基于深度学习方法的在线动作检测技术综述  

A Review of Online Action Detection Techniques Based on Deep Learning Methods

作　　者：张婉 张睿萱 谢昭[1] 刘家仁 金宇奇 沈玉龙 

机构地区：[1]合肥工业大学计算机与信息学院,安徽 宣城

出　　处：《计算机科学与应用》2023年第3期626-634,共9页Computer Science and Application

摘　　要：动作检测技术,是在算法观测整个视频后自动识别出其中出现的动作类别和始末时间,在机器人、智能家居、城市安防等领域均有应用。然而实际生活中,很多场景需要在某些事件刚发生时给予反馈,这需要检测算法以一种在线形式接收视频信息,传统的动作检测算法因为观测信息不完全,效果很差。本文基于当前在线动作检测算法的研究现状,概述了目前用于在线检测的主流方法,总结了目前研究将遇到的挑战。Action detection technology, in which an algorithm observes the entire video and then automatically identifies the type of action that occurs in it and the start and end times, is used in robotics, smart homes, urban security and other areas. However, in real life, many scenarios require feedback when certain events first occur, which requires detection algorithms to receive video information in an online format. Traditional action detection algorithms are ineffective because of incomplete observation information. Based on the current state of research in online action detection algorithms, this paper provides an overview of the mainstream methods currently used for online detection and summarises the challenges that current research will encounter.

关 键 词：在线动作检测 机器视觉 深度学习 

分 类 号：TP3[自动化与计算机技术—计算机科学与技术]