Perceptual-Based Real-Time Cloud Game Streaming

Sammanfattning

Cloud gaming has emerged as a popular platform for interactive entertainment, but its widespread adoption is hindered by bandwidth limitations, particularly in regions where average internet speeds fall below the minimum bandwidth requirements for popular cloud gaming solutions. Traditional bitrate control mechanisms, such as CBR, ABR, and CRF, focus primarily on technical metrics, often neglecting user-centric quality metrics based on perceptual parameters. This thesis explores the feasibility of implementing a real-time encoding bitrate control mechanism for cloud gaming that prioritizes perceptual quality by leveraging the ITU-T Quality of Experience (QoE) estimation model P.1204.3. The proposed system integrates three key components: a screen capturing and encoding/decoding software that uses NVENC and NVDEC for real-time video encoding and decoding, a simplified version of the P.1204.3 model reference implementation for per-second quality scoring, and a reimplementation of a video parser capable of extracting metadata in real-time using a modified instance of FFmpeg. The system adjusts the encoding bitrate based on the P.1204.3 score, aiming to maintain a target quality level of 4.0 (± 0.2) by iteratively adjusting the bitrate in response to the perceived quality. Evaluation was conducted on a set of 12 4K and 4 FHD video game clips, each 10 seconds long, featuring diverse content and graphics. The control mechanism successfully managed to maintain the target quality for 4K content, although it slightly undershot the target for FHD videos. A key achievement of this implementation is its high adaptability, as it requires no input parameters beyond the video bitstream and a target quality level, and finds the optimal bitrate for maintaining this quality level. Despite the adaptability, the system's performance is constrained by the need to sequentially process video one second at a time, which limits its responsiveness to real-time changes. Comparisons with traditional bitrate control methods (CBR, ABR, and CRF) highlight the potential of this approach to achieve more efficient bandwidth usage while maintaining user-perceived quality. This work demonstrates the viability of a perceptually-driven bitrate control mechanism for cloud gaming, marking a significant step toward more user-centric streaming solutions. However, further optimization is required before it can be deployed in a live cloud gaming environment. The findings suggest that this approach could revolutionize bandwidth management in cloud gaming, ensuring high-quality experiences even under variable network conditions. As cloud gaming continues to grow, solutions like this will be essential to making high-quality gaming accessible and sustainable across diverse global contexts.