Weakly-supervised Video Anomaly Detection with Domain Adaptation

Video anomaly detection is a crucial yet challenging computer vision task aiming to automatically identify abnormal events in real-world videos. While it has broad applications in surveillance like detecting criminal activities, violent behaviors, and traffic accidents, it remains difficult due to the lack of fine-grained abnormal labeling data and the ambiguity of defining abnormal behaviors. This paper proposes a novel weakly-supervised method called Domain Adaptation Video Anomaly Detection (DA-VAD). DA-VAD treats video-level and clip-level features as source and target domains respectively, training a model in source domain to detect abnormalities in target domain. It uses domain adaptation to prevent overfitting and introduces a domain aligner to select representative target samples. These samples are used to build the domain adaptive aligning loss to address sample imbalance and perceive target cues. Experiments demonstrate that DA-VAD achieves good efficiency and effectiveness and outperform current methods on UCF-Crime and XD-Violence datasets.