结合双流I3D和注意力机制的视频异常事件检测  

Video Anomaly Event Detection Combining Dual-Stream I3D and Attention Mechanism

作　　者：程相贵 刘钊[1] 郭放 CHENG Xianggui;LIU Zhao;GUO Fang(School of Information and Cyber Security,People's Public Security University of China,Beijing 100038,China)

机构地区：[1]中国人民公安大学信息与网络安全学院,北京100038

出　　处：《信息与电脑》2022年第24期65-68,共4页Information & Computer

基　　金：国家重点研发计划项目(项目编号:2020YFC1522600)。

摘　　要：为了减少视频异常事件检测过程中冗余帧对检测效果的影响,更好地利用视频中关键帧包含的有用信息,提出了一种结合双流膨胀卷积神经网络(Two-stream Inflated 3D ConvNets,I3D)模型和压缩-激励注意力机制多示例异常检测算法。首先,利用双流膨胀卷积神经网络提取视频时空特征;其次,通过双向长短期记忆(Bidirectional Long Short Term Memory,Bidirectional LSTM)神经网络获取视频特征长时序信息;再次,借助压缩-激励注意力机制分配特征权重;最后,通过多示例排序损失函数得到异常排序模型,并在排序损失函数中加入稀疏损失和平滑损失,更好地预测视频异常分数。实验表明,在公开数据集UCF-Crime上检测准确率达到了82.84%,高于基线模型7.43%。In order to reduce the impact of redundant frames on the detection effect in the process of video anomaly event detection and to make better use of the useful information in the video, a multi example anomaly detection algorithm combining a dual-stream Two-stream Inflated 3D ConvNets(I3D) model and a compression-incentive attention mechanism is proposed. Firstly, the video spatio-temporal features are extracted using a dual-stream expanded convolutional neural network. Secondly, the video sequence long time series information is obtained by Bidirectional Long Short Term Memory(Bidirectional LSTM). Thirdly, assigning feature weights with the help of a compressionincentive attention mechanism. Finally, the anomaly ranking model is obtained by a multiple example ranking loss function, and sparse loss and smoothing loss are added to the ranking loss function to better predict video anomaly scores. The experiments showed that the detection accuracy reached 82.84% on the publicly available dataset UCF-Crime, which is 7.43% higher than the baseline model.

关 键 词：多示例学习 注意力机制 双向长短期记忆(Bidirectional LSTM)神经网络 视频异常检测 

分 类 号：TP301.6[自动化与计算机技术—计算机系统结构]