Memory-Guided Coordinate Encoding Network for Anomaly Detection
Xingang Wang, Hong Zhang, Rui Cao, Jinyan Zhou, Xingchao Lu · 2023
Video anomaly detection remains challenging due to the complexity of visual scenes. Video surveillance is often fixed lens, and existing approaches, either using autoencoder architectures or generative adversarial network models, lack encoding of dynamic and static information in feature extraction to emphasize information-rich features. Based on this, we improve the encoder architecture and propose a memory-guided coordinate encoding network based on extensive experiments to introduce a coordinate attention module to improve the U-Net network and enhance dynamic entity representation. Considering the diversity of abnormal events, we use the memory module to record the prototype patterns of normal features and propose feature discretization loss and feature aggregation loss to make a compact representation between features and separation between feature terms to improve the accessibility and prediction accuracy of the memory module. Our experimental results on three open standard datasets show that our model outperforms the state-of-the-art methods.