•  
  •  
 

العنوان

NA

أسماء الباحثين باللغة العربية

NA

Abstract

For public safety in crowded public places, less human detection is insufficient without interpretability and actionability of decision assistances. Although deep learning–based video surveillance systems have achieved accurate detection recently, the majority are restricted to passive perception and do not convert visual products into executive decisions for crowd status. This study presents an agentic AI-based decision support framework for interpretable real-time crowd status monitoring, to link vision-based perception and public safety decision. The proposed framework incorporates a YOLOv5-derived human detection module, an agentic state evaluation by considering persistability and a transparent rule-based decision making, which distinguishes scenes into crowd versus non-crowd. The human detections are temporally accumulated by counting and persistence measures to maintain stable and interpretable decision signals. The framework is tested on CCTV-style surveillance data and demonstrate consistent performance for detection, with a mean average precision of 95.3% (at IoU = 0.5) while achieving real-time processing speed of 25–30 fps on commodity hardware. The results interpretable decision logic explicitly into speech-and-video based monitoring of crowd status in safety-critical public facilities, which is practical, scalable, and explainable.

Article Language

English

Download

Share

COinS