Cross-Scale Spatiotemporal Memory-Augmented Network for Unsupervised Video Anomaly Detection

Video anomaly detection (VAD), a critical task in intelligent surveillance systems, faces two key challenges: Dynamic behavioral characterization under complex scenarios and robust spatiotemporal context modeling. Existing methods face limitations, such as inadequate cross-scale feature fusion, weak channel-wise dependency modeling, and sensitivity to background noise. To address these issues, we propose a novel multi-scale spatiotemporal feature augmentation framework. Our approach introduces three core innovations: Hierarchical feature pyramid architecture for multi-granularity representation learning, capturing both local motion patterns and global scene semantics; A channel-adaptive attention mechanism that dynamically models long-range spatiotemporal dependencies; A spatiotemporal Gaussian difference module to enhance anomaly response through frequency-domain feature reconstruction, effectively suppressing noise interference. Extensive experiments on UCSD Ped1/2, CUHK Avenue, and ShanghaiTech benchmarks demonstrate that our method achieves state-of-the-art performance, outperforming existing approaches in both accuracy and robustness.