Research on Fire Target Detection Algorithm Based on YOLOv11 Architecture

Abstract

Fire is a highly hazardous public safety event that not only causes significant casualties and property damage but also results in a profound negative impact on the ecological environment. Consequently, investigating methods to enhance the accuracy and detection speed of fire detection models holds substantial social and economic significance. Traditionally, fire detection technology has primarily relied on physical sensor components. Although these devices demonstrate high sensitivity in practical applications, their overall detection efficiency remains limited, and they are frequently susceptible to interference from complex external environments. With the continuous advancement of computer hardware performance and the increasing maturity of object detection technology based on computer vision algorithms, deep learning models with robust generalization capabilities and high detection precision have emerged as superior alternatives. This study explores a fire detection algorithm based on a one-stage object detection framework, specifically focusing on the You Only Look Once version 11 architecture. This advanced model can automatically extract fire characteristics from input images, facilitating an end-to-end detection methodology that significantly improves both the speed and accuracy of fire identification. By implementing optimization strategies such as cosine learning rate scheduling and mixed precision training, the research aims to achieve a balance between lightweight deployment and high-performance detection. Experimental results indicate that the proposed approach effectively identifies fire targets in diverse scenarios, providing a reliable technical foundation for real-time monitoring and early warning systems in large buildings, warehouses, and forest areas.