QoE-Driven Resource Allocation Optimized for Uplink Delivery of Delay-Sensitive VR Video Over Cellular Network

Junchao Yang, Jiangtao Luo, De Meng, Jenq-Neng HwangIEEE Access 2019 · DOI

Abstract

Uploading virtual reality (VR) video over cellular networks is expected to boom in the near future, as general consumers generate high-quality VR content with portable 360-degree cameras. However, the uplink bandwidth and delay requirements for tile-based VR video streaming present significant challenges. This paper proposes a novel scheme for uplink delivery where the encoding bitrate of each tile is determined by uplink resource allocation (RA). By jointly considering the Quality of Content (QoC) contribution of each tile and the channel quality of User Equipments (UEs), we formulate the RA problem as a frequency and time-dependent NP-hard problem. We introduce three algorithms, including a low-complexity "One-Shot" approximate convex algorithm, which achieves higher total Quality of Experience (QoE) for viewers.

While most research in Virtual Reality (VR) focuses on the downlink—streaming content from servers to users—the rise of User-Generated Content (UGC) on platforms like YouTube and Facebook has shifted the spotlight to the uplink. Uploading 360-degree video from a mobile device to the cloud is a critical bottleneck due to massive data sizes (often 4K+) and high frame rates.

System architecture for uplink delivery of user-generated VR content over cellular networks.

VR videos are uniquely sensitive to bandwidth constraints and latency. In live VR broadcasts, high latency can ruin the immersive experience. Current systems often treat all video data equally, but in a 360-degree sphere, not all regions are equally important to the viewer. This paper addresses how to intelligently allocate limited cellular radio resources to the most "important" parts of a VR video during the upload process to maximize the final viewer's experience.

The core of the proposed method is a tile-based uplink delivery scheme. Instead of treating the video as a single monolithic file, the Equirectangular Projection (ERP) frame is divided into a grid of independent tiles. This allows the network to prioritize specific regions based on their visual importance.

QoC (Quality of Content) Weighting

Each tile is assigned a weight based on two primary factors:

  1. Saliency: Using saliency detection to identify regions likely to attract a viewer's attention (e.g., moving objects or faces).
  2. Vertical Position: Prioritizing the "equatorial" region of the 360-view, as viewers rarely spend time looking at the extreme poles (top or bottom) of the sphere.
Saliency map overlaid on a tiled 360-degree frame, used to determine Quality of Content (QoC) weights for resource allocation.

By integrating these weights into the cellular scheduler (the eNodeB), the network can identify which packets are more critical for human perception and prioritize them during periods of congestion.

Resource Allocation Algorithms

The resource allocation problem is formulated as a non-deterministic polynomial (NP)-hard optimization task. The authors compare three distinct approaches:

  • Greedy Algorithm: Iteratively assigns resource blocks (RBs) to the tile providing the highest utility gain. While it offers high performance, it is computationally expensive.
  • Two-Stage Algorithm: A naive approach that allocates resources to users first, then to tiles. This is fast but results in suboptimal quality.
  • One-Shot Algorithm (Novel): This is the paper's primary contribution. It relaxes the complex integer programming problem into a continuous convex problem, solves it using a convex optimization toolbox (CVX), and then maps the results back to discrete resource blocks. This allows for near-greedy performance with significantly lower computational overhead.

Experimental Results

The study utilized the STJU immersive video sequence Dataset, comprising nine uncompressed 360-degree video sequences. Simulations were conducted using an LTE-compliant radio module in MATLAB.

Key findings include:

  • Utility Gains: The proposed One-Shot and Greedy algorithms achieved at least 17% higher system utility compared to standard Proportional Fairness (PF) methods.
  • QoE Improvement: When measuring "Viewport-Weighted Utility" (the quality actually perceived by the viewer), the proposed schemes showed a 25-26% improvement over baseline methods.
  • Computational Efficiency: The One-Shot algorithm completed the optimization in just 0.67 seconds, compared to 5.78 seconds for the Greedy algorithm. This makes it viable for real-time cellular scheduling where decisions must happen in milliseconds.

Conclusion

The quality of the uplink serves as the "base version" for VR content; if a tile is uploaded at a low bitrate, no downlink optimization can recover that lost detail. By utilizing saliency-aware scheduling and the efficient One-Shot algorithm, cellular networks can significantly enhance the QoE for VR viewers. Future work will explore the impact of saliency dispersion and the integration of real-time viewport prediction from the uploader's perspective.

Summary updated

BibTeX
@article{yang2019qoe,
  author = {Junchao Yang and Jiangtao Luo and De Meng and Jenq-Neng Hwang},
  title = {{QoE-Driven Resource Allocation Optimized for Uplink Delivery of Delay-Sensitive VR Video Over Cellular Network}},
  journal = {IEEE Access},
  year = {2019},
  doi = {10.1109/access.2019.2915370},
}
Figure