A novel reconstruction-based video anomaly detection with idempotent generative network

Video anomaly detection (VAD) is vital in intelligent security for public safety. Reconstruction-based VAD has received increasing research attention, but faces challenges such as missing anomalies for the reconstruction error as a criterion, and information loss when suppressing anomalous data, existing methods also struggle to detect unseen anomalies. We propose a novel reconstruction-based video anomaly detection with idempotent generative network (RVADIGN), which is composed of the novel reconstruction module namely PSVAE and an idempotent loss term (IGN). Specifically, video frames are reconstructed within PSVAE. During this process, skip connections are established between the encoder and decoder to enhance contextual understanding. Finite Scalar Quantization (FSQ) layer is designed to discretize the encoder’s output, preserving key discriminative features. Meanwhile, the Pyramid Deformation Module (PDM), as an integral part of PSVAE, computes offset maps of original video frames for anomaly detection supplementation. Alongside PSVAE, idempotence is introduced as a regularity term, which projects the anomaly information back to the estimated manifolds of the target distribution, improves the adaptability and stability of the reconstruction method in different videos. Extensive experimental results demonstrate that our method outperforms other state-of-the-art VAD methods, achieving 99.03%, 92.40%, and 77.20% AUC on UCSD Ped2, CUHK Avenue, and ShanghaiTech, respectively.