Real-Time Firearm Detection in Video Using Enhanced YOLO Models and a Custom Incident Dataset
Catargiu Constantin, Iulian B. Ciocoiu · 2025
Gun violence remains a global threat, with attacks in schools, public events, and religious gatherings causing widespread loss of life. As these incidents grow in frequency and severity, there is a critical need for real-time firearm detection systems that enable rapid response. This study compares the performance of YOLOv, YOLOv10, YOLOv11, and YOLOv12 (medium versions) on firearm detection, evaluating accuracy, inference speed, and training time. We introduce a high-quality custom dataset of approximately$\mathbf{1 2, 5 0 0}$annotated images from 700 real-life video clips featuring gun-related incidents. Captured in varied indoor and outdoor environments under challenging conditions, the dataset is publicly available on the Roboflow platform. Advanced augmentation techniques were used to enhance model generalization capabilities. Experimental results show that all models achieved$\sim 90 \%$accuracy, with YOLOv12 outperforming others in complex scenes. Our findings confirm that modern YOLO architectures, when paired with realistic data, are well-suited for deployment in surveillance systems across domains such as smart cities, schools, law enforcement, and public transportation.