第15卷第6期 智能系统学报 Vol.15 No.6 2020年11月 CAAI Transactions on Intelligent Systems Nov.2020 D0L:10.11992tis.202007003 深度自编码与自更新稀疏组合的异常事件检测算法 王倩倩，苗夺谦，张远健 (同济大学嵌入式系统与服务计算教育部重点实验室，上海201804) 摘要：基于深度学习的异常检测算法输入通常为视频帧或光流图像，检测精度和速度较低。针对上述问题， 提出了一种以运动前景块为中心的卷积自动编码器和自更新稀疏组合学习(convolutional auto-encoders and self- updating sparse combination learning,CASSC)算法。首先，采用自适应混合高斯模型(gaussian mixture model,, GMM)提取视频前景，并以滑动窗口的方式根据前景像素点占比过滤噪声：其次，构建3个卷积自动编码器提 取运动前景块的时空特征；最后，使用自更新稀疏组合学习对特征进行重构，依据重构误差进行异常判断。实 验结果表明，与现有算法相比，该方法不仅有效地提高了异常事件检测的准确性，且可以满足实时检测需求。 关键词：深度学习：稀疏组合；自动编码器：自更新：异常事件检测；卷积神经网络：无监督学习；稀疏学习 中图分类号：TP391文献标志码：A文章编号：1673-4785(2020)06-1197-07 中文引用格式：王倩倩，苗夺谦，张远健.深度自编码与自更新稀疏组合的异常事件检测算法.智能系统学报，2020,15(6)： 1197-1203. 英文引用格式：VANG Qiangian,MIAO Duogian,ZHANG Yuanjian..Abnormal event detection method based on deep auto- encoder and self-updating sparse combinationJ.CAAI transactions on intelligent systems,2020,15(6):1197-1203. Abnormal event detection method based on deep auto-encoder and self-updating sparse combination WANG Qianqian,MIAO Duoqian,ZHANG Yuanjian (Key Laboratory of Embedded System and Service Computing,Tongji University,Shanghai 201804,China) Abstract:In the construction of a deep learning model for abnormal event detection,frames or optical flow are con- sidered but the resulting accuracy and speed are not satisfactory.To address these problems,we present an algorithm based on convolutional auto-encoders and self-updating sparse combination learning,which is centered on the move- ment of foreground blocks.First,we use an adaptive Gaussian mixture model to extract the foreground.Using a sliding window,the foreground blocks that are moving,are filtered based on the number of foreground pixels.Three convolu- tional auto-encoders are then constructed to extract the temporal and spatial features of the moving foreground blocks. Lastly,self-updating sparse combination learning is applied to reconstruct the features and identify abnormal events based on the reconstruction error.The experimental results show that compared with existing algorithms,the proposed method improves the accuracy of abnormality detection and enables real-time detection. Keywords:deep learning;sparse combination;auto-encoder,self-updating,abnormal event detection;convolution neur- al network;unsupervised learning:sparse representation 异常事件检测是指通过图像处理、模式识别和 因光照、背景和视角等因素的影响以及缺少异常数 计算机视觉等技术，分析视频中的有效信息，判断 据，异常事件检测仍是一项具有挑战性的任务②。 异常事件检测通常包含特征提取和建立检测 异常事件。作为智能视频监控系统的重要应用之 模型。特征分为底层和深度学习特征。底层特征 一，异常事件检测受到了国内外学者的广泛关注。 主要有方向梯度直方图)、三维时空梯度、光流 收稿日期：2020-07-01 基金项目：国家自然科学基金项目(61976158,61673301). 直方图的等。近年来，部分学者提出基于深度学 通信作者：苗夺谦.E-mail:dqmiao@tongji..edu.cn. 习的检测算法6-。Zhou等设计了一个特征提DOI: 10.11992/tis.202007003 深度自编码与自更新稀疏组合的异常事件检测算法 王倩倩，苗夺谦，张远健 （同济大学 嵌入式系统与服务计算教育部重点实验室，上海 201804） 摘 要：基于深度学习的异常检测算法输入通常为视频帧或光流图像，检测精度和速度较低。针对上述问题， 提出了一种以运动前景块为中心的卷积自动编码器和自更新稀疏组合学习 (convolutional auto-encoders and self￾updating sparse combination learning, CASSC) 算法。首先，采用自适应混合高斯模型 (gaussian mixture model, GMM) 提取视频前景，并以滑动窗口的方式根据前景像素点占比过滤噪声；其次，构建 3 个卷积自动编码器提 取运动前景块的时空特征；最后，使用自更新稀疏组合学习对特征进行重构，依据重构误差进行异常判断。实 验结果表明，与现有算法相比，该方法不仅有效地提高了异常事件检测的准确性，且可以满足实时检测需求。 关键词：深度学习；稀疏组合；自动编码器；自更新；异常事件检测；卷积神经网络；无监督学习；稀疏学习 中图分类号：TP391 文献标志码：A 文章编号：1673−4785(2020)06−1197−07 中文引用格式：王倩倩, 苗夺谦, 张远健. 深度自编码与自更新稀疏组合的异常事件检测算法 [J]. 智能系统学报, 2020, 15(6): 1197–1203. 英文引用格式：WANG Qianqian, MIAO Duoqian, ZHANG Yuanjian. Abnormal event detection method based on deep auto￾encoder and self-updating sparse combination[J]. CAAI transactions on intelligent systems, 2020, 15(6): 1197–1203. Abnormal event detection method based on deep auto-encoder and self-updating sparse combination WANG Qianqian，MIAO Duoqian，ZHANG Yuanjian (Key Laboratory of Embedded System and Service Computing, Tongji University, Shanghai 201804, China) Abstract: In the construction of a deep learning model for abnormal event detection, frames or optical flow are con￾sidered but the resulting accuracy and speed are not satisfactory. To address these problems, we present an algorithm based on convolutional auto-encoders and self-updating sparse combination learning, which is centered on the move￾ment of foreground blocks. First, we use an adaptive Gaussian mixture model to extract the foreground. Using a sliding window, the foreground blocks that are moving, are filtered based on the number of foreground pixels. Three convolu￾tional auto-encoders are then constructed to extract the temporal and spatial features of the moving foreground blocks. Lastly, self-updating sparse combination learning is applied to reconstruct the features and identify abnormal events based on the reconstruction error. The experimental results show that compared with existing algorithms, the proposed method improves the accuracy of abnormality detection and enables real-time detection. Keywords: deep learning; sparse combination; auto-encoder; self-updating; abnormal event detection; convolution neur￾al network; unsupervised learning; sparse representation 异常事件检测是指通过图像处理、模式识别和 计算机视觉等技术，分析视频中的有效信息，判断 异常事件。作为智能视频监控系统的重要应用之 一，异常事件检测受到了国内外学者的广泛关注。 因光照、背景和视角等因素的影响以及缺少异常数 据，异常事件检测仍是一项具有挑战性的任务[1-2]。 异常事件检测通常包含特征提取和建立检测 模型。特征分为底层和深度学习特征。底层特征 主要有方向梯度直方图[3] 、三维时空梯度[4] 、光流 直方图[5] 等。近年来，部分学者提出基于深度学 习的检测算法[6-9]。Zhou 等 [8] 设计了一个特征提 收稿日期：2020−07−01. 基金项目：国家自然科学基金项目 (61976158，61673301). 通信作者：苗夺谦. E-mail：dqmiao@tongji.edu.cn. 第 15 卷第 6 期 智 能 系 统 学 报 Vol.15 No.6 2020 年 11 月 CAAI Transactions on Intelligent Systems Nov. 2020