Application of Multiple Description Coding in 4G Wireless Communication Systems

Transcription

1 Application of Multiple Description Coding in 4G Wireless Communication Systems Frank H.P. Fitzek, Başak Can, Ramjee Prasad, DS Park, Youngkwon Cho Department of Communications Technology, Aalborg University Neils Jernes Vej 12, 9220 Aalborg Øst, Denmark; ph: , [ff bc SAMSUNG ELECTRONICS CO., LTD.; [youngkn Abstract In this paper we introduce and advocate the use of multiple description coding (C) in presence of channel diversity for fourth generation (4G) wireless communication networks. C allows to use the spectrum efficiently and simultaneously supporting heterogenous terminals. In this paper the investigations are made for voice and video type traffic. We present the drawbacks of the C scheme in terms of the overhead generated by the network layer and the encoding process. I. INTRODUCTION Coding in presence of channel diversity can make a wireless link very robust to propagation errors. It randomizes errors over several channels whereby the error rate is greatly reduced. Multiple description coding (C) divides the serial data stream into n data streams, with n>1. These n data streams are coded separately and transmitted over n channels. At the receiver these n data streams are decoded individually and independently. The quality of the reception increases as the number decodable channels increases. If there is at least one channel which is decodable, a minimum of service can be guaranteed through the advantage of coding. As given in Figure 1, the channel quality will change over time. In case a multimedia source has to be transmitted over this kind of channel the question arises at which rate the information has to be coded. Adaptive schemes may be used but a feedback channel is needed in such case. Three possibilities are given in Figure 1: i. The source code is encoded with a low rate (dashed line) and therefore the quality degrades significantly. ii. The source code is encoded at a high rate (dotted line). The quality is very bad when the data rate of the encoded bit stream exceeds that of channel. iii. The source is encoded using C scheme. Therefore, the spectrum can be used in a very efficient manner (solid line). The more capacity is avaible for the terminal, the more descriptors will be received. Fig. 1. Channel condition, data rate and application quality for C and non C source coding Fig. 2. Protocol stack in case C is used

2 Figure 2 shows a general description of an C system with n=3. By splitting the source data into several descriptors and transmitting the encoded streams over several channels, each channel encounters almost flat fading and the probability of receiving some information from at least one channel is greatly increased with the price of increased overhead [1]. A. Support of Heterogeneous Terminals 4G mobile systems focus on seamlessly integrating the existing wireless technologies including existing technologies such as GSM, Wireless LAN, personal area networks (PANs) and Bluetooth. It is envisioned that the offered services will be independent of the underlaying network and will be able to support heterogeneous terminals. In this section, we introduce the possibility of supporting heterogeneous terminals by C. Two key features to be considered in mobile communications are power consumption and data rate. Let device grade represent the level of battery power and the maximum data rate that a terminal can support. It is very obvious that a laptop has a higher device grade than a mobile phone. For example a laptop connected to a wireless LAN can receive data at a rate of 54 Mbps [2]. However a cellular phone which has access to GPRS (General Packet Radio Service) may receive data at a rate of 100 kbits/second [2]. Therefore in an C system like the one depicted in Figure 2, a laptop is able to receive all of the information generated by D1, D2 and D3 but a mobile phone may not be able to receive all of them. For instance, it may receive data generated by only one of the descriptors. An C system can grant different number of descriptors to mobile terminals having different device grades. Mobile terminals which have a high device grade can be granted more descriptors than the ones which have a low device grade. The quality of received packets increases as the number of the granted descriptors increases. encoded at three different quantization values (1, 31, and 51) are given. If C is not used, that is, if only one stream is generated, then the network overhead is zero. If the quantization parameter (QP) is low, e.g. 1, than this means that the video stream consumes a large bandwidth and thus the effect of the network overhead to increasing number of sub-streams becomes less. However, for high values of the QP, the network overhead versus the number of sub-streams generated is larger. It is obvious in these figures that the network overhead is increasing with the number of sub-streams generated by the application layer. In the IP world, the transmission of multiple descriptors to the receiver is mostly done by using the real time protocol (RTP) together with the real time control protocol (RTCP). RTP and RTCP protocols transport their packet via the user datagram protocol (UDP) to the underlying Internet protocol (IP) layer. As given in Figure 5, the multiple description coded streams are sent to the RTP layer by the application layer. RTP layer packetizes each data segment into an RTP packet. Each layer adds an additional overhead to each data stream that is transmitted. Using RTP/UDP/IP an overhead of 40 byte per stream has to be taken into account. In Figures 3 and 4, the network overhead in addition to the encoding overhead for the foreman video sequence encoded at three different quantization values (1, 31, and 51) is also given for IP version 4 and IP version 6, respectively. Splitting the video sequence into 20 sub streams leads to an overhead of 1.3 introduced by the encoder. If the network overhead of IPv6 is also taken under consideration, the overhead increases up to 9.2. Thus some scenarios the network overhead becomes dramatically larger than the overhead produced by the encoding process. The overhead generated by an C system used for voice communications depends on the number of descriptors incorporated and the type of the encoders used. To reserve B. Overhead caused by C The outstanding capabilities of C come at the price of larger overhead θ due to the encoding process. The overhead θ is defined as Foreman QCIF J i=1 θ = B i,k B 1 (1) network overhead where B i,k is the bandwidth of each descriptor i describing source k and B is the bandwidth for the single descriptor. For video communication the overhead is a function of the number of descriptors, the codec used and the content itself. This makes it hard even to give an approximate value of the overhead. In Figures 3 and 4, the overhead produced by the encoding process for the foreman video sequence number of sub-streams QP 1 QP 31 QP 51 QP 1 + Network QP 31 + Network QP 51 + Network Fig. 3. The network overhead for the foreman video sequence and three different quantization values (namely 1, 31, and 51) for IP version 4 [3]

3 network overhead Foreman QCIF number of sub-streams QP 1 QP 31 QP 51 QP 1 + Network QP 31 + Network QP 51 + Network Fig. 4. The network overhead for the foreman video sequence and three different quantization values (namely 1, 31, and 51) for IP version 6 [3] stream 1 stream 2 Application stream... RTP UDP IP stream J Audio stream 1 Audio stream 2 RTCP rep RTCP rep RTCP rep RTCP rep RTCP rep RTCP rep socket socket socket socket socket socket Fig. 5. Multiple description coded sources arriving at the RTP layer [3] bandwidth, the overall bit rate of the C system must not exceed the bandwidth required for single stream schemes. If voice coding based on pulse code modulation (PCM) is used, the overhead depends only on the codec used for each of the descriptors. If voice coding based on silence detection is used, then the overhead may depend on the type of conversation. For example, if the conversation is taking place in a police department, people always tend to speak very quickly which leads to smaller silence intervals (leading to more transmitted bits), however, if the conversation is just a chat, than it is more probable to have longer silence times. This means that fewer bits are transmitted. In addition to the encoding overhead, an increased network overhead is expected. Assuming IP transport, each descriptor will carry the IP overhead. Some work has already investigated the possibility to use real time control protocol (RTCP) messages for MCP [4]. II. QUALITY MEASUREMENTS As an example for the video quality in dependency of the number of received descriptors, we refer to Figure 6. In case all descriptors are received, the quality is always high. In case only a subset of the descriptors is received, the quality goes up and decreases rapidly. In case only one out of twenty streams is received, the PSNR values are less than 20dB. III. LITERATURE REVIEW In [5], the performance of a single stream based voice communication is compared with that of multiple description coding (multistream communications) using Forward Error Correction (FEC) scheme. Therefore, the payload data rate of single stream and multistream schemes are identical and C does not increase the overall data rate of the payload compared to single stream scheme. In this way, the added overhead of packet headers as a result of transmitting multiple streams are compared. In [5], it was reported that with 30 milliseconds packetization time, 8-bit Pulse Code Modulation (PCM) / 4 bit Adaptive Differential Pulse Code Modulation (ADPCM) coded payload, and 40 bytes of Real Time Protocol (RTP), User Datagram Protocol (UDP) and IPv4 headers for each packet, the overall data rate of multiple streaming is 11 % more than single- stream FEC. Wireless links consist of bandwidth limited and error prone channels. Each channel may have different bandwidths. communication over wireless links therefore requires high compression and robustness against errors. Achieving both high compression and robustness against errors is a challenge because there is a big trade off between these two [6]. The study carried out in [6] combats errors encountered in the wireless channel by coding the highly compressed video data into multiple independently decodable streams. The proposed C coding schemes in the literature does not take receiving device characteristics and bandwidth variations into account [7]. Such C techniques can be referred as non-scalable coding techniques. In [7], a new C technique called Multiple Description Scalable Coding (SC) is proposed. SC can simply be described as a combination of C and scalable coding. This technique addresses receiving device characteristics and bandwidth variations of the channel and also enables tradeoffs between throughput, redundancy and complexity which is not possible with non-scalable C schemes. In SC, the number and the composition of descriptions are changed dynamically to make the proposed system very robust to changing channel characteristics [7]. In the context of SC, authors introduce a new C scheme called Multiple Description Motion Compensated Temporal Filtering (-MCTF)in [7]. As introduced in [7], MCTF is designed for the removal of redundancies of video sequences. The proposed - MCTF scheme is used together with wavelet based video

4 35 PSNR comparison foreman 30 video quality (PSNR [db]) frame number 20/20 10/20 1/20 Fig. 6. quality of the foreman video sequence for 1, 10, and 20 out of twenty descriptors [3] coding schemes whereby providing a scalable and highly error resilient video transmission over wireless links [7]. Authors include the quality comparisons with previously proposed C schemes and their results show that, - MCTF outperforms them by approximately db in PSNR. C uses path diversity in order to combat with errors. The available bandwidth of each path in a packet network may be different and time varying [8]. communication over those kind of links is obstructed by limited bandwidth and packet loss. For a good performance over such kind of paths, an unbalanced C scheme is proposed in [8]. This scheme creates unbalanced Multiple Description () streams by adjusting the frame rate, quantization or spatial resolution of each stream and can achieve unbalanced rates of up to 2:1. Authors compare the conventional C system with their proposed system. Their results show that unbalanced C scheme provides improved reliability over multiple paths with unequal bandwidths. Authors examine the effectiveness of unbalanced C and path diversity for transmission of Foreman (144 x 176 pixels/f, 30 f/s ) and Bus (240 x 352 pixels/f, 30 f/s) sequences. To make a fare comparison of the proposed C scheme with conventional SD scheme, the Foreman and Bus sequences are coded at the same frame rate (22.5 f/s) [8]. The recovered video quality for unbalanced rates of up to 1.75:1 and 1.76:1 for Foreman and Bus are compared with that of SD scheme. Their results obviously show that the proposed unbalanced C scheme is very robust to bursty errors and network congestions than SD scheme. Content delivery networks (CDN) are developed to overcome overloads and delays that can arise when many users overload the network to access popular content. The performance of an -Streaming Mode (SM)-CDN network is compared with that of conventional Single Description (SD)-SM-CDN in [9]. The simulations in [9] show that -SM-CDN has a better performance than a conventional SD-SM-CDM. In [10], an algorithm for optimal splitting of DCT coefficients in an system is proposed. In the algorithm, the authors use redundancy rate-distortion (RRD) criteria and split one compressed video stream into two correlated streams optimally. Their results show better performance than the conventional splitting algorithms based on RRD criteria

5 at all rates. In [11], an C system with cross layer optimization scheme is proposed. The application and network layers cooperate to provide more robustness against severe network conditions. The authors propose a multi-path selection method that chooses a set of paths to achieve more robust media transmission. The media content is transmitted over these intelligently selected multiple paths instead of shortest path or maximally disjoint paths. The simulation results in this paper show an average peak-signal-to-noise ratio (PSNR) improvement of up to 8.1 db than conventional C scheme. The challenge that the authors faced in this paper was finding a set of paths to minimize the cost function that estimated average streaming distortion in terms of network statistics, media characteristics and application requirements [12]. To remedy this problem, they proposed a fast heuristics-based solution by exploiting the infrastructure features of the internet in [12]. Their simulations run over several random internet topologies show that the proposed heuristic approach is capable of finding a good set of paths in a much shorter time than the methods they used in [11]. This heuristic is therefore very well suited for time critical applications such as VoIP and video conferencing. In [13] layered multiple description scheme is proposed. Layered codes incorporate sequences of layers in order to provide bandwidth heterogeneity and to cope with dynamic network congestion[13]. Authors use base layer descriptions for low bandwidth clients whereas they use both base and enhancement layer descriptions for high bandwidth clients. Vinay Anant Vaishampayan [14] designed a multiple description scalar quantizer (SQ) which is designed for operation in a diversity system. In this scheme, the encoder sends the information over several wireless channels which are subject to bandwidth constraint. An encoder-decoder pair that minimizes the average distortion between quantized and original data streams is designed in [14]. The effectiveness of the combination of C with multiple path transport (MPT) for video and image transmission over multiple hop wireless links is studied in [4]. Some applications have high bandwidth and stringent reliability requirements. For supporting these kind of applications, the sender may need to send the data stream over several paths to the destination [4]. The study in [4] addresses this need and the authors compare their proposed C-MPT system with a system which uses layered coding and asymmetrical paths for the base and enhancement layers. In [15], a practical C scheme have been developed for video, and its performance over two state Markov channels and Rayleigh fading channels have been characterized. An C technique based on matching pursuits (MP) signal decomposition, called MP-C [15], has been shown to outperform Discrete Cosine Transform (DCT) based C [15]. In the case of Rayleigh fading channels, interleaving applied to Single Description Coding (SDC) has been shown to outperform C at the expense of additional delay [15].It has also been shown in [15] that C outperforms SDC in the case of bursty and slowly varying environments. The motion vector data is the most important part of the compressed video stream [16]. It is therefore very crucial to send it without errors. In [16], a multiple description motion coding (MC) algorithm is proposed. This algorithm ensures robust delivery of the motion vector data to the destination. The performance of C is evaluated over UMTS channels in [17]. transmission over UMTS channels requires both good compression rates and effectiveness in the presence of channel failures[17]. Therefore C is a good solution for video over UMTS channels. Authors propose a balanced C scheme for 3D scan-based DWT (Discrete Wavelet Transform) video coding. The dispatch of the source video overhead between different channels is the novel property of this proposed scheme [17]. In the proposed scheme, authors adapt the redundancy between descriptors according to the channel model and time varying BER. They also compare their proposed scheme with SDC (Single Description Coding).The results in [17] show that C outperforms SDC for indoor, pedestrian and vehicular environments. With this proposed scheme, long bursty errors or the loss of an entire description does not have a catastrophic effect on the received video quality. A novel C scheme which is very well suited for video traffic over TCP is proposed in [18]. This system is based on coaction of source packetization strategy and TCP-friendly congestion control protocol [18]. The proposed source packetization strategy generates scalable layered video bitstreams while providing robustness against packet drops within the network. Authors strengthen their proposed algorithms by cooperating application and transport layers to maximize the expected delivered video quality at the receiver.the simulations in this paper show that the proposed system can gracefully tolerate and quickly react to sudden changes in the available connection capacity due to congestion [18]. The performance of wireless video conferencing using C is explored in [19]. The wireless video conferencing requires the transmission of highly compressed, delay sensitive streams over wireless links which are error prone [19]. These requirements are met through the use of C. A leaky prediction mechanism is proposed to recover from the errors very quickly. A rate control scheme based on Least Mean Square (LMS) algorithm is also proposed in this work. The main issue in voice communication is the tight time constraints. It is therefore very important to make

6 the number of retransmissions and packet losses as low as possible. In [20], an C system which adapts its number of descriptors to network loss conditions in order to conceal packet losses and retransmissions is proposed. Without increasing the transmission bandwidth, the results in [20] show that the packet loss problem is almost completely solved. IV. CONCLUSION AND OUTLOOK In this paper we have introduced and examined the performance of multiple description coding in combination with channel diversity for wireless communication systems. We have presented quality measurements for the received video quality. We have made a brief literature review on the performance of various C schemes. Consequently, we have shown that, this scheme provides scalable and very robust video transmission over lossy wireless environments and it is able to support heterogeneous terminals easily. We have also given the drawbacks of C in terms of overhead. ACKNOWLEDGMENT This paper is based upon work supported by SAMSUNG, Korea. [13] P. A. Chou, H. J. Wang, and V. N. Padmanabhan, Layered Multiple Description Coding, in Packet Workshop, Nantes, France, April [14] V. A. Vaishampayan, Design of Multiple Description Scalar Quantizers, IEEE Transactions on Information Theory, vol. 39, pp , May [15] X. Tang and A. Zakhor, Matching Pursuits Multiple Description Coding for Wireless, IEEE Transactions on Circuits and Systems for Technology, vol. 12, no. 6, pp , June [16] C. S. Kim and S. U. Lee, Multiple Description Coding of Motion Fields for Robust Transmission, IEEE Transactions on Circuits and Systems for Technology, vol. 11, no. 9, pp , September [17] M. Pereira, M. Antonini, and Michel Barlaud, Multiple Description Coding for UMTS, in Picture Coding Symposium, Saint Malo, France, April [18] R. Puri, K. W. Lee, K. Ramchandran, and Vaduvur Bharghavan, An Integrated Source Transcoding and Congestion Control Paradigm for Streaming in the Internet, IEEE Transactions on Multimedia, vol. 3, no. 1, pp , April [19] A. Sehgal, A. Jagmohan, and N. Ahuja, Wireless Conferencing Using Multiple Description Coding, IEEE Int l Symposium on Circuits and Systems, vol. 5, pp , May [20] B. W. Wah and D. Lin, Lsp-Based Multiple-Description Coding for Real-time Low Bit-Rate Voice Transmissions, in IEEE Int l Conf. on Multimedia and Expo, vol. 2, Swiss Federal Institute of Technology, EPFL, Lausanne, Switzerland, August 2002, pp REFERENCES [1] Y. Wang, M. T. Orchard, V. Vaishampayan, and A. R. Reibman, Multiple Description Coding Using Pairwise Correlating Transforms, IEEE Trans. Image Processing, vol. 10, pp , March [2] R. Prasad and S. Hara, Multicarrier Techniques for 4G Mobile Communications, ser. Universal Personal Communications. Artech House, [3] F. Fitzek, Application and Network Measurements for Multiple Description Coded Sequences, Aalborg University, Niels Jernes Vej 12, room A5 204 DK-9220 Aalborg Denmark, Tech. Rep. ISBN , ISSN , February [4] N. Gogate, D. M. Chung, S. S. Panwar, and Y. Wang, Supporting Image and Applications in a Multihop Radio Environment Using Path Diversity and Multiple Description Coding, IEEE Transactions on Circuits and Systems for Technology, vol. 12, no. 9, pp , September [5] Y. J. Liang, E. G. Steinbach, and B. Girod., Multi-stream Voice over IP Using Packet Path Diversity, in IEEE 4th Workshop on Multimedia Signal Processing, [6] J. G. Apostolopoulos, Error Resilient Compression Through The Use of Multiple States, in IEEE Int l Conf. on Image Processing, vol. 3, Vancouver, BC, Canada, September 2000, pp [7] M. Schaar and D. S. Turaga, Multiple Description Scalable Coding Using Wavelet-Based Motion Compensated Temporal Filtering, in IEEE Int l Conf. on Image Processing, vol. 3, Barcelona, Spain, September 2003, pp [8] J. G. Apostolopoulos and S. J. Wee. [9] J. G. Apostolopoulos, W. Tan, and S. J. Wee, Performance of a Multiple Description Streaming Media Content Delivery Network, in IEEE Int l Conf. on Image Processing, vol. 2, Rochester, New York, September 2002, pp. II 189 II 192. [10] I. K. Kim and N. I. Cho, Error Resilient Coding Using Optimal Multiple Description of DCT Coefficients, in IEEE Int l Conf. on Image Processing, vol. 2, Barcelona, Spain, September 2003, pp [11] A.C. Beǧen, Y. Altunbaşak, and Ö. Ergün, Multi-path Selection for Multiple Description Encoded Streaming, in IEEE Int l Conf. on Communications, vol. 3, Anchorage,Alaska,USA, May 2003, pp [12] A. C. Beǧen, Y. Altunbaşak, and Ö. Ergün, Fast Heuristics for Multi- Path Selection for Multiple Description Encoded Streaming, in IEEE Int l Conf. on Multimedia and Expo, Baltimore, Maryland, July 2003, pp