YOLO-HF: Early Detection of Home Fires Using YOLO

Domestic fires (residential and indoor) result in substantially higher deaths compared to other types of conflagrations, making early detection crucial to minimizing the loss of life and property. Existing studies and datasets predominantly focus on wildfire detection, with few datasets that capture the complexity of indoor environments and the subtle characteristics of early-stage fires. This gap causes object detection algorithms to produce high false positive and false negative rates in research on the early detection of such fires. To address this, we developed a domestic-fire dataset and proposed YOLO-HF, an early home-fire detection model based on YOLOv5s. To enhance the detection performance for small flames and smoke in a fire’s early stages, we proposed the parametric boosted channel attention (PBCA) mechanism. PBCA fuses global information and adaptively adjusts weights, improving the model’s ability to select and represent key features with minimal parameter overhead. Additionally, we incorporated the RepNCSPELAN4 feature extraction module, which preserves and leverages multi-scale feature information through multi-stage convolution and multi-layer feature fusion. To mitigate information loss caused by the stride setting in traditional downsampling methods, we employed space-to-depth (SPD) downsampling, which retains more spatial details. The experimental results demonstrate that YOLO-HF achieves a 4.6% improvement in Recall, 2.8% in mAP50, and 4.2% in mAP50-95 compared to existing models, all while maintaining its lightweight design. The model size is 13.8 MB, and its frame rate of 66.9 fps, meets real-time detection requirements. Detection experiments using content captured with domestic cameras validated the model’s generalization capability and robustness.