45, no. 2, pp. 163-171, 1995. [122] M. P. Wand and M. C. Jones, “Proportional differentiated services: delay differentiation and packet scheduling,” IE E E /A C M Trans, on Networking, vol. 10, no. 1, pp. 12-26, 2002. [123] F. Wu, S. Li, and Y.-Q. Zhang, “A framework for efficient progressive fine granularity scalable video coding,” IEEE Trans, on Circuits and Systems for Video Technology, vol. 11, no. 3, pp. 332-344, 2001. [124] D. Wu and R. Negi, “Effective capacity: A wireless link model for support of quality of service,” to appear in IEEE Transactions on Wireless Communica tions. [125] L. Xiao et al., “Joint optimization of communication rates and linear systems,” Proc. IEEE Conf. Dec. Cntl, vol. 3, pp. 2321-2326, 2002. 183 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. [126] Q.F. Zhu, Y. Wang, and L. Shaw, “Coding and cell-loss recovery in dct-based packet video,” IEEE Trans, on Circuits and Systems for Video Technology, vol. 3, no. 3, pp. 238-247, 1993. [127] M. Zoonizi and P. Dassanayake, “Correlation model for shadowing fading in mobile radio systems,” Personal, Indoor and Mobile Radio Communications, vol. 2, 1996. [128] M. Zorzi, R.R. Rao, and L.B. Milstein, “Arq error control for fading mobile radio channel,” IEEE Trans. Veh. Technol, vol. 46, no. 2, pp. 445-455, 1997. [129] M. Zorzi, “Some results on error control for burst-error channels under delay constraints,” IEEE Trans. Veh. Technol., vol. 50, no. 1, pp. 12-24, 2001. [130] Z.-L. Zhang, “End-toend support for statistical quality-of-service guarantees in multimedia networks,” Ph.D. Dissertation,Department of Computer Sci ence, University of Massachusetts, 1997. 184 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.

Linked assets

University of Southern California Dissertations and Theses

Conceptually similar

PDF

Error resilient techniques for robust video transmission

PDF

Contributions to image and video coding for reliable and secure communications

PDF

Algorithms and architectures for robust video transmission

PDF

Design and analysis of server scheduling for video -on -demand systems

PDF

Color processing and rate control for storage and transmission of digital image and video

PDF

Intelligent systems for video analysis and access over the Internet

PDF

Design and analysis of MAC protocols for broadband wired/wireless networks

PDF

Dynamic radio resource management for 2G and 3G wireless systems

PDF

Contributions to coding techniques for wireless multimedia communication

PDF

Contributions to efficient vector quantization and frequency assignment design and implementation

PDF

Advanced video coding techniques for Internet streaming and DVB applications

PDF

Design and performance analysis of low complexity encoding algorithm for H.264 /AVC

PDF

Design and applications of MPEG video markup language (MPML)

PDF

Contributions to content -based image retrieval

PDF

Blind multiuser receivers for DS -CDMA in frequency-selective fading channels

PDF

Information hiding in digital images: Watermarking and steganography

PDF

Complexity -distortion tradeoffs in image and video compression

PDF

Efficient acoustic noise suppression for audio signals

PDF

Fading channel equalization and video traffic classification using nonlinear signal processing techniques

PDF

Call admission control and resource allocation for quality of service support in wireless multimedia networks

Asset Metadata

Creator Kumwilaisak, Wuttipong (author)

Core Title Adaptive video transmission over wireless fading channel

School Graduate School

Degree Doctor of Philosophy

Degree Program Electrical Engineering

Publisher University of Southern California (original), University of Southern California. Libraries (digital)

Tag engineering, electronics and electrical,OAI-PMH Harvest

Language English

Contributor Digitized by ProQuest (provenance)

Advisor Kuo, C.-C. Jay (committee chair), Ortega, Antonio (committee member), Zimmermann, Roger (committee member)

Permanent Link (DOI) https://doi.org/10.25549/usctheses-c16-369706

Unique identifier UC11334941

Identifier 3103921.pdf (filename),usctheses-c16-369706 (legacy record id)

Legacy Identifier 3103921.pdf

Dmrecord 369706

Document Type Dissertation

Rights Kumwilaisak, Wuttipong

Type texts

Source University of Southern California (contributing entity), University of Southern California Dissertations and Theses (collection)

Access Conditions The author retains rights to his/her dissertation, thesis or other graduate work according to U.S. copyright law. Electronic access is being provided by the USC Libraries in agreement with the au...

Repository Name University of Southern California Digital Library

Repository Location USC Digital Library, University of Southern California, University Park Campus, Los Angeles, California 90089, USA