An Efficient And Precise Fire Detection Model Based on Improved RT-DETR

For fires, especially urban fires, early warning is crucial and can significantly reduce the loss of life and property among affected populations. Traditional fire detection algorithms suffer from issues such as missed detection, false detection, and inadequate real-time performance. Moreover, the scarcity of high-quality fire detection datasets severely limits the performance of these algorithms. To address these challenges, this paper presents a dual-pronged approach. On one hand, we develop the CFS fire dataset for model training. On the other hand, we propose a detection model called FDC-DETR, which based on improved RT-DETR architecture for detecting flames and smoke in complex environment. It incorporates Fasternet_Rep backbone, reconstruction of the Cross-Scale Feature Fusion Module(CCFM) component, Deformable Attention and WIoUv3 loss function. Experiments demonstrate that the detection model achieves an average precision (mAP) of 89.6% on the self-generated fire dataset. Compared to the original RT-DETR network, the model has reduced its parameters by 27.1%, increased accuracy by 2.7%, and improved average precision by 5.7%, without significant decrease in detection speed. These enhancements meet the requirements for practical security applications in fire detection.