SCMU-Campus for Video Anomaly Detection

Moving cameras are widely employed across surveillance domains for their mobility and broad field of view, generating enormous amounts of untrimmed videos. Analyzing these videos to detect anomalies is time-consuming and labor-intensive, necessitating efficient and automated video anomaly detection (VAD) techniques. However, most of the existing VAD methods were designed for videos captured by static cameras and fail to generalize well to moving camera videos due to challenges like perspective changes, dynamic backgrounds, and small objects. While some methods have attempted to handle moving camera videos, they often struggle to detect anomalies that resemble normal patterns and are not data-efficient due to the extensive annotations required. In this paper, we propose Dynamics-Aware LEarning (DALE), a novel weakly supervised approach for moving camera video anomaly detection, which captures global-local spatiotemporal representations in moving camera videos through a dynamics-aware feature learning module and controls the representation structure of normality and abnormalities with a dual discrepancy loss. Extensive experiments on moving camera video datasets show that DALE consistently outperforms previous SOTA methods across multiple evaluation metrics, highlighting its effectiveness for moving camera video analysis.