面向多设备协同场景的实时视频流分析系统

Toward cooperative multi-agent video streaming perception

实时视频流分析在智能监控、智能制造、自动驾驶等场景中具有重要价值,然而其存在计算负载高、带宽需求大和延迟要求严格等特点,难以通过传统的本地计算模式或者云计算模式进行部署.近年兴起的边缘计算范式,将复杂的计算任务从终端设备上传到物理临近的边缘服务器上,能够有效解决设备层面的部署问题.然而,例如无人机编队飞行、车队自动驾驶和多机器人协同等不断涌现的多设备协同场景,新增了系统层面的综合性能要求,包括智能分析的实时准确率、设备之间的性能一致性和系统容纳的设备数量上限.当前的边缘计算范式对多设备协同场景的优化尚显不足,未能有效解决设备之间对上传带宽和服务器算力的竞争问题,所以难以满足这类场景的要求.本文设计了MASSIVE系统,能够在多设备协同场景中,全面提升实时视频分析的综合性能.首先,MASSIVE系统提出了适合多设备协同场景中度量视频流分析系统综合性能的评价体系.其次,MASSIVE系统设计了帕累托改进调度器来计算帕累托最优的系统调度策略,使得系统在3个维度上同时取得了相比已有系统更好的性能表现.最后,MASSIVE设计了虚拟流量整形器来保证各个设备在无线网络中按照调度策略上传视频流数据.实验结果表明,MASSIVE在多种典型的视频分析任务中,相比于当前的代表性系统,至少达到了122.7%的实时准确率、1.8倍的系统容量和更好的系统一致性,并达到了帕累托最优.

Video streaming perception ability is critical for AI applications on resource-constrained devices(agents),which prefers to offload video streams from devices to edge servers for real-time inference by deep neural networks(DNNs).Meanwhile,the multi-agent system(MAS) community is attempting to run DNNs on multiple cooperative agents to enable improved swarm intelligence-based tasks(e.g.,drone swarm intelligence,selfdriving fleet collaboration,and multi-agent robot cooperation).However,transferring video streaming perception capability from single-agent systems to MASs is extremely difficult due to spontaneous competition-induced trade-offs between the desired goals of accuracy,consistency,and capacity,which are three critical but conflicting measuring indexes.In this paper,we present the design and implementation of MASSIVE,an edge-assisted cooperative multi-agent video streaming perception system that simultaneously achieves all three desired goals.In our design,we consider the performance characteristics of video streaming perception and the insight of its periodic offloading pattern.On this basis,we develop a Pareto improvement scheduler to eliminate spontaneous competition among agents,allowing multi-objective optimization to achieve an ideal Pareto optimal state.Finally,we propose a virtual traffic shaper based on the mainstream 802.11 MAC protocol to ensure deterministic periodic video stream offloading in an uncertain wireless network.Our experiments demonstrate that MASSIVE achieves 122.7% accuracy and 1.8x capacity compared to the closest baseline on multiple actual cooperative vision tasks with even better consistency,and achieves an ideal Pareto optimal state in a wireless environment.