Hidden Weapon Detection and Tracking in Thermal Video Using Auto-Labeling and Deep Learning Approaches

This study proposes an automated framework for detecting and tracking concealed weapons in thermal video, aimed at real-time surveillance in high-security public areas like airports and stadiums. Due to the lack of relevant public datasets, thermal videos were recorded using an infrared camera in scenarios simulating concealed weapons. Frames extracted from the videos were automatically annotated using the GroundedSAM model, which aligns textual prompts ("gun", "knife") with image content, eliminating manual labeling. A YOLOv11n model was trained on over 6,200 labeled thermal images, achieving 81% precision, 82% recall, and 87% mAP@50. For tracking across frames, a Graph Neural Network (GNN) connected detected objects over time, 0.88 consistency for guns, and 0.66 consistency for knives. The integration of smart annotation, thermal-aware detection, and GNN tracking demonstrates strong potential for real-time, robust weapon detection in crowded, security-sensitive environments.