A model for dynamic monitoring of urban fire with improved YOLOv8

Context: Urban fires are characterized by high suddenness and rapid spread, therefore dynamic monitoring is of great significance to urban safety. However, the existing detection methods for urban fire scenarios have problems such as poor real-time performance and low accuracy of small target recognition, which are difficult to meet the needs of early warning and emergency response for urban fires. Objective: In this paper, the existing YOLOv8 architecture is improved to deal with the dynamic monitoring of urban fires, so that the model's detection performance in multi-scale and multi-target fire scenarios is improved. Method: First, we introduce the Coordinate Attention mechanism to enhance the sensitivity to spatial localization of fire sources, and then we adopt the Transformer module to replace the SPPF structure in the original YOLOv8 to strengthen the feature modeling capability, and we add a new shallow detection head to improve the recognition accuracy of small fire sources. Result: We conduct a five-fold cross-validation experiment on a self-constructed urban fire image dataset, and the results show that the improved model significantly outperforms the original YOLOv8 in key metrics, including accuracy, recall and mAP, where the accuracy improves by about 4.0%, the recall improves by about 2.0%, and the mAP improves by about 5.0%, respectively. Meanwhile, in the comparison experiments with YOLOv5 and YOLOv11, the present model shows better detection stability and adaptability in different complex scenarios. Conclusion: This study is expected to provide technical support for intelligent firefighting dispatching and decision making in fire monitoring missions.