Error resilient stereoscopic video streaming using model-based fountain codes
Tan, A Serdar
Item Usage Stats
MetadataShow full item record
Error resilient digital video streaming has been a challenging problem since the introduction and deployment of early packet switched networks. One of the most recent advances in video coding is observed on multi-view video coding which suggests methods for the compression of correlated multiple image sequences. The existing multi-view compression techniques increase the loss sensitivity and necessitate the use of efficient loss recovery schemes. Forward Error Correction (FEC) is an efficient, powerful and practical tool for the recovery of lost data. A novel class of FEC codes is Fountain codes which are suitable to be used with recent video codecs, such as H.264/AVC, and LT and Raptor codes are practical examples of this class. Although there are many studies on monoscopic video, transmission of multi-view video through lossy channels with FEC have not been explored yet. Aiming at this deficiency, an H.264-based multi-view video codec and a model-based Fountain code are combined to generate an effi- cient error resilient stereoscopic streaming system. Three layers of stereoscopic video with unequal importance are defined in order to exploit the benefits of Unequal Error Protection (UEP) with FEC. Simply, these layers correspond to intra frames of left view, predicted frames of left view and predicted frames of right view. The Rate-Distortion (RD) characteristics of these dependent layers are de- fined by extending the RD characteristics of monoscopic video. The parameters of the models are obtained with curve fitting using the RD samples of the video, and satisfactory results are achieved where the average difference between the analytical models and RD samples is between 1.00% and 9.19%. An heuristic analytical model of the performance of Raptor codes is used to obtain the residual number of lost packets for given channel bit rate, loss rate, and protection rate. This residual number is multiplied with the estimated average distortion of the loss of a single Network Abstraction Layer (NAL) unit to obtain the total transmission distortion. All these models are combined to minimize the end-toend distortion and obtain optimal encoder bit rates and UEP rates. When the proposed system is used, the simulation results demonstrate up to 2dB increase in quality compared to equal error protection and only left view error protection. Furthermore, Fountain codes are analyzed in the finite length region, and iterative performance models are derived without any assumptions or asymptotical approximations. The performance model of the belief-propagation (BP) decoder approximates either the behavior of a single simulation results or their average depending on the parameters of the LT code. The performance model of the maximum likelihood decoder approximates the average of simulation results more accurately compared to the model of the BP decoder. Raptor codes are modeled heuristically based on the exponential decay observed on the simulation results, and the model parameters are obtained by line of best fit. The analytical models of systematic and non-systematic Raptor codes accurately approximate the experimental average performance.