Research on Smoke and Fire Detection with Tracking Based on Deep Learning

Fires represent one of the most urgent global challenges, threatening human lives, ecosystems, and economic stability. The development of robust early detection systems are critical for minimizing casualties, reducing property damage, and preserving ecological balance. This paper proposes a novel video-based smoke and fire detection algorithm that synergistically exploits both dynamic and static features. First, a smoke and fire detection framework is constructed based on YOLOv10m, we re-engineer the Backbone by integrating DCNv3, replacing conventional CIB modules. This empowers the model to adaptively capture irregular fire morphologies, thereby enhancing feature extraction capability. Second, the SEAM is introduced to improve the Head, addressing occlusion challenges and strengthening feature fusion. Additionally, to reduce false alarms caused by static luminous interferences that exhibit fire-like color and brightness (e.g., streetlights and vehicle lights), we propose an improved BoT-SORT algorithm incorporating the MSE. This method leverages dynamic characteristics of fire (e.g., spreading and diffusion) to distinguish it from static interferences. Experimental results demonstrate that the improved detection algorithm achieves a 1.6% increase in mAP. The enhanced BoT-SORT reduces the misreport rate by 88% in interference videos while maintaining the detection rate of fire videos, effectively resolving false alarms in the detection stage.