Swin‐3DART: An Efficient and Robust Lightweight Transformer for Video Anomaly Detection with TG‐RGB +
Intissar ZIANI, Gueltoum Bendiab, Mourad Bouzenada, Meriem Guerar · IET Image Processing · 2026
ABSTRACT Video anomaly detection is vital for public safety but remains challenging due to complex motion patterns, limited robustness to motion‐related perturbations, and the heavy computation demands of modern transformers. To address these challenges, swin‐3DART is introduced as a unified framework that improves both efficiency and resilience. First, the proposed modality enhances RGB frames with temporal‐gradient motion cues, improving motion sensitivity. Second, it designs T‐GAP (temporal gradient adaptive perturbation), which generates worst‐case perturbations to expose vulnerabilities and strengthen the model through adversarial training. Third, an adversarial defence mechanism is embedded to ensure robustness, achieving consistently low attack success rates (4%–7%) across datasets. Finally, the framework incorporates the 3DART (3D adaptive receive transformer), which reduces memory footprint by ~12% and FLOPs by ~11.9%, making it suitable for deployment in real‐time surveillance or edge computing scenarios. Comprehensive evaluations show that swin‐3DART achieves state‐of‐the‐art AUCs of 95% on UBI fights, 86% on UCF‐crime, and 99% on RLVS. These results highlight swin‐3DART's potential as an efficient and robust solution for real‐time, safety‐critical video anomaly detection.