Abstract
Modern network communication systems extensively utilize video data for various applications, creating a pressing need for efficient Video Anomaly Detection (VAD) mechanisms. The essence of VAD lies in the detection of discrepancies in the appearance-motion dynamics of normal and abnormal events. However, existing methods primarily focus on the isolated analysis of appearance or motion anomalies, thereby overlooking the crucial aspect of appearance-motion consistency semantics. To this end, we introduce a Memory-enhanced Appearance-Motion Consistency (MAMC) framework, which is focused on understanding complex appearance-motion consistency patterns in video data for the discernment of anomalies. The first stage of our model involves the design of an Appearance-Motion Fusion (AMF) module, engineered to generate a robust representation of scene dynamics that captures appearance-motion consistency. This consistency data is then processed through the memory module, augmenting the distinction between normal and anomalous events. Experimental results on three benchmark datasets validate the effectiveness of our approach, which achieves AUCs of 96.7%, 87.6%, and 71.5% on three benchmark datasets (UCSD Ped2, CUHK Avenue, and ShanghaiTech datasets). Additional experiments and visualization analysis confirm the effectiveness of the proposed MAMC framework in anomaly detection scenarios.
Published Version
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call