摘要
随着智慧高速公路的建设,现有的视频云平台在海量数据处理与多任务实时处理能力方面面临更高的要求。为了提升视频云平台的数据处理能力和道路路况监控、预警以及应急事件分析与决策支持能力,文章研究了一种基于硬件加速的实时视频流多任务处理架构,在边缘端利用GPU硬件加速进行实时转码、压缩和轻量化检测模型,并将检测结果推送至云端进行进一步处理和分析,该架构支持多视频流并发任务处理。边缘侧的检测算法采用端到端检测算法结合SORT跟踪并结合寒武纪芯片的特性进行优化。基于Tensor Core并行与Layer Fusion降内存。采用ONNX格式兼容、INT8量化+剪枝提速以及Docker+Kubernetes容器化部署与弹性扩展,为视频云平台提供高效实时多任务流处理方案。
With the construction of smart highways,existing video cloud platforms face higher requirements in terms of massive data processing and real-time multi-task processing capabilities.In order to enhance the data processing capabilities of video cloud platforms,and the capabilities of road condition monitoring,early warning,and emergency event analysis and decision support,this paper studies a real-time video stream multi-task processing architecture based on hardware acceleration.This architecture uses GPU hardware acceleration at the edge for real-time transcoding,compression,and lightweight detection model operations,and pushes the detection results to the cloud for further processing and analysis.It supports concurrent task processing of multiple video streams.The detection algorithm on the edge side adopts an end-to-end detection algorithm combined with SORT tracking,and is optimized based on the characteristics of Cambrian chips.It achieves parallel processing based on Tensor Cores and reduces memory usage through Layer Fusion.The architecture adopts ONNX format compatibility,INT8 quantization+pruning for speed improvement,and Docker+Kubernetes for containerized deployment and elastic scaling,providing an efficient real-time multi-task stream processing solution for video cloud platforms.
作者
杨焙
卢春静
王全松
陈双
YANG Bei;LU Chunjing;WANG Quansong;CHEN Shuang(Guizhou Zhongnan Transport Technology Co.,Ltd.,Guiyang 550018,China;Guizhou Door To Time Science and Technology Co.,Ltd.,Guiyang 550081,China;Guizhou Institute of Technology,Guiyang 550025,China)
出处
《现代信息科技》
2025年第22期92-97,共6页
Modern Information Technology
