Real-Time Early Warning of Incipient Fire in Multiple Urban Scenarios: A Deep Learning-Based Monitoring Method

Urban fire incidents in complex built environments pose severe threats to public safety. However, the unstructured nature of urban scenes presents substantial challenges for existing detection algorithms in reliably identifying incipient flames and diffuse smoke under dynamic visual interference. To address this issue, we propose YOLO-Fire, a lightweight and high-precision detection algorithm based on YOLOv11. Specifically, a Hybrid Feature Fusion Module (HFFM) adopts a parallel dual-stream architecture to structurally decouple high-frequency flame boundaries from low-frequency smoke textures. A Dual-Scale Contextual Diffusion (DCD) mechanism establishes global contextual constraints through an additive diffusion strategy, effectively suppressing fire-like background interference while enhancing semi-transparent smoke features. In addition, a Gaussian Spatial Pyramid Pooling Fast (GSPPF) module further improves multi-scale receptive field aggregation. Evaluated on a self-constructed large-scale urban fire dataset, YOLO-Fire achieves an mAP50 of 75.7%, mAP50-95 of 53.3%, and an F1-score of 73.7%, with only 10.02 M parameters, surpassing the YOLOv11 baseline by 2.4%, 4.5%, and 2.9%, respectively. Ablation studies confirm that each proposed module contributes both independently and synergistically to the overall performance gains. Comprehensive comparisons with mainstream detectors and specialized fire detection models further demonstrate that YOLO-Fire achieves superior overall performance, outperforming YOLO-FireAD and FireSmoke-YOLO by 2.7% and 2.4% in mAP50, respectively, while maintaining lower computational complexity. Furthermore, inference evaluation on a single-core CPU achieves 17.28 FPS, validating the practical deployment potential of YOLO-Fire in resource-constrained environments and offering an efficient, lightweight solution for real-time urban fire surveillance and early warning.