Video anomaly detection based on both normal and anomaly reconstruction with Siamese Network
Xi Luo, Shifeng Li, Cheng Yan, Xiaoru Liu · 2025
Traditional autoencoders (AEs) often use only normal data for training in video anomaly detection (VAD) tasks, which makes them unable to better distinguish normal samples from abnormal samples and can also have better results in the reconstruction of abnormal data. To address this challenge, we present a novel unsupervised learning framework that integrates the ideas of Siamese networks and pseudo anomalies to effectively learn feature differences in data distribution. By leveraging noise injection and explanatory note, the method enables robust anomaly detection without labeled data. The model learns a discriminative feature space where normal and abnormal samples are well-separated, and Kullback-Leibler Divergence (KLD) is used to quantify the distribution differences between them, enhancing anomaly discrimination. Experimental results on the UCSD Ped2 and CUHK Avenue datasets demonstrate the effectiveness of the proposed method, achieving superior performance in anomaly detection tasks.