Detecting rare events in videos is crucial for surveillance, industry, and healthcare, but remains a major challenge. Designed primarily for fixed-view surveillance environments, Existing unsupervised approaches often suffer from overfitting to narrowly defined notions of normal behavior, resulting in high false-positive rates and poor generalization performance. To address these limitations, this study proposes STARGA, a novel unsupervised rare-event detection framework that leverages a dual-stream generative architecture and domain-specific augmentation module to model normal spatiotemporal patterns. STARGA consists of a spatial generator-discriminator pair that learns structural regularities via edge-preserving representations and a temporal generator-discriminator pair that captures dynamic variations through background-subtracted motion patterns. A domain-informed augmentation module further expands the support of “normal” behavior by introducing subtle variations during training, which improves generalization to rare events. This study introduces a Feature-Aware Edge-Modified Gradient Difference Loss (FA-EM-GDL) to enforce high-frequency consistency and deep feature alignment between generated and real sequences. To stabilize training, this study employed an adaptive adversarial loss (combining relativistic hinge and energy-based objectives) with a decaying schedule, alongside spectral normalization, gradient penalties (R1, DRAGAN, and WGAN-LP), and contrastive feature regularization. Extensive evaluations across seven benchmarks, including UCSD Ped1/Ped2, CUHK Avenue, ShanghaiTech, UCF-Crime, Subway, and UBnormal, demonstrate that STARGA achieves state-of-the-art accuracy (frame-level AUC up to 99.9% on UCSD Ped2, 87.9% on ShanghaiTech, etc.) while running in real-time ( ≈ 40 FPS on 256 × 256 videos). It outperformed recent baselines (e.g., STAN, STemGAN, and CVAD-GAN) and remained robust across diverse scenes. This study further provides detailed ablation studies, error analyses, and comparisons with diffusion- and transformer-based approaches to substantiate the novelty and effectiveness of our approach. The source code for this research will be publicly available at https://github.com/aasimwadood/STARGA .
Wadood et al. (Thu,) studied this question.