An integration of Pseudo Anomalies and Memory Augmented Autoencoder for Video Anomaly Detection
Anh Le, Quang Uy Nguyen, Trần Nguyên Ngọc, Hai‐Hong Phan, Thi Huong Chu · 2022
Video anomaly detection (VAD) has received a lot of attention from the research community in recent years. The purpose of VAD is to identify the anomalous appearance and behavior of objects in videos. Due to the difficulty in collecting anomalous data, video anomaly detection is often based on a one-class classification (OCC) problem. Among the methods, deep autoencoders have been shown to be effective for anomaly detection in video. Specifically, autoencoders are only trained on normal images during the training time, then it is expected to reconstruct or predicted well for normal frames but poorly for anomalous ones at the test time. Contrary to the expectation, abnormal frames may have a chance of being well constructed by trained autoencoders because of the powerful representation capacity of the deep neural network as well as the diversity of normal patterns. To address the issue, we propose a strategy called and shortened as PA-MAE (Pseudo Anomalies-Memory-augmented Autoencoder) which feeds pseudo anomalies into a memory-augmented autoencoder network during training time. Therefore, the proposed model is able to take advantage of both memory-based autoencoder networks and pseudo-anomaly synthesizers which can store the prototypical features of the normal frames and produce high reconstruction errors on dummy anomaly examples. Experimental results on several benchmark video datasets (i.e. Ped2, Avenue, and ShanghaiTech) demonstrate that our method outperforms some state-of-the-art memory-augmented methods as well as several recent models using pseudo-anomaly synthesizers.