A Fire Smoke Detection Method Based on YOLOv11s

As climate change progresses, the frequency of wildfires has increased dramatically, causing severe ecological damage. Timely and accurate fire detection has emerged as a pressing concern. This paper explores the effects of different parameters on model performance based on four datasets, namely the FIRE Dataset, Home Fire Dataset, Forest Fire Dataset and Wildfire Prediction Dataset. Meanwhile, by adopting the strategy of dataset fusion and progressive ablation, the contribution of each individual dataset is quantitatively analyzed. This work introduced the YOLOv11s architecture for the task of fire and smoke identification and conducted comparative experiments with several representative YOLO-series detectors. Extensive experimental findings fully verify that the presented YOLOv11s model displays superior performance in inference efficiency, detection precision, and model lightweight characteristics. Experimental findings demonstrate that the YOLOv11s framework designed in this study realizes optimal model compression efficiency and the lowest inference delay, processing each frame at a speed of merely 3.5 ms, with a detection precision of 93.2%, coupled with a recall ratio of 88.9%, a mean average precision (mAP) of up to 93.1%, and a corresponding F1-score of 91%. The proposed model is easy to deploy on the RK3588 platform. Field tests on the board have verified that the frame rate and peak memory usage meet the requirements for long-term, uninterrupted monitoring applications.