VigilFormer: Deformable Attention for Video Anomaly Detection with Causal Risk Inference

A new video anomaly detection framework called VigilFormer achieves high accuracy while maintaining real-time speed, according to a preprint posted on arXiv. The system combines deformable spatio-temporal attention with causal temporal modeling to process untrimmed surveillance footage at 41.5 frames per second on a single GPU [1][2]. The framework, detailed in a paper submitted on 31 May 2026, addresses a persistent tension in surveillance video analysis: methods that improve detection accuracy often sacrifice processing speed, and vice versa [1][2]. VigilFormer's architecture includes a Deformable Spatio-Temporal Encoder, or DSTE, which attends to a sparse set of informative locations across frames rather than processing every pixel. This approach avoids the quadratic computational cost of dense attention while still capturing irregular motion patterns [1][2]. A second component, the Causal Anomaly Classifier, applies dilated causal convolutions over snippet-level features and uses a contrastive multiple-instance learning objective to separate anomalous and normal representations without requiring frame-level labels [1][2]. The paper also introduces an Adaptive Confidence Scheduler that dynamically skips low-information frames during inference, reducing redundant computation in static scenes [1][2]. On three benchmark datasets, VigilFormer recorded AUC scores of 87.83% on UCF-Crime, 97.21% on ShanghaiTech, and 89.74% on CUHK Avenue [1][2]. The authors report that these results outperform recent weakly-supervised methods in both accuracy and speed [1][2]. The preprint appears on arXiv, an open-access repository that hosts scientific papers across mathematics, physics, computer science, and other fields. As of November 2024, the repository receives about 24,000 submissions per month and has surpassed two million articles [6]. Papers on arXiv are moderated but not peer-reviewed before posting [6].