Abnormal Behavior Detection in Surveillance Systems Using a Hybrid EfficientNet-Transformer Model

Abstract

Anomaly detection in video surveillance is vital for public safety, but challenges arise from the unpredictability of abnormal behaviors and large-scale systems. We propose a hybrid architecture combining EfficientNetV2S for efficient feature extraction with a transformer encoder to capture long-range dependencies through self-attention. This model robustly detects abnormal events by modeling local and global patterns in video frames. Evaluated on UCSD Ped1, UCSD Ped2, Avenue and SHTech datasets, our approach achieved accuracies of 99.51%, 99.80%, 94.82% and 78.125%, outperforming existing methods and proving their suitability for real-time smart surveillance applications.