Presentation Title

H.265 Video Traffic Prediction Using Neural Networks

Location

Room 2911

Session Format

Paper Presentation

Research Area Topic:

Engineering and Material Sciences - Electrical

Abstract

By the year 2019, the Cisco Virtual Networking Index predicts that video traffic of all types will constitute 80%-90% of global internet traffic, amounting to 1.24- 1.44 zettabytes of data passing through existing internet infrastructure. Video bandwidth forecasting systems can significantly improve the video transport mechanisms by limiting waste in Dynamic Bandwidth Allocation algorithms. With increasing popularity, high-definition videos such as 4K resolution media, even using high compression ration encoding standards such as H.265, are requiring new methods of predicting bandwidth requirements to sustain global network demand. Although the physical network infrastructure is growing rapidly, user demand is increasing exponentially, more rapidly than updated technology can be deployed, due to increasing commonality of large data transfers and high-definition video streaming. To provide increased data throughput on the network without allowing the physical limitations of the infrastructure to limit the network usage, highly adaptive neural-computational network models are used to predict the bandwidth requirements for transmission across a network and reduce idle time on the network. Network traffic can be structured to maximize efficiency for certain time-sensitive information, such as uninterrupted video streaming, by eliminating the majority of network idle time. We propose the use of non-linear, autoregressive (NAR) neural networks to predict the bandwidth requirements for streaming high-definition video encoded with the H.265 standard. We have chosen the NAR method as results have shown a high degree of accuracy in predicting the regression values of the network after training and testing. Our approach is divide a high-definition H.265 video trace file in half, with each half consisting of the beginning and end of the video. These traces, further separated by I, B, and P frame types, are then used to train and to test, respectively, the NAR network. The NAR network consists of the input layer with varying delays based on frame type, two hidden layers, and an output layer. Each hidden layer consists of varying numbers of neurons to determine the optimal network size to deliver an accuracy over 90% when tested with video trace files.

Presentation Type and Release Option

Presentation (Open Access)

Start Date

4-16-2016 1:30 PM

End Date

4-16-2016 2:30 PM

This document is currently not available here.

Share

COinS
 
Apr 16th, 1:30 PM Apr 16th, 2:30 PM

H.265 Video Traffic Prediction Using Neural Networks

Room 2911

By the year 2019, the Cisco Virtual Networking Index predicts that video traffic of all types will constitute 80%-90% of global internet traffic, amounting to 1.24- 1.44 zettabytes of data passing through existing internet infrastructure. Video bandwidth forecasting systems can significantly improve the video transport mechanisms by limiting waste in Dynamic Bandwidth Allocation algorithms. With increasing popularity, high-definition videos such as 4K resolution media, even using high compression ration encoding standards such as H.265, are requiring new methods of predicting bandwidth requirements to sustain global network demand. Although the physical network infrastructure is growing rapidly, user demand is increasing exponentially, more rapidly than updated technology can be deployed, due to increasing commonality of large data transfers and high-definition video streaming. To provide increased data throughput on the network without allowing the physical limitations of the infrastructure to limit the network usage, highly adaptive neural-computational network models are used to predict the bandwidth requirements for transmission across a network and reduce idle time on the network. Network traffic can be structured to maximize efficiency for certain time-sensitive information, such as uninterrupted video streaming, by eliminating the majority of network idle time. We propose the use of non-linear, autoregressive (NAR) neural networks to predict the bandwidth requirements for streaming high-definition video encoded with the H.265 standard. We have chosen the NAR method as results have shown a high degree of accuracy in predicting the regression values of the network after training and testing. Our approach is divide a high-definition H.265 video trace file in half, with each half consisting of the beginning and end of the video. These traces, further separated by I, B, and P frame types, are then used to train and to test, respectively, the NAR network. The NAR network consists of the input layer with varying delays based on frame type, two hidden layers, and an output layer. Each hidden layer consists of varying numbers of neurons to determine the optimal network size to deliver an accuracy over 90% when tested with video trace files.