Design of an Improved Model for Anomaly Detection Using Long-Term Transformer Networks and Graph-Augmented Recurrent Neural Networks

The detection of anomalies in multiple camera feeds is crucial in the latest surveillance system. However, the methods proposed till date failed to effectively capture the long dependencies and spatial relations over the feeds that generally cause suboptimal performance with high false positives. We propose a novel deep learning architecture designed for incorporation of temporal contexts from multiple camera feeds. Here we introduce the Long Term Transformer Networks, which seize on the both short and long temporal dependencies efficiently through sparse attention mechanisms without increasing the memory usage over $30-40 \%$. Thus, their precision improves by 510%. We therefore develop Graph-Augmented Recurrent Neural Networks, using graph neural networks fused with recurrent layers to capture the temporal dependencies and spatial layoutsa $\mathbf{1 0 - 1 5 \%}$ increase in recall, especially in the overlapping fields of view for different scenarios. To deepen the time modeling, we propose Dual-Path Temporal Attention Networks (DPTAN) that make use of parallel short-term and long-term paths towards capturing dynamic behaviors. DPTAN improved F1 scores of $\mathbf{7 - 1 2 \%}$ and reduced false positives by $\mathbf{2 0 \%}$. Last but not least, Hierarchical Temporal Graph Networks (HTGN) is proposed for capturing multiple-scale temporal dependencies that improve the accuracy of anomaly detection by $\mathbf{1 5 - 2 0 \%}$ more than state-of-theart methods on a wide range of real-world scenarios. Extensive experiments on large-scale surveillance datasets have shown that the architectures developed can considerably enhance the performance and generalization of anomaly detection in previously unseen anomalies. Indeed, by having substantial improvements in terms of precision, recall, and F1-score, these architectures are rich platforms for enhanced anomaly detection performances. By introducing new benchmarks for multiple camera anomaly detection, it further enables scalable and robust solutions to realworld surveillance applications.