Five shot and Ten Shot Analysis for Anomaly Detection of Video Scenes

Abstract

Anomaly detection in video surveillance is essential for maintaining public safety and identifying irregular events in real-world settings. This paper introduces a novel few-shot learning framework for video anomaly detection, titled Five-Shot and Ten-Shot Analysis for Anomaly Detection from Video Scenes. The proposed method leverages a meta-learning approach to achieve effective and adaptive performance under limited training data conditions. Specifically, this paper evaluated detection performance across 11 distinct anomaly categories—including water incidents, traffic accidents, shootings, fires, assaults, vandalism, explosions, fights, object falls, robberies, and human falls—captured across 14 diverse real-world scenarios such as malls, streets, offices, highways, and public parks. Our approach utilizes a five-shot one-query scheme during training and a ten-shot one-query scheme during testing, enabling strong generalization from minimal examples. The framework incorporates a hybrid backbone that combines spatial-temporal feature extraction using the Swin Transformer with Model-Agnostic Meta-Learning (MAML) for rapid task adaptation. Experimental results demonstrate that our method achieves robust anomaly detection performance with limited data and delivers competitive Area Under the Curve (AUC) scores.