Towards Secure Video Surveillance: A Few-Shot Spatiotemporal Perception Transformer for Unseen Behavioral Anomalies
IEEE AVSS 2025 (CORE A) — International Conference on Advanced Visual and Signal-Based Systems, 2025
FewShot-SPT is a transformer-based framework for few-shot anomaly detection in surveillance systems. It introduces Event-Guided Keyframe Extraction (EGKE), Adaptive Modality Gating (AMG) for multimodal fusion, and adaptive prototypical learning for improved generalization to unseen anomalies. The model achieves 91.6% AUC on UCF-Crime and 84.2% accuracy on XD-Violence, outperforming state-of-the-art methods.
This work was carried out in collaboration with The Ohio State University and Ming Chi University of Technology.
Deployed on an NVIDIA Jetson edge device, the system enables real-time park surveillance under edge constraints, detecting incidents such as falls, intrusions, and weapon-related activities in real-world environments.