Next Generation Packet Video Networking with Applied AI

点击翻译按钮获取中文摘要

It is safe to say that packet video has already taken over the Internet, when considering that video traffic constituted 75% of all IP traffic in 2017, and will constitute 82% of all IP traffic by 2022. Sandvine reports that 6 of the top 7 biggest consumers of global IP traffic in 2018 were packet video systems: NetFlix (15% of all IP traffic), HTTP Media Streaming (13%), YouTube (11%), MPEG-TS (4.4%), HTTPS (not video, 4.1%), the video steaming protocol QUIC (3.9%), and Amazon Prime Video (3.7%). These video services are also known as Over-The-Top (OTT) video, because they directly offer video to consumers over the Internet, instead of the traditional distribution approaches of cable, broadcast or satellite. A portion of this traffic is from Internet gaming, such as cloud gaming and online gaming, taking about 7.8% of all Internet traffic in 2018. Despite their relatively fast market penetration, packet video systems still face fundamental challenges that prevent their continuous growth: Quality of Experience (QoE) which relates to the satisfaction of users with their video service experience, delay which is important for conversational application, and fairness among multiple users sharing the same bottleneck network link. To address these challenges, in this project we study, design, and develop Artificial Intelligence based methods to eliminate or significantly reduce unsatisfactory experience of users in packet video systems and networks. The vision behind this goal; i.e., Applied AI for packet video systems and networks, will in fact continue beyond the 5 years of this proposal, since AI applied to packet video systems is a relatively young research area with huge potential and impact. The said project goal decomposes into the following objectives: fairness among multiple OTT video users; efficient resource utilization within the cloud, which will also lead to lower delay for live video; system parameter estimation, such as bandwidth, delay, buffer levels, etc.; saliency-based video coding for higher Quality of Experience (QoE); and fault detection and root-cause analysis. Beyond the 5-year period, the vision will continue with applying AI to the following: traffic classification which helps us further improve QoE and resource utilization; self-organizing networks that intelligently make routing and operation decision; automatic recommender systems, once the root-cause has been detected, to aid the network operator to more quickly fix the problem, or in some cases for the system itself to autonomously repair itself; and congestion control, instead of widely-used exact TCP algorithms. The project will also contribute to the training of HQP, industrial collaboration, and support of outreach activities.