A cascade reconstruction model with generalization ability evaluation for anomaly detection in videos. (February 2022)

• Record Type: Journal Article
• Authors: Zhong, Yuanhong; Chen, Xia; Jiang, Jinyang; Ren, Fan
• Is Part Of: Pattern recognition. Volume 122(2022)
• Publication Date: 2022-02
• Subjects: Anomaly detection -- pixel reconstruction -- optical flow prediction -- generalization ability evaluation; Pattern perception -- Periodicals; Perception des structures -- Périodiques; Patroonherkenning; 006.4
• DOI: 10.1016/j.patcog.2021.108336
• Languages: English
• ISSNs: 0031-3203

Abstract

Highlights: We propose a cascaded model composed of a frame reconstruction network and an optical flow prediction network. By predicting optical flow based on reconstruction frame, the model increases the prediction error of optical flow containing abnormal events. We propose to choose a model for anomaly detection according to the reconstruction error of pseudo abnormal sample, which restricts the ability of test model to represent abnormal frames and achieves the tradeoff between the generalization of test model for normal and anomaly. We propose a two-step calculation function of abnormality score, which takes into account the spatial locality and temporal continuity of abnormal events. Experiments on three datasets demonstrate the competitive performance of our model compared with state-of-the-art methods. Abstract: Anomaly detection plays an important role in surveillance video since it maintains public safety efficiently with low cost. In current works, anomaly detection methods based on reconstruction with deep learning has been extensively studied for the powerful representation capacity. These methods use convolutional neural networks to learn model for describing normality at training and detect anomalies according to reconstruction error at testing. However, excessive representation capacity of neural networks will also bring disadvantages to anomaly detection when it is powerful enough to reconstruct abnormal information. For this reason, we proposed two solutions; firstly, a cascade model which conducts pixel reconstruction followed by optical flow prediction is designed. The conversion from frame to optical flow learns the correlation between object appearance and motion, while pixel reconstruction enlarges the optical flow prediction error to conduct effective anomaly detection. Secondly, the generalization ability evaluation based on pseudo-anomaly is proposed, which is used to evaluate the ability of model to represent anomaly, thus selecting an optimal model for anomaly detection. The selected model achieves AUC 88.9% on Avenue, 82.6% on Ped1, 97.7% on Ped2, and 70.7% on ShanghaiTech datasets. Extensive ablation experiments have verified the effectiveness of our method. Code will be released at https://github.com/Xia-Chen/Cascade_Reconstruction.