Computer Engineering and Applications ›› 2022, Vol. 58 ›› Issue (14): 114-120.DOI: 10.3778/j.issn.1002-8331.2202-0181

• Pattern Recognition and Artificial Intelligence • Previous Articles     Next Articles

FPN-Centernet Helmet Wearing Detection Algorithm

ZHAO Jianghe, WANG Hairui, WU Lei   

  1. Faculty of Information Engineering and Automation, Kunming University of Science and Technology, Kunming 650504, China
  • Online:2022-07-15 Published:2022-07-15

FPN-CenterNet安全帽佩戴检测算法

赵江河,王海瑞,吴蕾   

  1. 昆明理工大学 信息工程与自动化学院,昆明 650504

Abstract: As the safety guarantee of workers in the construction site, the wearing of safety helmet affects the life safety of workers. In terms of wearing detection, the introduction of deep learning can effectively remind workers to wear safety helmets. However, because the image of safety helmet in the construction image is too small, CenterNet does not perform well. Therefore, in view of this situation, FPN-CenterNet framework is proposed. Then, ACNet(asymmetric convolution kernel)is utilized to enhance the feature extraction of the backbone network. Finally, DIoU loss function is used to optimize the accuracy of frame prediction. Compared with the original CenterNet algorithm mAP, the final modified algorithm improves 4.99 percentage points, and the FPS on the GTX GeForce 1050 GPU reaches 25.81. Experimental results show that the modified algorithm has good accuracy and efficiency in helmet wearing detection.

Key words: safety helmet wearing detection, feature pyramid net, asymmetric convolution kernel, DIoU loss function

摘要: 安全帽作为施工场所工人的安全保障,佩戴与否影响着工人的生命安全。在佩戴检测方面引入深度学习可以高效地提醒工人佩戴安全帽。但由于施工图像中安全帽的图像过小,CenterNet表现得并不好。因此针对这个情况,提出了FPN-CenterNet框架;使用ACNet非对称卷积核来对主干网络的特征提取进行增强;使用DIoU损失函数来优化边框预测的准确度。最终修改的算法相较于原始的CenterNet算法mAP提升了4.99个百分点,在GTX GeForce 1050的GPU上的FPS达到25.81。实验结果表明修改之后的算法在安全帽佩戴检测上有良好的准确性和效率。

关键词: 安全帽佩戴检测, 特征金字塔, 非对称卷积核, DIoU损失函数