Proactive Scheduling and Caching for Wireless VR Viewport Streaming

TL;DR

This paper presents a proactive wireless VR streaming system that uses deep learning to predict user orientation, enabling efficient viewport caching and scheduling, significantly reducing latency and improving delivery success under challenging network conditions.

Contribution

It introduces a novel proactive VR streaming framework combining neural network-based prediction with viewport-aware caching and scheduling, enhancing performance in wireless VR applications.

Findings

Latency reduced by over 80%

Delivery success rate close to 100%

Significant performance gains in fluctuating channels

Abstract

Virtual Reality (VR) applications require high data rate for a high-quality immersive experience, in addition to low latency to avoid dizziness and motion sickness. One of the key wireless VR challenges is providing seamless connectivity and meeting the stringent latency and bandwidth requirements. This work proposes a proactive wireless VR system that utilizes information about the user's future orientation for proactive scheduling and caching. This is achieved by leveraging deep neural networks to predict users' orientation trained on a real dataset. The 360{\deg} scene is then partitioned using an overlapping viewports scheme so that only portions of the scene covered by the users' perceptive field-of-view are streamed. Furthermore, to minimize the backhaul latency, popular viewports are cached at the edge cloud based on spatial popularity profiles. Through extensive simulations, we show that the proposed system provides significant latency and throughput performance improvement, especially in fluctuating channels and heavy load conditions. The proactive scheduling enabled by the combination of machine learning prediction and the proposed viewport scheme reduces the mean latency by more than 80% while achieving successful delivery rate close to 100%.