面向智慧城市的视频边缘协同计算理论与关键技术研究

The amount of video data grows exponentially. The centralized way that videos are streamed to and processed at a cloud becomes inefficient and problematic and cannot satisfy the need of a smart city for large-scale video analytics. The reason is two-fold. First, the bandwidth of a cloud strictly constrains the incoming video streams. Second, the computational capability of a cloud is far from enough to perform large-scale video analytics using deep learning. The increase of computational capability of end devices and the emerging of edge computing makes it possible to perform video analytics elsewhere, not only on the cloud. In this research, we propose cooperative edge computing where end (i.e., camera), edge, and cloud cooperatively perform video analytics. The idea is to perform video analytics on the way a video is streamed from end to cloud via edge, i.e., reallocating computation from cloud to end and edge. Such a way will significantly reduce data transmission and accelerate video analytics. To enable cooperative edge computing for video analytics using deep learning, this research focuses on adaptive computing of convolutional neural networks, modeling compute time and energy of convolutional neural networks, the scheme of multi-level cooperative edge computing, and the design and implementation of cooperative edge computing systems for video surveillance.

视频数据正在以几何倍数增长。将视频汇聚到云并做处理分析的集中式计算越来越无法满足智慧城市对大规模视频处理分析的需求。其原因有两点：首先由于网络带宽是一定的，能同时接入云的视频流是非常有限的；其次云很难满足基于深度学习的大规模视频处理的计算需求。终端越来越强的计算能力和边缘计算的兴起使得云已经不是进行视频处理分析的唯一选择。本课题提出面向智慧城市的视频边缘协同计算，其原理是在视频传输过程中（从终端到边缘到云）就对其进行处理，且处理分析后只传输结果或者中间结果。这样不仅可以减少视频数据传输量，同时可以利用终端、边缘的计算能力加速视频处理分析。为了实现视频边缘协同计算，本课题开展了复杂度感知的卷积神经网络自适应计算，异构平台的卷积神经网络时延和能耗模型，多层次边缘协同计算机制以及智慧城市视频监控应用的研究。视频边缘协同计算将大幅提升基于深度学习的大规模视频处理分析的吞吐量，满足智慧城市的需求。

视频数据正在以几何倍数增长。将视频汇聚到云并做处理分析的集中式计算越来越无法满足智慧城市对大规模视频处理分析的需求。其原因有两点：首先由于网络带宽是一定的，能同时接入云的视频流是非常有限的；其次云很难满足基于深度学习的大规模视频处理的计算需求。终端越来越强的计算能力和边缘计算的兴起使得云已经不是进行视频处理分析的唯一选择。本项目研究了面向智慧城市的视频边缘协同计算理论与关键技术，包括异构平台的神经网络计算时延和能耗模型，基于神经网络的视频自适应计算，多层次边缘协同计算机制，构建了两个面向智慧城市的视频协同计算系统，并探索了基于强化学习的计算节点自主协同。项目通过算法分析和实验验证了视频边缘协同计算可以大幅提升基于深度学习的大规模视频处理分析的吞吐量，满足智慧城市中视频处理需求。项目的研究成果，包括算法和系统设计有潜力被用于智慧城市中的端边云协同计算。