融合梯度差信息的稀疏去噪自编码网络在异常行为检测中的应用  

Application of Sparse Denoising Auto Encoder Network with Gradient Difference Information for Abnormal Action Detection

作　　者：袁静[1,2] 章毓晋[1] 

机构地区：[1]清华大学电子工程系,北京100084 [2]防灾科技学院灾害信息工程系,三河101601

出　　处：《自动化学报》2017年第4期604-610,共7页Acta Automatica Sinica

基　　金：国家自然科学基金(61171118,61673234,U1636124)资助~~

摘　　要：本文是在稀疏去噪自编码网络的基础上,增加梯度差约束条件改进了自编码网络的解码效果,并成功地应用于全局异常行为检测的领域.基于稀疏自编码网络异常行为的检测过程是通过训练非异常行为的视频帧数据得到自编码网络模型,将待测视频帧输入模型,根据前向传播算法得到模型的输出,计算输出与输入之间的损失值,当该值高于某个阈值时,判定该视频帧中存在异常行为.通过在标准异常行为库开展的实验表明融合梯度差信息的稀疏去噪自编码网络算法较传统的稀疏去噪自编码网络算法在全局异常行为检测中更加有效.The paper proposes an improved sparse denoising auto encoder network by adding a gradient difference constraint, which has been successfully applied to frame-level abnormal action detection. Firstly, the abnormal action detection algorithm is trained using normal frames in video to get an auto encoder network model. Then after inputting a test frame into the network model, the forward propagation algorithm is used to get the output. Finally, if the loss between input data and output data is higher than a threshold, then it is concluded that some abnormal action shave occurred in the test frame. Experimental results on UMN datasets show that the improved network with gradient difference information is more effective than the traditional one in global abnormality detection.

关 键 词：自编码网络 稀疏编码 梯度差信息 异常行为 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]