Back to results list
Please use this identifier to cite or link to this item:
|Title:||Towards a QoE-aware video streaming system||Authors:||Mok, Ka-pui||Advisors:||Chang, Rocky (COMP)||Keywords:||Streaming technology (Telecommunications)
|Issue Date:||2016||Publisher:||The Hong Kong Polytechnic University||Abstract:||Online video streaming is one of the most popular web application nowadays. Billions of users watch videos on YouTube, while Netflix has 60 millions of subscribers around the world. These video service providers use HTTP streaming to transfer video data to the users over the existing web architecture. HTTP streaming gains its popularity rapidly, because it can help video data traverse the firewall and the Network Address Translation(NAT).The Content Delivery Network (CDN) can also be leveraged to largely increase the scalability. HTTP Adaptive Streaming (HAS),a newer version of HTTP streaming, can support video bitrate adaptation, which can adjust the video bitrate according to the network condition. Although video service providers can easily obtain the network Quality of Service (QoS) data, ordinary Internet users do not have sufficient knowledge about the network QoS.They concern more about their Quality of Experience (QoE), which represents their overall perceived quality. However,it is very challenging for service providers to evaluate the QoE because of its the subjective nature. In this research, we contribute to three aspects of the QoE of HTTP video streaming systemQoE measurement,QoE improvement, andQoE assessment. QoE measurement is to understand the impact from the network quality on the QoE of HTTP video streaming. We investigate the relationships among the network QoS, application layer events, and the QoE. A set of Application Performance Metrics (APM) is proposed to quantify the application layer events. The APM is further correlated with the QoE collected from the subjective assessments. Our results show that the rebuffering events canadversely affect the QoE. Our follow-up studies further show that both the initial and the abrupt change of video bitrate in HTTP adaptive streaming can affect the QoE. The MOS can be reduced by 17% if a sub-optimal initial bitrate is chosen.
One of the key issues leading to sub-optimal QoE in HTTP adaptive streaming is that the streaming system lacks accurate network measurement data to support the selection of video bitrate. This can result in rebuffering events or unnecessary bitrate switching. However, it is challenging to perform lightweight and accurate network measurement on the clients' browsers. To improve the QoE, we propose a server-side measurement paradigm, which executes the main measurement logic in a middlebox installed in front of the streaming server. With this framework, we design and implement two systems, IRate and QDASH, to support measurement before and during the video stream (i.e., pre-stream and mid-stream stages), respectively. IRate exploits the pre-stream time window to probe the network, so that a quick estimation of the network condition can be performed, and the best initial video bitrate can be estimated at the onset of the streaming. Our results show that IRate can achieve an accuracy of 80% by performing 10s of measurement. After the video streaming starts, QDASH hijacks the video flow to carry out inline measurement. By carefully designing the packet sending order and rate, we can conduct packet train-based available bandwidth measurement and estimate the video bitrate the network can support. We compare the amount of time required for obtaining the correct throughput between QDASH and throughput averaging method commonly used in video players. Our testbed experiment shows that QDASH can respond quicker than the harmonic mean of throughput data for at least 5s. The final part of this research is on enhancing the scalability and reliability of QoE assessments by employing user behavior analytics. Traditional subjective assessment in a controlled environment using Mean Opinion Score (MOS) does not scale. Distributing a set of customized video can increase the participation, but the subjective assessment alone has its limitations. Crowdsourcing is aimed at further scaling the QoE measurement. However, screening out low-quality workers is an inherent and unsolved challenge for this approach. Subjective assessment is also hard to be conducted in real-world environment,because users are not responsive to the assessment. We propose user-behavior analytics to improve the video QoE assessments. User behaviors, such as pause events, mouse click events and cursor trajectory, contain rich information reflecting users' cognitive processes. We record and analyze these user behaviors while they review videos in customized video players or crowdwsourcing platform. We found that these user-behavior data can significantly improve the explanatory power of QoE model for the MOS by 8%. Our worker behavior based approach can detect around 80% of low-quality users in crowdsourcing platforms.
|Description:||PolyU Library Call No.: [THS] LG51 .H577P COMP 2016 Mok
xx, 197 pages :illustrations (some color)
|URI:||http://hdl.handle.net/10397/55634||Rights:||All rights reserved.|
|Appears in Collections:||Thesis|
Show full item record
Files in This Item:
|b29041570_link.htm||For PolyU Users||208 B||HTML||View/Open|
|b29041570_ira.pdf||For All Users (Non-printable)||2.43 MB||Adobe PDF||View/Open|
Citations as of Jul 15, 2019
Citations as of Jul 15, 2019
Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.