Scholarlink

<p>Fire detection and alarm technologies are critical components of intelligent safety systems, particularly in the domain of embodied AI, where agents must perceive, interpret, and respond to physical environments. Despite progress from traditional image processing to deep learning-based methods, existing fire detection approaches still suffer from low detection accuracy for small fire targets, high false negative rates, suboptimal real-time performance, and missing fire alarm information. To address these challenges within the context of embodied visual perception, we propose FA-YOLO, a novel fire detection and information extraction model based on the YOLOv5 framework. FA-YOLO incorporates a small-target detection head and a customized SPD-Conv module to enhance feature extraction for small-scale flames in low-resolution scenes. To enable efficient real-time deployment, we adopt Ghost modules and the CARAFE upsampling operator to significantly reduce computational cost. Furthermore, we design a lightweight fire information extraction module that supports auxiliary multi-object recognition, facilitating semantic scene understanding for embodied agents. Experimental results show that FA-YOLO achieves an average precision of 88.63%, with 74% fewer parameters and 71% lower computational complexity compared to YOLOv5s. These findings suggest that FA-YOLO is well-suited for deployment in real-time, vision-based fire alarm systems within embodied AI frameworks.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

FA-YOLO: fire alarm based on YOLO algorithm

  • 期刊论文

  • 发表日期：

    2025年8月13日

查看全文
  • Binglei Yue,
  • Yinming Shen,
  • Peihong Zhang,
  • Aili Jiang,
  • Yin Zhang

摘要

Fire detection and alarm technologies are critical components of intelligent safety systems, particularly in the domain of embodied AI, where agents must perceive, interpret, and respond to physical environments. Despite progress from traditional image processing to deep learning-based methods, existing fire detection approaches still suffer from low detection accuracy for small fire targets, high false negative rates, suboptimal real-time performance, and missing fire alarm information. To address these challenges within the context of embodied visual perception, we propose FA-YOLO, a novel fire detection and information extraction model based on the YOLOv5 framework. FA-YOLO incorporates a small-target detection head and a customized SPD-Conv module to enhance feature extraction for small-scale flames in low-resolution scenes. To enable efficient real-time deployment, we adopt Ghost modules and the CARAFE upsampling operator to significantly reduce computational cost. Furthermore, we design a lightweight fire information extraction module that supports auxiliary multi-object recognition, facilitating semantic scene understanding for embodied agents. Experimental results show that FA-YOLO achieves an average precision of 88.63%, with 74% fewer parameters and 71% lower computational complexity compared to YOLOv5s. These findings suggest that FA-YOLO is well-suited for deployment in real-time, vision-based fire alarm systems within embodied AI frameworks.