A Hybrid U-Net and Vision Transformer approach for Video Anomaly detection
Anisha Gupta, Vidit Kumar · 2024
The growing dependence on surveillance videos highlights the need for automating anomalous events detection in videos, to improve real-time response and reduce manual workload. However, the diverse nature of abnormal activities across various public and private settings poses a challenge. This paper proposes integrating a U-Net-based architecture with vision transformers, leveraging their capabilities in capturing long-range dependencies, for frame-level anomaly detection in surveillance videos. The model learns inherent patterns by training on normal event videos and identifies deviations during testing for unseen video frames. Experiments on UCSD - Ped1 Dataset shows better results compared to others.