面向智慧高速的视频云多任务处理架构研究

打开文本图片集

中图分类号：TP311 文献标识码：A 文章编号：2096-4706（2025）22-0092-06

Research on Multi-task Processing Architecture of Video Cloud for Smart Highway

YANG Bei，LU Chunjing， WANG Quansong， CHEN Shuang （1.Guizhou Zhongnan Transport TechnologyCo.，Ltd.，Guiyang 550o18，China; 2.GuizhouDoorTo Time Science and TechnologyCo.，Ltd.，Guiyang550081，China; 3.Guizhou Institute ofTechnology，Guiyang 550025，China）

Abstract： With the construction ofsmart highways，existing video cloud platforms face higher requirements in terms of massive dataprocessingandreal-timemulti-task processngcapabilities.Inordertoenhancethedata processingcapabilities of videocloud platforms，andthecapabilitiesofroadcondition monitoring，early warning，andemergencyevent aalysisand decision support，this paper studies areal-time video stream mult-task processingarchitecture basedonhardware aceleration. ThisarchitectureusesGPUadwareacceleratioattheedgeforreal-timetranscoding，compressionndlghtweightetection model operations，and pushes thedetectionresults tothecloud for further procesingand analysis.It supports concurrent task procesingofmultiplevideostreams.Thedetectionalgorithmontheedgesideadoptsanend-to-enddetectionalgorithm combined with SORT tracking，and is optimized based onthecharacteristicsof Cambrian chips.Itachieves paralel processing based on Tensor Cores and reduces memory usage through Layer Fusion.The architecture adopts ONNX format compatibility， INT8 quantization+pruning forspeedimprovement，andDocker +Kubermetes forcontainerizeddeploymentand elastic scaling， providing an efficient real-time multi-task stream processing solution for video cloud platforms.

Keywords：multi-card parallel processing; CambrianMLU;Tensor Core;modeloptimization;containerizeddeployment

0 引言

为满足交通管理需求，各地正建设省级视频云平台，实现全国高速公路视频资源集中管理与共享，提升视频监测覆盖率和智能化水平，进而增强运营效率和服务质量。（剩余6998字）