视频中动作识别任务综述  

A survey of action recognition in videos

作　　者：卢修生[1] 姚鸿勋[1] LU Xiusheng;YAO Hongxun(School of Computer Science and Technology,Harbin Institute of Technology,Harbin 150001,China)

机构地区：[1]哈尔滨工业大学计算机科学与技术学院,哈尔滨150001

出　　处：《智能计算机与应用》2020年第3期406-411,共6页Intelligent Computer and Applications

基　　金：国家自然科学基金面上项目(61772158,61472103)。

摘　　要：随着移动通讯技术的发展,网络上视频数据呈爆炸性增长,对于智能视频分析技术的需求日益增加。随着深度学习技术的应用,视频理解和分析领域近年来得到了快速发展。作为视频分析领域的核心任务,对动作识别的研究不但能够提供更好的视频表达模型,也能够促进其它视频相关任务的进展。在本文中,首先给出了视频中动作识别任务的定义,并区分了短时动作、动作、行为、事件等概念。其次,从传统方法和深度学习方法两方面介绍了动作识别任务的研究进展,其中传统方法又包括了基于全局表示与局部表示的识别方法。最后,介绍了具有代表性的动作识别数据集,并着重阐述了数据集的发展趋势。With the development of mobile communication technology,the online video data is exploding,and the demand for intelligent video analytics is increasing.With the application of deep learning technology,the field of video understanding and analysis has developed rapidly in recent years.As a core task in the field of video understanding,the research on action recognition not only provides better video representation,but also promotes the research of other video-related tasks.In this paper,the definition of action recognition task in videos is given and the four concepts of act,action,behavior,and event are distinguished.Secondly,the research progress of action recognition task is introduced from two aspects:the traditional methods and the deep learning based methods.The traditional methods include the recognition methods based on global representation and local representation.Finally,representative action recognition datasets and the development trend of these datasets are also described.

关 键 词：视频理解 动作识别 行为分析 深度学习 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]