Content-based VBR Video Tra c Modeling and its Application to Dynamic Network Resource Allocation

In bandwidth limited networks and network interfaces, dynamic resource allocation can substantially increase the link utilization and also decrease the required network bu ering. In general, there are two important tradeo s e ecting the network utilization. First, tradeo between e ciency of the real-time dynamic resource allocation and the policy controlling the renegotiation frequency. Second, tradeo between accuracy of stream resource prediction and prediction delay. The latter is the focus of our work. We propose a new Content-based Video Tra c Model which uses the visual content as an important indicator of the video stream resource requirements. The model has two components: the visual content is characterized by Video Structure Model and the particular compression mechanism is described by Scene Resource Model. The object-based video content classi cation scheme, rst introduced in our work, maps individual video scenes into their bandwidth resource requirements. We validate our classi cation scheme for the MPEG-2 stream using a real 0.5 hour VBR MPEG-2 video trace. Also, using a trace-driven network simulator we show main advantages of our new technique: improvement of network utilization. From the results we nd that while the performance of o -line contentbased dynamic network resource allocation scheme shows marginal improvement compared to existing approaches (e.g., RVBR), its on-line performance shows signi cant improvements (55% 70%).