摘要
视频异常检测已成为当前研究的热点问题,具有深刻的实际应用价值。针对视频异常检测中3D卷积计算复杂度高、难以训练以及使用重构方法进行检测时仅利用正常数据容易导致过拟合的问题,提出一种新型的深度可分离卷积异常驱动网络。首先,通过手工特征提取的方式抽取跳跃帧,并将其作为伪异常样本进行辅助训练;其次,设计深度可分离卷积网络,降低3D卷积的计算参数量;最后,通过最小化正常数据的重构误差和最大化异常数据的方式让网络学习以区分异常数据和正常数据。实验结果表明,该模型在各大公开数据集上均表现出具有竞争力的性能,其中在UCSDped1、UCSDped2、Avenue和UMN数据集上的准确率分别达91.3%、99.2%、87.4%和98.6%。此外,该模型对异常检测具有较强的灵敏度,且具有较强的泛化能力和鲁棒性。
Video anomaly detection has become a hot issue in current research with profound practical application value.Aiming at the problems of high computational complexity of 3D convolution in video anomaly detection,difficulty in training,and easy overfitting by utilizing only normal data when using reconstruction methods for detection,a novel deeply deparable convolutional anomaly-driven network is proposed.The network firstly extracts jump frames as pseudo-anomaly samples through manual feature extraction to assist training,secondly designs the deeply deparable convolutional network to reduce the number of computational parameters for 3D convolution,and finally allows the network to learn to differentiate between anomalous and normal data by minimizing the reconstruction error of normal data and maximizing the anomalous data.Experimental results show that the model exhibits competitive performance on all major public datasets,with accuracy rates of 91.3%,99.2%,87.4%and 98.6%on UCSDped1,UCSDped2,Avenue and UMN datasets,respectively.In addition,the model has strong sensitivity to anomaly detection,and has strong generalization ability and robustness.
作者
李新
宋刘广
孙钰奇
曾佳全
LI Xin;SONG Liuguang;SUN Yuqi;ZENG Jiaquan(College of Computer Science and Engineering,Guilin University of Technology;Guangxi Key Laboratory Fund of Embedded Technology and Intelligent System,Guilin 541000,China)
出处
《软件导刊》
2024年第10期187-192,共6页
Software Guide
基金
广西科技计划项目(2020GXNSFAA297255)
广西嵌入式技术与智能系统重点实验室开放基金项目(2020-2-6)。
关键词
深度可分离卷积
伪异常
帧重构
视频异常检测
deeply separable convolution
pseudo-anomaly
frame reconstruction
video anomalies detection
作者简介
李新(1973-),女,硕士,CCF会员,桂林理工大学计算机科学与工程学院副教授、硕士生导师,研究方向为计算机视觉;通讯作者:宋刘广(1997-),男,CCF会员,桂林理工大学计算机科学与工程学院硕士研究生,研究方向为视频异常检测;孙钰奇(1998-),男,CCF会员,桂林理工大学计算机科学与工程学院硕士研究生,研究方向为点云分割;曾佳全(1998-),男,桂林理工大学计算机科学与工程学院硕士研究生,研究方向为点云语义分割。
