Multimedia Systems Computer Networks and Communications
NMGM-VAD: normal memory-guided MIL for weakly-supervised multimodal video anomaly detection
Abstract
Video anomaly detection (VAD) is critical for intelligent surveillance, yet capturing diverse normal patterns in complex scenes remains a significant challenge. This paper proposes a unified framework via normal memory-guided MIL for weakly-supervised multimodal video anomaly detection (NMGM-VAD). Central to our framework is a Normal Memory Module that models and stores normal prototypes to provide normality-aware contextual representations. Rather than relying on memory augmentation alone, the proposed framework jointly exploits prototype-based feature enhancement and memory-deviation-guided snippet selection to improve anomaly discrimination under weak supervision. Furthermore, the discrepancy between input features and normal memory prototypes is combined with normalized feature magnitude to guide Top- k snippet selection, encouraging the model to prioritize informative instances during weakly supervised learning. Experimental results demonstrate that NMGM-VAD achieves strong performance, with AUC values of 99.65%, 98.14%, and 85.51% on UCSD Ped2, ShanghaiTech, and UCF-Crime, respectively. This work underscores the effectiveness of normal memory mechanisms in semantic modeling and anomaly detection for complex real-world scenarios. Our code and datasets are publicly available at: https://github.com/SVIL2024/NMGM-VAD .