Research on Visual Algorithm for Fire Detection of Firefighting UAVs Based on Infrared Imaging

Firefighting drones are an effective means for high-rise building fire rescue. While extensive research has been conducted on flame detection using visible light sensors, these methods are significantly influenced by environmental factors such as weather conditions and thick smoke, thereby diminishing the effectiveness of firefighting drones in rescue operations. Due to limitations in available datasets, studies utilizing non-visible light sensors predominantly employ traditional multi-feature extraction algorithms, resulting in compromised detection accuracy and speed, which fail to match the real-time requirements of firefighting drones. Addressing these challenges, this study proposes a thermal imaging flame detection model, YOLOv5-EfficientNet, which is an improvement upon YOLOv5s v6.0. The backbone of YOLOv5 is replaced with EfficientNet, and the detected target information is further verified using infrared sensing technology. Additionally, a high-quality dataset comprising 6652 images of thermal flames and 292 images of thermal-like fires is constructed based on the COCO dataset standard. Results obtained from testing with this dataset demonstrate a 3.3% increase in mean Average Precision (mAP) relative to YOLOv5s. Both the detection accuracy and parameters of this model meet the real-time detection requirements of firefighting drones. Compared to traditional visible light sensors, infrared thermal imaging sensors exhibit superior environmental adaptability and interference resistance. This detection system combines the advantages of thermal imaging and deep learning, offering a novel approach for flame detection in firefighting unmanned devices.