Enhanced Real-Time Multi-Person Smoking Event Detection for Public Health Surveillance Using YOLOv8 Deep Learning Architecture

Abstract

Smoking remains a pervasive global health issue, claiming millions of lives annually and posing significant challenges to public health and safety. This study presents an innovative real-time multi-person smoking event detection system leveraging the advanced YOLOv8 architecture to address global health concerns associated with smoking. Smoking is a major contributor to preventable diseases and fatalities worldwide, with significant impacts on individual health and public safety. The proposed system aims to enhance surveillance capabilities by accurately detecting and classifying smoking behaviors in diverse, real-world scenarios. A dataset comprising 20,540 labeled images was utilized for training, validation, and testing, ensuring robust model performance. Experimental results indicate an overall mean average precision (mAP) of 84.35%, with smoking detection achieving 93.1% precision and non-smoking detection achieving 95.4% precision. Statistical analysis confirmed significant differences in detection accuracy across datasets. The findings demonstrate the system's potential for practical applications in public health initiatives, enforcement of no-smoking policies, and real-time monitoring of smoke-free environments. Despite the promising results, the study acknowledges limitations, including dataset diversity and environmental factors, and highlights future directions for model optimization to improve adaptability in broader applications.