基于深度高斯过程回归的视频异常事件检测方法

Video anomaly detection and localization via deep Gaussian process regression

针对现有异常检测方法忽视异常事件发生概率小而造成虚警这个问题,基于高斯过程回归(GPR)的框架,将GPR核函数非参数化所具有的灵活性与深度神经网络的结构特性相结合,并将卷积神经网络封装在GPR的核函数中,以同时实现异常检测任务中特征提取和检测两个步骤。在测试阶段,相对于训练样本集的后验概率的对数似然较小的被判定为异常。方法在一个模拟数据集和一个完全真实的数据集上进行了实验验证,实验结果证明所提出的方法在两个数据集上分别达到了83.9%的帧级AUC和34.4%的帧级AUC,在性能上达到了现有技术发展水平。

Aiming at the problem of false alarms caused by the low probability of abnormal events in existing anomaly detection methods, a novel video anomaly detection approach is proposed based on the Gaussian process regression framework. By integrating the structural properties of deep learning with the flexibility of kernel methods, a new deep learning technology called deep Gaussian process regression that fully encapsulates CNN structure is introduced to extract features and detect anomaly in one model. The results on the popular Avenue dataset and on a recently introduced real-event video surveillance dataset show that the detection model based on deep Gaussian process regression has achieved 83.9% frame-level AUC and 34.4% frame-level AUC on the two dataset, respectively, and has reached the state of the art in performance.