摘要
动作检测已经成为计算机视觉领域的一个重要研究方向,特别是对特定事件的检测,如打架斗殴等暴力事件。暴力行为检测任务在视频监控等场景中是十分有用的,例如监狱、学校等公共场所。其实用性使得越来越多的研究者研发相应的检测算法。传统方法提取的时空特性同时考虑运动和外观信息,进而达到不错的准确率,但是他们的计算量却令人望而却步。近年来,随着深度学习广泛应用于计算机视觉领域,暴力行为检测任务也得到了很好地解决。我们根据two-stage目标检测架构的启发,提出一个用于暴力行为检测的方案,通过结合目标检测和行为识别的算法,更加准确地关注到运动执行者本身,避免无关背景和其他信息对检测结果的影响,进而实现了高效的暴力行为检测。此外,我们的方案还能有效地扩展到时空行为检测任务上,进而实现对更多类别的行为进行检测。
Action detection has become an important research direction in the field of computer vision,especially the detection of specific events,such as fights and other violent events.Violent action detection task is very useful in video surveillance scenarios,such as prisons,schools and other public places.Its practicality makes more and more researchers develop corresponding detection algorithms.The temporal and spatial features extracted by traditional methods consider motion and appearance information at the same time to achieve impressive accuracy,but they are computationally expensive.In recent years,as deep learning is widely used in the field of computer vision,the task of violent action detection has also been well solved.Based on the inspiration of the two-stage object detection architecture,we propose a scheme for violent action detection.By combining object detection and action recognition methods,we can more accurately focus on the action performer itself,and avoid the influence of irrelevant background and other information on the detection results,and then realizing efficient violent action detection.In addition,our scheme can be easier extended to spatiotemporal action detection task,and achieve detection of more types of action.
作者
王怡明
WANG Yiming(College of Computer Science,Sichuan University,Chengdu 610065)
出处
《现代计算机》
2021年第12期83-86,91,共5页
Modern Computer
基金
国家重点研发计划:高质高效的审判支撑关键技术及装备研究(No.2018YFC0830300)。
关键词
暴力检测
时空行为检测
行为识别
深度学习
Violence Detection
Spatiotemporal Action Detection
Action Recognition
Deep Learning
