STC-VAD: spatio-temporal collaborative weakly-supervised video anomaly detection

Abstract

Weakly supervised video anomaly detection is of great importance to security surveillance. The task detects anomaly with solely video-level labels. Despite significant progress made by existing methods, they have the following limitations: On one hand, they pay little attention to the specific spatial regions where anomalies occur and spatial contextual information. On the other hand, they struggle to capture anomaly features across different temporal scales and overlook the role of anomalous feature enhancement. In response to these problems, we propose a novel spatio-temporal collaborative video anomaly detection model (STC-VAD). The model employs a dual-branch architecture to separately model spatial and temporal information. The spatial compression adapter (SCA) is designed to solve the corresponding spatial problem, while the multi-scale temporal modeling (MTM) and anomaly feature amplification (AFA) modules are developed to address the corresponding temporal problem. STC-VAD demonstrates promising performance on UCF-Crime and XD-Violence datasets, which achieves 87.49% AUC and 85.05% AP and outperforms the current state-of-the-art methods.