结合重构和预测模型的无监督视频异常检测算法

Unsupervised video anomaly detection algorithm combining reconstruction and prediction model

针对目前视频异常检测领域使用单个重构模型无法完整重构图像、单个预测模型易受噪声扰动等问题,提出一种结合重构和预测模型的无监督视频异常检测算法。使用预测模型,输入当前多帧视频帧来精准预测下一帧图像,能够扩大正常和异常的区分度;使用重构模型提高网络的鲁棒性;使用结合残差网络和U-net网络的生成对抗网络(GAN)来处理异常,避免网络出现梯度爆炸和梯度消失等问题。实验证明:提出的算法能提高视频异常检测的准确性和鲁棒性,实现了监控视频中异常事件的自动监督。

Aiming at the problems that the current video anomaly detection field cannot reconstruct the image completely using a single reconstruction model, and is susceptible to noise disturbance using a single prediction model, an unsupervised video anomaly detection algorithm that combining reconstruction and prediction models is proposed.Using the prediction model to input the current multi-frame video frame to accurately predict the next frame image, which can expand the distinction between normal and abnormal;using the reconstruction model to improve the robustness of the network;using the generative adversarial network(GAN) which is the combination of residual network and U-net network to deal with anomalies to avoid problems such as gradient explosion and gradient disappearance.Experiments prove that the proposed algorithm can improve the accuracy and robustness of video anomaly detection, and realize the automatic supervision of abnormal events in the surveillance video.