A Scene-Aware Collaborative Optimization Method for Multi-Stream Transmission in End-Edge-Cloud Systems
摘要
随着智慧城市监控和智能交通等场景下视频数据的爆发式增长,大量帧级冗余信息给传输系统带来了巨大压力.传统集中式云计算架构在应对多路并发视频上传时,受限于带宽不足、带宽波动以及核心网拥塞,难以保证高效传输.现有自适应码率与压缩算法虽能降低码率,但仍依赖整帧传输,难以在兼顾低时延的同时保障关键区域质量.同时,多数研究聚焦于单视频流优化,忽视了人机视觉差异化需求及结构化数据在低带宽下的潜力.针对上述问题,本文提出一种基于场景感知的端边云多数据流协同优化传输方法.该方法首先构建了一种新颖的端边云多流协同传输架构.随后,将多流自适应传输建模为带约束的非凸组合优化问题,并引入场景感知状态划分模型,通过感知可用带宽及其波动,将复杂连续的网络环境映射为常规、干扰和恶劣三类典型场景,从而生成候选流集合与压缩率范围,有效降低优化维度.在此基础上,提出权重驱动的多流调度与压缩率联合优化方法,通过计算各流偏好得分与调度权重,自适应激活最优流类型,并在候选集合内动态配置传输策略与压缩率.同时结合拉格朗日对偶方法,实现高效求解,以最小化边云传输时延并保障视频质量.实验结果表明,本文所提方法在高压缩率下能够有效保持视频焦点区域结构相似性指标(Structural Similarity Index,SSIM)的图像质量,在直播场景中焦点区域平均 SSIM 达到 0.882,显著优于其他基线方法.在 4G/LTE 真实带宽数据集上,本文方法较各类基线方法传输时延降低 44.0%~85.4%;在 5G 带宽数据集构建的直播场景下,可以将传输时延降低 31.82%~63.57%.综上,本方法能够在复杂多变的网络条件下兼顾视频质量与低时延,确保端边云系统能够高效地处理大规模视频任务.
The explosive growth of video data in smart city surveillance and intelligent transpor-tation has placed heavy pressure on transmission systems due to massive frame-level redundancy.Traditional centralized cloud architectures struggle to support concurrent video uploads because of limited bandwidth,fluctuations,and core network congestion.Existing adaptive bitrate and com-pression methods can reduce bitrate but still rely on full-frame transmission,making it difficult to ensure both low latency and critical region quality.Furthermore,most studies focus on single-stream optimization,overlooking the differentiated needs of human-machine vision and the poten-tial of structured data under constrained bandwidth.To address these issues,this paper proposes a scene-aware collaborative optimization method for multi-stream transmission in End-Edge-Cloud systems.First,a novel End-Edge-Cloud multi-stream collaborative transmission architec-ture is constructed.The problem is then modeled as a constrained non-convex combinatorial opti-mization problem,where a scene-aware state partitioning model maps dynamic network condi-tions into normal,disturbed,and harsh scenarios.This reduces optimization complexity by gen-erating candidate stream sets and compression ranges.A weight-driven joint scheduling and com-pression strategy is then introduced,which adaptively activates optimal stream types and dynami-cally configures transmission policies.The Lagrangian dual method is applied for efficient sol-ving,minimizing latency while maintaining video quality.Experimental results show that the proposed method can effectively preserve the Structural Similarity Index(SSIM)quality of video focus regions under high compression ratios.In the live-streaming scenario,the average SSIM of focus regions reaches 0.882,significantly outperforming other baseline methods.On a real 4G/LTE bandwidth dataset,the proposed method reduces transmission latency by 44.0%-85.4%compared with various baselines.In the live-streaming scenario constructed using a 5 G bandwidth dataset,it further reduces transmission latency by 31.82%-63.57%.Overall,the proposed method achieves a favorable balance between video quality preservation and low-latency transmis-sion under complex and dynamic network conditions,providing effective support for edge-cloud systems to efficiently process large-scale video tasks.