西安电子科技大学学报 ›› 2022, Vol. 49 ›› Issue (2): 155-163.doi: 10.19665/j.issn1001-2400.2022.02.018

• 计算机科学与技术 & 网络空间安全 • 上一篇    下一篇

一种改进R(2+1)D网络的暴力行为检测方法

王勇(),靳伟昭(),冯伟(),全英汇()   

  1. 西安电子科技大学 电子工程学院,陕西 西安 710071
  • 收稿日期:2020-08-15 出版日期:2022-04-20 发布日期:2022-05-31
  • 通讯作者: 冯伟
  • 作者简介:王 勇(1976—),男,副教授,E-mail: ywangphd@xidian.edu.cn;|靳伟昭(1994—),男,西安电子科技大学硕士研究生,E-mail: jin_weizhao@163.com;|全英汇(1981—),男,教授,博士,E-mail: yhquan@mail.xidian.edu.cn
  • 基金资助:
    国家自然科学基金(61772397);国家自然科学基金(12005169);国家重点研发计划(2016YFE0200400);陕西省自然科学基础研究发展计划(2021JQ-074);陕西省社科界重大理论与现实问题研究项目2020(20ST-81);榆林市科技局项目(CXY-2020-094);多模态认知计算安徽省重点实验室开放基金(2021)

Improved violent behavior detection method for the R(2+1)D network

WANG Yong(),JIN Weizhao(),FENG Wei(),QUAN Yinghui()   

  1. School of Electronic Engineering,Xidian University,Xi’an 710071,China
  • Received:2020-08-15 Online:2022-04-20 Published:2022-05-31
  • Contact: Wei FENG

摘要:

公共安全中复杂的暴力行为检测具有重要的研究价值。传统的研究方法主要基于手工设计的特征,泛化能力较差,现有的深度学习网络模型泛化能力强但准确率较低。针对上述问题,提出了一个结合R (2+1)D改进网络和密集连接思想的暴力行为检测方法。由于原R(2+1)D残差模块支路中的步长为2的卷积操作忽略了特征图的3/4,所以将其优化为池化操作和步长为1的卷积操作。本实验的数据集共有1 500个视频样本,具体包括曲棍球比赛数据集和自制数据集。实验结果证明,改进后R(2+1)D网络相比原网络准确率分别提高了约2.30%和1.00%。另外,引入密集连接思想,将残差模块中的不同卷积层级间建立连接,使残差块中的卷积层输出特征图可重复使用,这在一定程度上减轻了训练过程中梯度消散的问题。通过在相同数据集上进行测试,发现改进后(2+1)D网络相比传统的方法,检测精度进一步提升了约1.47%和0.93%。因此,在公开的经典暴力行为检测数据集上的实验证明,相对于传统的3种网络学习方法,该算法能够更好地表示暴力行为信息,是一种更加简单有效的暴力行为检测方法。

关键词: 暴力行为检测, (2+1)D密集残差块, 残差网络, 深度学习

Abstract:

In public security,the complex violence behavior detection is of important research value.Traditional methods are mainly based on hand-crafted features,but they have a limited generalization ability.The existing deep learning network models have a better tolerance,but those kinds of methods face the challenge of obtaining a high accuracy.To solve the above problems,a novel violent behavior detection method is proposed in this paper by combining an improved R (2+1) D network and the dense connection idea.In the branch of the traditional R (2+1) D residual module,the three-quarters of the feature map is ignored due to the convolution operation with strides of 2.In this paper,the convolution operation is optimized for a pooling operation and a convolution operation with strides of 1,and the detection accuracy is increased by 2.3%.Besides,the dense connection idea is adopted into the residual module to establish the connection between different convolutional layers.The improvement could alleviate the problem of gradient dissipation during the training process,and the detection accuracy is further improved by 1.46%.Experimental results on one public dataset and one self-built dataset demonstrate the effectiveness of the proposed method for the complex violence behavior detection.

Key words: violence detection, (2+1) D dense residual block, residual network, deep learning

中图分类号: 

  • TP391.4