摘要
行为识别技术具有巨大的应用前景和潜在的经济价值,广泛应用于视频监控、视频检索、人机交互、公共安全等领域.图卷积网络表现出基于图数据的依赖关系进行建模的强大功能,成为行为识别领域的研究热点.基于此,主要概述基于图卷积网络的行为识别方法.图卷积网络主要有两大方法:基于频谱的方法和基于空间的方法.首先,从不同侧面分析两种方法的优缺点,概述两种方法在行为识别领域的应用与发展;然后,根据行为识别中图网络模型和算法设计的差异,总结网络构造的关键方面,对比不同算法对模型性能产生的影响;最后,针对图卷积网络在行为识别中存在的问题,对未来图卷积网络的发展进行展望.
Action recognition technology has great application prospects and potential economic value,and is widely used in video surveillance,video retrieval,human-computer interaction,public security and other fields.Graph convolutional networks show the powerful function of modeling based on graph data dependency,which have become a research hotspot in the field of action recognition.This paper mainly summarizes action recognition methods based on graph convolutional networks.There are two main methods of graph convolutional networks:the spectral-based method and the spacial-based method.Firstly,for the two methods,this paper analyzes advantages and disadvantages from different aspects,summarizes their application and development in the field of action recognition.Then,according to the differences of the design of graph network models and algorithms in action recognition,key aspects of network construction are summarized,and the influence of different algorithms on model performance is compared.Finally,according to the problems existing in the action recognition based on graph convolutional networks,future development of graph convolutional networks is prospected.
作者
孔玮
刘云
李辉
王传旭
KONG Wei;LIU Yun;LI Hui;WANG Chuan-xu(College of Information Science and Technology,Qingdao University of Science and Technology,Qingdao 266061,China)
出处
《控制与决策》
EI
CSCD
北大核心
2021年第7期1537-1546,共10页
Control and Decision
基金
国家自然科学基金项目(61702295,61672305)。
关键词
图卷积网络
行为识别
神经网络
深度学习
非欧氏空间
graph convolutional networks
action recognition
neural networks
deep learning
non-Euclidean space
