Video anomaly detection has long been a challenging task in computer vision due to the complexity of real-world scenarios and the rarity of abnormal events. To address the limitations of traditional methods, this study proposes a novel dynamic perception network for robust anomaly detection. Specifically, a spatiotemporally decoupled architecture based on omni-dimensional dynamic convolution is designed to process spatial and temporal features separately, thus overcoming the limited dynamic modeling capacity of conventional convolution. A two-level memory propagation module based on a variant of the multi-head attention mechanism is designed to dynamically enhance inter-frame feature correlations while reducing computational overhead. Finally, a manifold clustering method is developed using a Riemannian curvature metric, overcoming the limitations of traditional distance measures and providing more discriminative representations of different types of abnormal events. Extensive experimental results on three public datasets, including Avenue, ShanghaiTech, and UCSD Ped2, demonstrate the effectiveness of the proposed method, which achieves state-of-the-art performance.