Adaptive Layered Video Coding for Multi-time Scale Bandwidth Fluctuations

Quality of service of network video is determined by both the amount of available bandwidth and its temporal uctuations. Except for the premium service, other proposed Di Serv tra c classes such as assured and best e ort services are not immune to temporal bandwidth uctuations. We observe that uctuations in shared packet switching networks occur in di erent time scales, from less than a microsecond up to the order of minutes and longer. In the fast time scale (microseconds), bandwidth uctuations are caused by bit errors, cell loss, and packet corruption, which are associated with noise in physical devices and communication channels. In the medium time scale (milliseconds to seconds), uctuations are often originated from link layer operations such as routing, queueing, multiplexing and scheduling of communication links. In the slow time scale (seconds to minutes), uctuations are caused by connection migration among networks with di erent speeds (e.g. LAN vs. wireless). While temporal uctuations in the Di Serv model mainly happen at the medium time scale, in this paper we broaden the scope to investigate layered video coding techniques which adapt to multi-time scale uctuations of connection bandwidth, with emphasis on the medium time scale. We review prior work in dealing with fast and slow time scale uctuations and argue additional work is required for medium time scale in order to provide a complete solution. We propose delay cognizant video coding and demonstrate that it adapts e ectively to medium term bandwidth uctuations. Delay cognizant video coding segments and compresses video information into layers with variable delay tolerance. When a network link is congested, delay sensitive layers are prioritized over delay tolerant layers, whose packets are queued. Delay tolerant layers resume their transmission as soon as the congestion is cleared. Late arrivals from delay tolerant layers are displayed, according to decoding rules, along with delay sensitive data at the next picture refresh. We found the amount of tolerable delay exceeds 400 milliseconds for videoconferencing type sequences. Delay cognizant video coding complements prior work in layered coding for bandwidth adaptation. With our contribution, layered video now o ers a complete solution for adapting to multi-time scale uctuations.