Abnormal behavior detection in surveillance videos is necessary for public monitoring and safety. In human-based surveillance systems, it requires continuous human attention and observation, which is a difficult task. The autonomous detection of such events is of essential significance. However, due to the scarcity of labeled data and the low occurrence probability of these events, abnormal event detection is a challenging vision problem. In this paper, we introduce a novel two-stage architecture for detecting anomalous behavior in videos. In the first stage, we propose a 3D Convolutional Autoencoder (3D-CAE) architecture to extract spatio-temporal features from normal event training videos. In 3D-CAE, the encoder and decoder architectures are based on 3D convolutions, which can learn both appearance and the motion features effectively in an unsupervised manner. In the second stage, we group the 3D spatio-temporal features into different normality clusters, and then remove the sparse clusters to represent a stronger pattern of normality. From these clusters, one-class SVM classifier is used to distinguish between normal and abnormal events based on the normality scores. Experimental results on four different benchmark datasets show significant performance improvement compared to state-of-the-art approaches while providing results in real-time.

监视视频中的异常行为检测对于公共监视和安全是必要的。在基于人的监视系统中，它需要不断的人为关注和观察，这是一项艰巨的任务。自主检测此类事件至关重要。但是，由于标记数据的缺乏和这些事件的发生概率低，异常事件检测是一个具有挑战性的视觉问题。在本文中，我们介绍了一种新颖的两阶段体系结构，用于检测视频中的异常行为。在第一阶段，我们提出了3D卷积自动编码器（3D-CAE）架构，以从正常事件训练视频中提取时空特征。在3D-CAE中，编码器和解码器架构基于3D卷积，可以以无人监督的方式有效地学习外观和动作特征。在第二阶段，我们将3D时空特征分组到不同的正态性群集中，然后删除稀疏群集以代表更强的正态性模式。从这些群集中，使用一类SVM分类器根据正常性得分来区分正常事件和异常事件。与最新方法相比，在四个不同基准数据集上的实验结果显示出显着的性能改进，同时提供了实时结果。一类SVM分类器用于根据正常性得分区分正常事件和异常事件。与最新方法相比，在四个不同基准数据集上的实验结果显示出显着的性能改进，同时提供了实时结果。一类SVM分类器用于根据正常性得分区分正常事件和异常事件。与最新方法相比，在四个不同基准数据集上的实验结果显示出显着的性能改进，同时提供了实时结果。