摘要
根据云计算平台下智能视频分析的实时性需求,设计一个基于Storm流计算框架的实时视频分析系统。采用短照片组进行视频流分割,损失部分视频传输带宽以降低整体延时,采用合并解码单元和视频分析单元的方法避免耗尽集群带宽,并利用工作窃取机制加速算法执行。通过节点性能监控并利用贪心置换策略动态调节Worker Node负载,改进Storm的默认调度器,降低消息的处理延时。实验结果表明,在运行人脸检测算法的Storm集群中接入多路监控设备,实现100ms之内的消息处理延时和低于1s的整体延时,能够为云环境下多路监控终端提供实时稳定的视频分析服务。
On the demand for real-time intelligent video analysis,this paper proposes a real-time video analysis system based on Storm stream computing platform.It uses short group of pictures splitting video streams to low total latency while sacrifices some video transfer bandwidth,avoiding the exhaustion of cluster bandwidth by incorperate video decode unit with video analysis unit,and speeding up video algorithm by work-stealing mechanism.Haing reduces tuple latency by improving the default scheduler of storm for load monitor and dynamic adaption of loads using greedy and permulation strategy.Experimental results show the system can achieve 100 ms tuple processing delay and less than1 s total delay for multiple channel terminals in the Storm cluster running face detection,and can provide real-time and stable video analysis service for multiple terminals under cloud environment.
出处
《计算机工程》
CAS
CSCD
北大核心
2015年第12期26-29,35,共5页
Computer Engineering
基金
国家自然科学基金资助项目(40927001)
关键词
视频监控
Storm平台
流计算
智能分析
实时视频分析
集群
video surveillance
Storm platform
stream computing
intelligent analysis
real-time video analysis
cluster
