This dissertation presents advances in end-to-end distortion estimation and optimization techniques for error-resilient video networking. Two main scenarios are considered: live video streaming, and streaming of pre-compressed video.;The first contribution of this dissertation is concerned with redundant encoding in live video streaming. The encoder generates two representations of the video---a primary and a secondary representation---at different quality levels, which both contribute to the overall end-to-end rate and distortion costs. After reviewing current redundant encoding schemes, several issues central to achieving good rate-distortion (RD) performance are identified. Previously proposed schemes address these through various heuristics. In this work, redundant encoding is formulated as a joint optimization problem of the parameters for the primary and secondary encoding. Performance of the optimization depends critically on accurate estimation of the end-to-end distortion, which is determined jointly by the primary and secondary transmission. An extension of the well-known "recursive optimal per-pixel estimate" (ROPE) is presented, enabling accurate end-to-end distortion estimation of redundantly coded data. The estimate is then integrated into the macroblock (MB) parameter selection process via RD optimization. Encoding decisions are adaptive at the MB level, as opposed to coarse packet-level redundancy (e.g., forward error correction). Three encoding strategies are presented, enabling different gain-complexity trade-off options.;In the second contribution, this work targets streaming of pre-compressed video. First, the previously proposed "first-order distortion estimate" (FODE) is reviewed. Based on a first-order Taylor expansion of the true distortion around the reference point of no packet loss rate (PLR), it provides good estimation accuracy at low to medium PLR. However, at medium to high PLR, or during adaptive delivery using unequal error protection (UEP), estimation accuracy deteriorates with the distance from the no-loss reference point. After identifying several issues adversely affecting the estimate, a new and more exible end-to-end distortion estimate is detailed. It is the combination of the FODE technique with a ROPE module, which was previously only applicable for estimation in live streaming. The new FODE-ROPE technique enables accurate estimation at arbitrary reference PLR for each packet, while maintaining low complexity. Integration into a simple UEP-based delivery system enables performance gains in pre-compressed streaming.;A common simplifying assumption in error-resilient video networking is that losses are independent and identically distributed. The final contribution is the extension of the FODE-ROPE technique to account for more complex channel statistics, in particular bursty channels. This extension enables accurate end-to-end distortion estimation for pre-compressed video in practical channels, thereby significantly enhancing the range of applications of the FODE technique. Exploitation of this estimate in an adaptive UEP delivery system achieves significant gains over existing delivery algorithms.
展开▼
