ShielDir: AI-Powered Real-Time Threat Detection System to Reduce Crime Response Time
Abstract
The rapid growth of urban populations and increasing social inequalities have contributed to a rise in violent crimes in public spaces. Traditional surveillance systems, relying mainly on passive CCTV cameras, often fail to support timely interventions. Although these systems record incidents, their inability to interpret real-time behaviors-such as weapon use and violent acts-limits their effectiveness. As a result, critical crimes like armed robbery, assault, arson, vandalism, and domestic violence frequently go unnoticed, especially in areas without active human monitoring. To address this challenge, we propose ShielDir, a threat detection system powered by artificial intelligence (AI) that performs real-time analysis of human behaviors and weapon presence using deep learning models, identifying threats across 14 categories of criminal activity. The system provides instant alerts to authorities, reducing response times and enhancing public safety in live or recorded video streams. ShielDir integrates YOLOv11 for weapon detection and OPear, a VideoMAE-based model for behavior analysis, within a containerized microservice architecture supported by Kafka to enable seamless, real-time data streaming and processing.