Video Anomaly Detection via Multi-module Collaborative Adversarial Learning

Video anomaly detection is a critical task in intelligent surveillance and public safety, aiming to automatically identify rare or suspicious events. The main challenge lies in effectively modeling normal activities and distinguishing subtle anomalies with great precision. To address this, we propose a multi-module collaborative adversarial learning framework that integrates a generative adversarial mechanism, a memory module, the Self-Supervised Predictive Convolutional Attention Block (SSPCAB), and the Pyramid Deformation Module (PDM). This collaborative architecture enables the model to capture both spatial relationships and temporal changes in video sequences. Specifically, the framework leverages frame prediction to construct a spatiotemporal model of normal behaviors, while the memory module enhances long-term temporal modeling. SSPCAB explicitly learns global structural patterns through masked convolution and channel attention, and PDM improves motion perception for better anomaly discrimination. Combined with adversarial training, the framework achieves strong generalization and discriminative capability. Our approach delivers commendable performance, attaining AUC scores of 92.2% for the CUHK Avenue dataset and 99.1% when applied to the UCSD Ped2 dataset.