Unsupervised Video Anomaly Detection with Swin Transformer and Temporal-Context Modeling

Video anomaly detection has broad applications in public safety, health monitoring, and emergency response. Existing transformer-based methods often struggle to capture multi-scale spatiotemporal patterns effectively. In this paper, we propose an unsupervised video anomaly detection framework that enhances feature representation by combining a 3D Swin Transformer with temporal shift modules and dynamic large kernel (DLK) convolutions. The 3D encoder-decoder structure models normal behavior from video sequences, and anomalies are identified through reconstruction errors. We validate our method on three public datasets: ShanghaiTech, Avenue, and Ped2. Our model achieves competitive accuracy, demonstrating improved capacity for capturing temporal dynamics and contextual details.