基于变分自编码器的视频异常事件检测方法  

Video anomaly detection and localization via variational autoencoder

作　　者：苏鹏 王常顺[2] 卢萌萌[1] Su Peng;Wang Changshun;Lu Mengmeng(Shandong Labor Vocational and Technical College,Jinan 250022,China;School of Information Science and Electrical Engineering,Shandong Jiaotong University,Jinan 250357,China)

机构地区：[1]山东劳动职业技术学院,济南250022 [2]山东交通学院信息科学与电气工程学院,济南250357

出　　处：《电子测量与仪器学报》2020年第10期179-185,共7页Journal of Electronic Measurement and Instrumentation

基　　金：国家自然科学基金(61273144);山东劳动职业技术学院校级科技项目(2019KJ02)资助。

摘　　要：异常事件检测由于其在视频监控场景中的重要性而引起了广泛的关注。但是由于缺乏异常标注样本,使得这个问题较难解决。提出了一种新的部分监督学习方法,仅采用正常样本训练检测模型以进行视频异常事件检测和定位。假设所有正常样本的分布符合一个高斯分布,那么异常样本在这个高斯分布中将以较低的概率出现。该方法基于变分自编码器(VAE),通过端对端的深度学习技术,将正常样本的隐层表示约束成一个高斯分布。给定测试样本,通过变分自编码器获得其隐层表示,计算其隐层表示属于高斯分布的概率,并根据检测门限判断其是否异常。在两个公开的数据集(UCSD dataset和avenue dataset)上的实验结果表明,所提出的方法达到了92.3%的帧级AUC和82.1%的帧级AUC,以及571 fps的检测速度,在性能和效率上明显高于现有检测方法。Abnormal event detection has attracted wide attention due to its importance in video surveillance scenarios.The problem is particularly hard to crack because the lack of abnormal ground truth data.A novel partially supervised learning framework is presented for video anomaly detection and localization by training with the normal samples.The motivation is that the normal samples occur in high probability of a stochastic model,while the test samples occur in the low probability is regarded as anomaly.The method is based on variational auto-encoder(VAE),which can learn feature representations of the normal samples as a Gaussian distribution with deep learning technology.Experimental results of two popular benchmarks(UCSD dataset and avenue dataset)show that the proposed method achieved 92.3%and 82.1%frame-level AUC at a speed of 571 frames per second on average,which demonstrate the effectiveness and efficiency of the framework compared with other state of the arts approaches.

关 键 词：视频监控 异常检测 深度学习 变分自编码器 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]