计算机工程与应用 ›› 2021, Vol. 57 ›› Issue (10): 241-245.DOI: 10.3778/j.issn.1002-8331.2002-0189

• 工程与应用 • 上一篇    下一篇

三维卷积和视频帧采样算法下斗殴检测技术

黎晓昀,贾杰   

  1. 1.江西应用科技学院 人工智能学院,南昌 330100
    2.南昌航空大学 信息工程学院,南昌 330063
  • 出版日期:2021-05-15 发布日期:2021-05-10

Technology of Fighting Detection Based on Three-Dimensional Convolution and Video Frame Sampling Algorithm

LI Xiaoyun, JIA Jie   

  1. 1.College of Artificial Intelligence, Jiangxi University of Applied Sciences, Nanchang 330100, China
    2.College of Information, Nanchang Hangkong University, Nanchang 330063, China
  • Online:2021-05-15 Published:2021-05-10

摘要:

针对监控视频中斗殴行为检测的需求,提出了一种新的基于三维卷积神经网络和视频帧采样算法的斗殴行为检测方法。针对监控视频行为检测起始定位的难点,提出了一种利用基于人体姿态信息的关键区域检测算法定位斗殴行为起始帧的方法,形成了斗殴行为预识别空间。针对深度学习训练数据冗余和优化程度不够的问题,提出了基于时间采样的视频帧采样算法,并且搭建了一个三维卷积神经网络,使网络学习到整个行为动作的时空信息。实验结果证明了所提方法在两个公共数据集上取得了优越的性能。

关键词: 斗殴检测, 三维卷积, 预识别空间, 时空信息

Abstract:

Aiming at the demand of fighting action detection in surveillance video, a novel method of fighting action detection based on three-dimensional convolution neural network and video sampling algorithm is proposed. In view of the difficulty in the initial location of surveillance video action detection, a method based on the key region detection algorithm with human pose information is proposed to locate the initial frame of fighting action, thus forming the pre-recognition space of fighting action. As to the problem of redundancy and insufficient optimization degree of deep learning training data, a video sampling algorithm based on spatial temporal sampling is proposed. Also a three-dimensional convolution neural network is proposed, the network can learn the spatial temporal information of the whole action. Experimental results show that the proposed method achieves superior performance on two common datasets.

Key words: fighting detection, three-dimensional convolution, pre-recognition space, spatial temporal information