改进 YOLOv5 算法在停车场火灾检测中的应用-《郑州大学学报(工学版)》

文章信息/Info

Title:: Application of Improved YOLOv5 Algorithm in Parking Lot Fire Detection

作者:: 张震¹; 晋志华¹; 陈可鑫²; 1.郑州大学电气与信息工程学院,河南郑州 450001, 2.郑州大学计算机与人工智能学院,河南郑州 450001

Author(s):: ZHANG Zhen¹; JIN Zhihua¹; CHEN Kexin²; School of Electrical and Information Engineering, Zhengzhou University, 450001, Zhengzhou, Henan, School of Computer and Artificial Intelligence at Zhengzhou University, 450001, Zhengzhou, Henan

Keywords:: underground parking; fire detection; YOLOv5; coordinate attention; CIoU loss function

摘要:: 针对传统传感器对于地下停车场火灾检测不及时、目标检测对小型火焰目标检测效果较差等问题,提出了一种改进的 YOLOv5 火灾检测算法。为了提高检测算法对小型火焰目标的检测效果,在 YOLOv5s 网络骨干中增加小目标检测层;为了增强火焰特征的表达,提出了一种基于 CA 注意力机制的间隔注意力结构;为了提升定位精度、降低目标漏检率,将 GIoU 替换为 CIoU。设计了 3 组消融实验以及 1 组对比实验用来验证所提算法的有效性。实验结果表明,所提算法在自定义数据集上的 mAP0. 5 、召回率 R 分别为 92%、96. 9%。与 YOLOv5s 模型相比,所提算法在自定义火焰数据集上的 mAP0. 5 提升了 1. 8 百分点,R 提升了 2. 0 百分点。所提算法权重大小仅为 16. 4 MB, 帧率能达到 113 帧 / s,具有较小的模型体积以及较快的检测速度,且针对小型火焰目标能够准确检出,有效提升了地下停车场火灾防范能力。

Abstract:: Aiming at the problems that the traditional sensors are not timely in detecting the fire of underground parking lot and the object detection is not effective in detecting small flame targets, an improved YOLOv5 fire detection algorithm was proposed. In order to improve the detection effect of the detection algorithm on small flame targets, small target detection layer was added to YOLOv5s network backbone; In order to enhance the expression of flame features, an interval attention structure based on CA attention mechanism was proposed; In order to improve the positioning accuracy and reduce the rate of missed target detection, GIoU was replaced by CIoU. Three groups of ablation experiments and one group of contrast experiment were designed to verify the effectiveness of this improvement. The experimental results show that the mAP0. 5 and R of the algorithm on the user-defined dataset are 92% and 96. 9%, respectively. Compared with YOLOv5s model, the mAP0. 5 of the proposed algorithm on the customized flame data set is increased by 1. 8 percentage, and R is increased by 2. 0 percentage. The weight size of the proposed algorithm is only 16. 4 MB, and the frame rate can reach 113 frames per second. It has smaller model volume and faster detection speed, and can accurately detect small flame targets, effectively improving the fire prevention capability of the underground parking lot.