Computer Engineering and Applications ›› 2023, Vol. 59 ›› Issue (17): 250-256.DOI: 10.3778/j.issn.1002-8331.2206-0097

• Graphics and Image Processing • Previous Articles     Next Articles

Helmet Wearing Detection Method Based on Improved CenterNet with Enhanced Associations

HUANG Pinchao, LIU Shijian, XU Ge, ZOU Zheng   

  1. 1.College of Computer Science and Mathematics, Fujian University of Technology, Fuzhou 350118, China
    2.Fujian Provincial Key Laboratory of Information Processing and Intelligent Control, Minjiang University, Fuzhou 350108, China
    3.College of Computer and Cyber Security, Fujian Normal University, Fuzhou 350117, China
  • Online:2023-09-01 Published:2023-09-01

关联增强改进的CenterNet安全帽检测方法

黄品超,刘石坚,徐戈,邹峥   

  1. 1.福建工程学院 计算机科学与数学学院,福州 350118
    2.闽江学院 福建省信息处理与智能控制重点实验室,福州 350108
    3.福建师范大学 计算机与网络空间安全学院,福州 350117

Abstract: Automatic helmet wearing detection on construction sites is an important way to ensure the safety of employees. The current challenges include that the helmet targets are often very small, dense and partially blocked, and it is difficult to balance accuracy and real-time, etc. To solve that, an improved CenterNet method with enhanced associations is proposed in this paper. Taking the advantage of the pixel-wise classification feature of the CenterNet, an association module is introduced to realize the fusion of deep and shallow features and compensate for information loss. Besides, a context-attention module is involved to guide the multi-level correlation enhancement, which further improves detection accuracy and reduces false detection rate. In addition, a lightweight strategy is adopted in stages to eliminate redundancy and streamline the network, which greatly reduces the model weights and improves the efficiency of the algorithm. The accuracy of the proposed method on a complex scene public data set is 88.6%, the average forward propagation time is 12?ms, and the average weight size is 19.5 MB, which outperforms the mainstream comparison methods. Experimental results prove that, with the excellent real-time performance and the high accuracy, the proposed method is suitable for helmet wearing detection in complex scenes.

Key words: object detection, CenterNet, attention mechanism, spatial pyramid pooling

摘要: 施工现场的安全帽佩戴状况自动化检测是保障员工安全的重要手段,目前所面临的挑战包括安全帽目标往往较小、密集且部分被遮挡,难以兼顾方法的精确度和实时性等。为此,提出一种关联增强的CenterNet改进方法。为充分发挥CenterNet逐像素分类的特点,引入关联融合模块来实现深、浅层特征的融合,弥补信息损失;同时使用上下文注意力提升模块来引导关联多级增强,进一步提升检测精度,降低误检率;此外,分阶段实施轻量化策略,剔除冗余、精简网络,极大降低权重规模、提升算法效率。该方法在复杂场景数据集上的准确率为88.6%,平均推理时间12?ms,平均权重大小19.5?MB,均优于主流对比方法。实验结果证明,该方法兼具强实时性与高准确度,适合复杂场景中的安全帽检测。

关键词: 目标检测, CenterNet, 注意力机制, 金字塔池化