Anomaly graph: leveraging dynamic graph convolutional networks for enhanced video anomaly detection in surveillance and security applications

Abstract

Video abnormality behavior identification plays a pivotal role in improving the safety and security of surveillance systems by identifying unusual events within video streams. However, existing methods face challenges in capturing complex spatiotemporal anomalies effectively. To address this, we propose an anomaly graph approach that leverages dynamic graph representations and dynamic graph convolutional networks (GCN) for video anomaly detection. Anomaly graph constructs dynamic graphs where nodes represent objects or regions of interest within video frames while the edges encode spatial and temporal relationships. The GCN architecture extracts spatiotemporal embeddings from the dynamic graphs and allows the model to identify anomalies involving both spatial and temporal cues. Anomaly graph introduces uniqueness scores to quantify frame distinctiveness for precise anomaly detection while it employs adaptive reconstruction errors to pinpoint spatially localized anomalies. Real-time alerts are generated for detected anomalies to ensure timely responses to security incidents, and online fine-tuning (OLT) is incorporated as a dynamic learning mechanism to adapt to evolving anomaly patterns existing in dynamic environments. The extensive experimentation on benchmark datasets including UCSD Ped2 and CUHK Avenue is carried out for diverse performance measures, and the anomaly graph exhibits enhanced performance in the detection of anomalous events by achieving a detection accuracy of 98.72% and precision of 97.18%. Moreover, it is found to have achieved better values compared to other related methods for video anomaly detection. Overall, the anomaly graph introduces a robust video anomaly detection framework that excels in identifying complex spatiotemporal anomalies and empowers surveillance and security systems to detect anomalies more effectively.