伪异常引导的卷积自编码网络视频异常检测

Pseudo-anomaly-guided convolutional autoencoder video anomaly detection

视频异常检测由于可以高效、低成本地维护公共安全,在国家安防、医疗监护中发挥着重要作用.基于重构的深度自编码网络异常检测方法因其强大的表示能力而得到了广泛的研究.然而,自编码网络通常也可以成功地重建异常行为,从而导致异常行为的漏检.针对这一问题,提出了一种伪异常引导的卷积自编码网络视频异常检测方法,模型使用3D卷积提取视频时空特征.首先,通过正常数据模拟异常数据分布生成伪异常,提出了两种生成伪异常的方法:基于跳帧的方法和基于补丁的方法;然后,使用正常数据和生成的伪异常数据训练模型,训练时较好地重建正常数据同时较差地重建伪异常数据,由此模型被鼓励为限制异常数据的重建;最后,在UCSD-Ped2、Avenue和ShanghaiTech三个公共视频异常检测数据集上与其他基于重建的模型进行比较,其检测精度获得了有效提升.

Video anomaly detection plays an important role in national security and medical monitoring because it can maintain public safety efficiently and at low cost.Reconstruction-based deep autoencoder anomaly detection methods have been extensively studied for their powerful representational capabilities.However,autoencoder can often also successfully reconstruct abnormal behavior,leading to missed detection of abnormal behavior.To solve this problem,a pseudo-anomaly guided convolutional autoencoder video anomaly detection method is proposed.The model uses 3D convolution to extract video spatio-temporal features.Firstly,the normal data is used to simulate the abnormal data distribution to generate pseudo-abnormal data.Two methods of generating pseudo-anomalies are proposed:skip frame and patch based.Then,the normal data and the generated pseudo-abnormal data were used to train the model.During the training,the normal data was reconstructed well while the pseudo-abnormal data was reconstructed poorly.Therefore,the model was encouraged to limit the reconstruction of abnormal data.Finally,compared with other reconstruction-based models on UCSD-Ped2,Avenue and ShanghaiTech public video anomaly detection data sets,the detection accuracy has been effectively improved.