点击复制
双交叉注意力自编码器改进视频异常检测(
)
《南京师大学报（自然科学版）》[ISSN:1001-4616/CN:32-1239/N]
- 卷:
-
第46卷
- 期数:
-
2023年01期
- 页码:
-
110-119
- 栏目:
-
计算机科学与技术
- 出版日期:
-
2023-03-15
文章信息/Info
- Title:
-
Improved Video Anomaly Detection with Dual Criss-Cross Attention Auto Encoder
- 文章编号:
-
1001-4616(2023)01-0110-10
- 作者:
-
戚小莎1; 曾 静2; 吉根林2
-
(1.南京师范大学数学科学学院，江苏 南京 210023)
(2.南京师范大学计算机与电子信息/人工智能学院，江苏 南京 210023)
- Author(s):
-
Qi Xiaosha1; Zeng Jing2; Ji Genlin2
-
(1.School of Mathematical Sciences, Nanjing Normal University, Nanjing 210023, China)
(2.School of Computer and Electronic Information/Artificial Intelligence, Nanjing Normal University, Nanjing 210023, China)
-
- 关键词:
-
- Keywords:
-
- 分类号:
-
TP391
- DOI:
-
10.3969/j.issn.1001-4616.2023.01.015
- 文献标志码:
-
A
- 摘要:
-
针对视频中包含的异常事件数量稀少,信息密集的特征容易被遗漏等问题,本文提出一种双交叉注意力自编码器的视频异常事件检测方法. 首先预处理视频集,提取视频帧中表观和运动特征,然后设计双交叉注意力模块融入自编码器中,使特征图在自编码器中能够更好地关联全局特征. 其次将提取后的特征放入各自的自编码器中学习正常行为,使含有正常事件的视频帧能被模型重构,含有异常事件的视频帧则无法被重构. 最后通过检测模型得到各个视频帧的重构误差从而进行异常事件判定. 该方法可以以局部特征关联全局特征的方式有效提高视频异常事件检测的准确率,通过在多个公开数据集中进行实验验证,证明该方法优于其他同类方法.
- Abstract:
-
To solve the problems such as sparse quantity of abnormal events contained in video and information-intensive features are easy ommitted,this paper proposes a dual criss-cross attention auto encoder for video abnormal detection. Firstly,we preprocess the video to extract the apparent and motion features in the video,then design the dual criss-cross attention module and incorporate it into auto encoder,in this way,the features can better correlate the global features. Further,we put the extracted features into the respective auto encoders to learn normal behavior,in this way,video frames containing normal events can be reconstructed by the model and those containing abnormal events cannot be reconstructed. Finally,reconstruction errors of each video frame are obtained by the model to determine the abnormal events. This method can effectively improve the accuracy of abnormal event detection by correlating global features with local features,and it is proved to be better than other similar methods through experimental validation in several public datasets.
更新日期/Last Update:
2023-03-15