摘要
为了减少智慧城市场景下多摄像头实时视频数据的处理量,提出了基于机器学习算法的边缘端视频协同分析方法。首先,针对各摄像头检测到的重要目标物体,设计了不同的关键窗口来筛选视频的感兴趣区域,缩减视频数据量并提取其特征。然后,根据提取的数据特征,对不同摄像头视频中的相同目标物体进行标注,并设计了摄像头之间关联程度值的计算策略,用于进一步缩减视频数据量。最后,提出了基于图卷积网络和重识别技术的GC-ReID算法,旨在实现多摄像头视频协同分析。实验结果表明,与现有的视频分析方法相比,所提方法能够有效降低系统时延和提高视频压缩率,并保证较高的准确率。
In order to reduce the processing volume of multi-camera real-time video data in smart city scenarios,a video collaborative analysis method based on machine learning algorithms at the edge was proposed.Firstly,for the important objects detected by each camera,different key windows were designed to filter the region of interest(RoI)in the video,reduce the video data volume and extract its features.Then,based on the extracted data features,the same objects in the videos from different cameras were annotated,and a strategy for calculating the association degree value between cameras was designed for further reducing the video data volume.Finally,the GC-ReID algorithm based on graph convolutional network(GCN)and re-identification(ReID)was proposed,aiming at achieving the collaborative analysis of multi-camera videos.The experimental results show that proposed method can effectively reduce the system latency and improve the video compression rate while ensuring the high accuracy,compared with the existing video analysis methods.
作者
期治博
杜磊
霍如
杨帆
黄韬
QI Zhibo;DU Lei;HUO Ru;YANG Fan;HUANG Tao(State Key Laboratory of Networking and Switching Technology,Beijing University of Posts and Telecommunications,Beijing 100876,China;Department of Industrial Internet Institute,China Academy of Information and Communication,Beijing 100083,China;Faculty of Information Technology,Beijing University of Technology,Beijing 100124,China;Purple Mountain Laboratories,Nanjing 211111,China)
出处
《通信学报》
EI
CSCD
北大核心
2023年第8期14-26,共13页
Journal on Communications
基金
国家重点研发计划基金资助项目(No.2018YFB1800602)
2020年工业互联网创新发展工程基金资助项目(工业互联网标识资源搜索系统)。
关键词
边缘计算
机器学习
视频协同分析
感兴趣区域标注
多摄像头关联性
edge computing
machine learning
video collaborative analysis
region of interest annotation
association between cameras
