Multiplexing Remote Driving Video Over Dual V2N Paths With Quality-of-Experience and Data Efficiency

초록

The purpose of this study is to tackle the critical challenge of delivering extremely reliable, low-latency video streaming for remote driving, where safety and maneuverability depend on uninterrupted and high-quality visual feedback. Reliance on a single communication path and static video frame encoding leaves remote driving systems exposed to fluctuations in network quality and to the inherent vulnerabilities of video inter-frame dependencies, often leading to degraded Quality of Experience (QoE) for human operators. The contribution of this work is a reinforcement learning (RL)-based video transmission framework that dynamically selects between dual communication paths while simultaneously adapting video encoding types. By employing an RL agent at the video sending side, the framework not only adjusts to rapidly changing network conditions but also opportunistically reduces data usage while preserving stringent QoE requirements. The findings from extensive experiments conducted over real-world 5G networks show that our framework achieves near-optimal reliability, remaining within 0.1% of full dual-path redundancy, while reducing data usage by up to 83%. In addition, the adaptive encoding strategy delivers up to 10.3% higher perceptual quality in SSIM compared to conventional all-I-frame dual-path transmission. Compared to a more practical alternative of selecting the minimum-latency path, SPAF reduced the unplayed frame ratio by up to more than 30%, respectively, while also achieving substantially higher visual quality (a 10.4% improvement in SSIM). These findings confirm the practicality and effectiveness of the proposed framework, underscoring its suitability for remote driving scenarios that can afford multiple communication paths.

키워드