基于深度自编码-高斯混合模型的视频异常检测方法

由于异常定义的模糊性和真实数据的复杂性，视频异常检测是智能视频监控中最具挑战性的问题之一。基于自动编码器(AE)的帧重建（当前或未来帧）是一种流行的视频异常检测方法。使用在正常数据上训练的模型，异常场景的重建误差通常比正常场景的重建误差大得多。但是，这类方法忽略了正常数据本身的内部结构，效率较低。基于此，提出了一种深度自动编码高斯混合模型(DAGMM)。首先利用深度自动编码器获得输入视频片段的生成低维表示和重构误差，并将其进一步输入高斯混合模型(GMM)。而估计网络则通过高斯混合模型预测能量概率，然后通过能量密度概率判断异常。DAGMM以端到端的方式同时联合优化深度自动编码器和GMM的参数，能够平衡自动编码重建、低维表示的密度估计和正则化，泛化能力强。在两个公共基准数据集上的实验结果表明，DAGMM达到了现有最高技术发展水平，在UCSD Ped2和ShanghaiTech两个数据集上分别取得了95.7%和72.9%的帧级AUC。

A video anomaly detection method based on deep autoencoding Gaussian mixture model

Due to the vagueness of anomaly definition and the complexity of real data, video anomaly detection is one of the most challenging problems in intelligent video surveillance. Frame reconstruction (current or future frame) based on autoencoder (AE) is a popular video anomaly detection method. Using a model trained on normal data, the reconstruction error of abnormal scenes is usually much larger than that of normal scenes. However, these methods ignore the internal structure of the normal data and are memory-consuming. Based on this, a deep auto-encoding Gaussian mixture model (DAGMM) was proposed. Firstly, the deep autoencoder was used to obtain the low-dimensional representation of the input video segment and the reconstruction error, and then further input into a Gaussian mixture model (GMM). The energy probability was predicted through the Gaussian mixture model, and then the anomaly was judged through the energy density probability. The proposed DAGMM can simultaneously optimizes the parameters of the deep autoencoder and GMM in an end-to-end manner, and balance auto-encoding reconstruction, density estimation and regularization of low-dimensional representation, and has strong generalization ability. Experimental results on two public benchmark datasets show that DAGMM has reached the highest level of technological development, achieving 95.7% and 72.9% frame-level AUC on the UCSD Ped2 and ShanghaiTech dataset, respectively.