OBJECT ROTATION-BASED PSEUDO-ANOMALOUS DATA GENERATION FOR VIDEO ANOMALY DETECTION

Abstract

In this paper, we propose a novel method for generating anomalous data to improve the performance of video anomaly detection. The proposed method, called the Rotational Anomaly Objects Generator (RAG), synthesizes anomalous objects exhibiting loss-of-balance behaviors, such as falling, based on rotational motion theory. The generated anomalous samples are combined with normal samples to construct an augmented training dataset. A memory augmented autoencoder model called MNAD [1] is then trained on this augmented dataset for video anomaly detection. The resulting model, referred to as RAG-MNAD, is evaluated on three popular benchmark datasets and compared with baseline models as well as several recent video anomaly detection methods that utilize pseudo-anomalous data. Experimental results demonstrate that the proposed method significantly enhances detection performance. Notably, RAG-MNAD consistently achieves state-of-the-art results across multiple benchmarks.