基于一类神经网络的视频异常事件检测方法

One-class neural network for video anomaly detection and localization

视频异常事件检测一直是一个具有挑战性的问题,现有的方法往往把视频特征提取和异常检测模型建立两个步骤独立设计,导致方法无法达到最优。针对该问题,设计了一种一类神经网络方法用于视频异常检测。该方法结合了自编码器的逐层数据表示形式能力以及一类分类能力,隐藏层的特征是针对异常检测的特定任务而构建的,从而获得了一个超平面以将所有正常样本与异常样本分开。实验结果表明,提出的方法在PED子集和PED2子集上分别达到了94.9%的帧级AUC和94.5%的帧级AUC,在Subway数据集上实现了80个正确事件检测,证实了该方法在工业和城市环境中的广泛适用性。

Due to the vague definition of abnormal events and the scarcity of its own samples, the detection of video abnormal events has always been a challenging problem. Existing methods often separate the two steps of video feature extraction and anomaly detection model establishment, it leads to the method that cannot reach the optimum. This paper follows the idea of distance-based anomaly detection, and proposes a one-class neural network method for video anomaly detection. This method combines the layer-by-layer data representation ability of the autoencoder and the one-class classification ability. The features of the hidden layer are constructed for the specific task of anomaly detection, thereby obtaining a hyperplane to separate all normal samples from abnormal samples. The experimental results on two benchmark data sets show that the proposed method achieves 94.9% frame-level AUC and 94.5% frame-level AUC on the PED subset and PED2 subset, respectively, and achieves 80 correct event detections on the Subway dataset, confirming the wide applicability of the method in industrial and urban environments.