Compared to traditional UDP-based video streaming such as RTP, HAS exhibits advantages such as easy firewall traversal, reuse of existing HTTP infrastructure, and adaptability.However,real-time UDP-based streaming still exhibits some advantages over HAS:a higher bandwidth utilization, the capability to use multicast to reduce the network load, alower start and channel-change delay, and a lower end-to-end delay for live streaming.To improve the QoE for HAS live streaming, a shorter segment duration could be used. This has following advantages:(1) Short segments are sooner available at the client because less time is required for encoding at the head-end and for transport. This results in a reduced live delay. (2)When the client starts a session, the first segment needs lesstime for transport, resulting in a reduced initial delay and afaster response to channel change. (3) At a channel change,there is less obsolete video data in the pipeline from serverto client that still needs to be received by the client. Asa result, the channel-change delay is further reduced. (4)Short segments increase the number of measuring and decisionpoints for the RDA. In case of deteriorating networkconditions, the RDA is able to react faster, possibly preventinga freeze.Unfortunately, it is generally infeasible to use super-short segments in current HAS systems because of the following reasons: (1) More HTTP-GET messages are required to retrievethe segments, resulting in a larger overhead for theserver and network. (2) Because more segments must beretrieved, more RTT cycles are lost between subsequent requests.This reduces the link utilization and the average quality. (3) In HAS, every segment starts with an IDR framein order to be decoded independently of other segments. For video coding standards such as H.264/AVC and High Effi-ciency Video Coding (HEVC), the encoding efficiency of an IDR frame is significantly lower than the efficiency of framescomposed of P or B slices. As a result, a higher bitrate isrequired to reach the same quality as for longer segments.A standard way to estimate this bitrate overhead requiredto achieve an equivalent average PSNR score over a set ofsegments is to measure the Bjontegaard Delta rate [5] forseveral test sequences. The estimated overheads for varioussegment lengths are reported in Section 5.However, the drawbacks of short segments can be mitigatedby applying additional methods. (1) The increasedoverhead of HTTP-GET messages can be avoided by usingthe full-push method. (2) Similarly, the problem caused bythe additional lost RTT cycles between subsequent segmentretrievals can be avoided by using full push since segmentsare pushed back-to-back. (3) The problem of the lost encodingefficiency could be mitigated by using quality-leveldependent encoding. Using this type of encoding, the highquality levels contain less IDR frames than the lower quality levels. Since all segments have the same duration in everylevel, not every segment in the highest level starts withan IDR frame. Consequently, upwards quality switching requiresa high quality segment that starts with an IDR frame.Downward quality switching can be done at every segmentsince every low-quality segment starts with an IDR frame.This type of encoding matches very well with the HAS characteristicssince a HAS client must be able to switch downquickly to avoid a freeze. However, clients act very conservativelywhen increasing the quality. Typically, a client waitsseveral seconds, evaluating the available bandwidth, beforeit decides to increase the quality. For now, quality-level dependentencoding was left as future work. Instead, we basedour evaluation on the typical HAS encoding scheme whereevery segment starts with an IDR frame.
đang được dịch, vui lòng đợi..