VUEBLOX: a semi-supervised hybrid deep learning framework for occlusion-aware and interpretable crowd anomaly detection

Abstract

Crowd tracking using computer vision technologies enhances public safety, but detecting crowd anomalies remains challenging due to issues like occlusion and interpretability models’ low accuracy [1]. Also, those are irregularly happening constitute a small percentage, so the datasets have lots of missing labeled outliers. This study address these gaps by proposing a semi-supervised hybrid architecture with dedicated modules for occlusion and interpretability.