近年来，随着视频流媒体的流量迅猛增长，互动直播、视频分析等新型低时延视频应用层出不穷，深刻地改变着人们的工作、学习和生活方式。然而，这些应用在保障用户体验的同时，如何在性能和成本之间做出权衡，成为学术界和工业界广泛关注的重要问题。这给当前网络传输机制提出了严峻的挑战，包括多对多场景中上下行链路资源相互制约，视频应用的公平和性能难以兼顾，海量信息的传输和分析造成成本高昂，以及复杂动态环境中的资源适配难题。本文针对以上机遇和挑战，围绕低时延流媒体传输开展研究，主要贡献如下：（1）提出了面向多方交互的流媒体传输优化方案。针对现有交互式流媒体系统的用户体验质量不能令人满意的问题，提出了面向多方互动直播的流媒体系统架构，进而设计了基于用户体验的自适应码率调节算法，将多对多码率选择问题建模成非线性规划问题，并通过缓存区反馈调节来减少建模和测量过程中的系统误差。实验结果表明，该算法优于固定码率算法，有效提升了用户体验，端到端时延降低到100 ms。在真实直播平台部署后，该算法服务大量用户，取得了良好效果。（2）提出了兼顾公平和性能的流媒体传输优化方案。针对低时延流媒体应用的公平性问题，以典型视频应用为例，建立时延成因视图，探究其在性能和公平之间的取舍。进而提出兼顾公平和性能的流媒体传输机制，通过观测自身应用增减速率对链路丢包的影响来甄别丢包成因，有效地竞争瓶颈链路。实验结果表明，该机制可以实现比现有视频应用更好的用户体验，同时和参与的其他流很好地共存。（3）提出了面向实时视频分析的传输成本优化方案。针对现有视频分析系统中传输成本巨大的问题，提出端边协同的实时视频分析架构，将计算任务合理地分配到端边两侧协同执行，进而设计了时空域数据冗余消除的传输机制，在空域上使用目标块所在区域作为传输单元；在时域上只在目标块处于最佳位置时传输一次。实验结果表明，该方案在满足准确率和时延的约束下可节省90%的带宽成本。（4）提出了面向实时视频分析的计算成本优化方案。针对现有视频分析系统计算成本高昂的问题，提出基于无服务器平台的视频分析架构，用弹性算力适配弹性计算资源需求，进而设计了基于马尔可夫近似的视频参数和计算资源联合配置优化算法，根据视频内容和网络状况的动态变化更新求解相应的最佳配置。实验结果表明，该方案在满足准确率的约束下有效地降低了计算成本。
In recent years, with the rapid growth of the traffic of video streaming, new low-delay video applications such as interactive live streaming and video analytics have emerged one after another, profoundly changing the way people work, study and live. However, how to balance performance and cost while ensuring user experience in these applications has become an important issue of widespread concern in academia and industry. This poses severe challenges to the current network transmission mechanism, including mutual constraints on uplink and downlink resources in many-to-many scenarios, difficulty in balancing the fairness and performance of video applications, high costs caused by the transmission and analytics of massive information, and difficulty of resource adaptation in complex dynamic environments. In view of the above opportunities and challenges, this dissertation conducts research on low-delay video streaming transmission. The main contributions are as follows:(1) A multi-party interactive video streaming transmission optimization scheme is proposed. Aiming at the unsatisfactory user experience quality of the existing interactive streaming systems, a multi-party interactive live streaming architecture is proposed, and then an adaptive bitrate adjustment algorithm based on user experience is designed. The problem of many-to-many bitrate selection is modeled as a nonlinear programming problem, and the system errors in the modeling and measurement process are reduced by buffer feedback adjustment. The experimental results show that the proposed algorithm is superior to the fixed bitrate algorithm, effectively improves the user experience, and reduces the end-to-end delay to 100 ms. After the deployment of real live streaming platform, the algorithm serves a large number of users and achieves good results.(2) A video streaming transmission optimization scheme which takes into account fairness and performance is proposed. Aiming at the fairness problem of low-delay video streaming applications, a typical video application is taken as an example to establish a view of the causes of delay and explore its trade-off between performance and fairness. Furthermore, a streaming transmission mechanism that takes into account both fairness and performance is proposed, and the cause of packet loss can be identified by observing the influence of its own application rate on link packet loss, so as to effectively compete for bottleneck links. Experimental results show that the mechanism can achieve a better user experience than existing video applications, while co-existing well with other participating streams.(3) A transmission cost optimization scheme for real-time video analytics is proposed. In view of the huge transmission cost in existing video analytics systems, an end-to-edge collaborative real-time video analytics architecture is proposed, and computational tasks are reasonably allocated to both sides for collaborative execution. Then, a transmission mechanism for spatial-temporal data redundancy elimination is designed. In the spatial domain, the region where the target object is located is used as the transmission unit; in the temporal domain, the transmission is only performed once when the target object is in the best position. Experimental results show that the scheme can save 90% of the bandwidth cost under the constraints of accuracy and delay.(4) A computational cost optimization scheme for real-time video analytics is proposed. In view of the high computational cost of existing video analytics systems, a video analytics architecture based on the serverless platform is proposed, and elastic computational resources are used to adapt to elastic computational requirements. Then, a joint configuration optimization algorithm of video parameters and computational resources based on Markov approximation is designed. The algorithm updates the corresponding optimal configuration according to the dynamic changes of video content and network conditions. Experimental results show that the scheme effectively reduces the computational cost while satisfying the constraint of accuracy.
