Computer Engineering and Applications ›› 2024, Vol. 60 ›› Issue (11): 298-308.DOI: 10.3778/j.issn.1002-8331.2307-0003

• Engineering and Applications • Previous Articles     Next Articles

Improved YOLOv5s Flame and Smoke Detection Method for Underground Garage

DU Chen, WANG Xing, DONG Zengshou, WANG Yilei, JIANG Zhonghao   

  1. 1.College of Computer Science and Technology, Taiyuan University of Science and Technology, Taiyuan 030024, China
    2.College of Electronic Information Engineering, Taiyuan University of Science and Technology, Taiyuan 030024, China
  • Online:2024-06-01 Published:2024-05-31

改进YOLOv5s的地下车库火焰烟雾检测方法

杜辰,王兴,董增寿,王亦雷,江忠浩   

  1. 1.太原科技大学 计算机科学与技术学院,太原 030024
    2.太原科技大学 电子信息工程学院,太原 030024

Abstract: Aimed at the problems that the traditional underground garage fire detection is not timely and can’t give detailed fire information, the target detection is difficult to detect small target smoke and flame, and the accuracy of other smoke and flame detection is low, an improved YOLOv5s underground garage smoke and flame detection algorithm is proposed. Attention mechanism is added to the last C3 module of YOLOv5s backbone network to help the network model extract the multi-scale spatial information and important features of smoke and flame more fully. The Neck part is improved to enhance the ability of feature interaction and small target smoke detection. The convolution module of backbone network is improved to improve the ability of smoke and flame feature extraction. WIoU (wise intersection over union) is introduced as a new bounding box loss function to enhance the generalization ability of the model. Soft NMS (soft non-maximum suppression) is introduced to enhance the detection ability of overlapping smoke and flame. Comparative experiments are carried out on the self-made smoke and flame data set, and the results show that the weight of the improved model is reduced by 0.2 MB, and the accuracy is improved by 6.8 percentage points, which can meet the requirements of smoke and flame detection in underground garages.

Key words: smoke and fire detection, YOLOv5s, attention mechanism, small target smoke and flame detection

摘要: 针对传统地下车库火灾检测不及时且无法给出火灾详细信息、目标检测对小目标烟火检测困难、其他烟火检测精度低等问题,提出了一种改进YOLOv5s的地下车库烟火检测算法。在YOLOv5s主干网络的最后一个C3模块中加入了注意力机制,帮助网络模型更充分地提取烟火多尺度空间信息和重要特征;对Neck部分进行改进,增强特征交互和小目标烟火检测能力;对主干网络卷积模块进行改进,提高烟火特征提取能力;引入WIoU(wise intersection over union)作为新的边界框损失函数,增强模型泛化能力;引入Soft NMS(soft non-maximum suppression)用以增强重叠烟火检测能力。在自制烟火数据集上进行对比实验,结果表明,改进后的模型权重减小0.2?MB,精度提升了6.8个百分点,能够满足地下车库烟火检测要求。

关键词: 烟火检测, YOLOv5s, 注意力机制, 小目标烟火检测