基于改进YOLO的安全帽佩戴检测算法研究

Research on Helmet Wear Detection Algorithm Based on Improved YOLO

高危场景下的安全帽佩戴检测任务是保障施工人员生命财产安全的重要一环。本文针对复杂场景下传统安全帽检测中小目标检测性能不佳、模型参数量大等问题,提出一种基于改进YOLOv7-Tiny的安全帽检测方法。该方法设计一种针对小目标检测的特征提取网络,从而加强网络对小目标的识别能力;引入轻量化上采样算子CARAFE,改进上采样过程从而提高特征融合质量;使用K-means++聚类算法对数据集中目标的锚框进行重新聚类,得到适应安全帽检测的锚框尺寸;并引入CA注意力机制,让网络更加关注有用的坐标信息,加强网络对小目标的识别能力。实验结果表明,在公开数据集SHWD上,改进后的算法相比YOLOv4-Tiny、YOLOX-Tiny等传统算法精度分别提升了14.1%和2.8%,相比原算法在参数量减少了31.7%的同时精度提升了1.5%,满足实际场景中对安全帽的检测要求。

The task of helmet wear detection in high-risk scenarios is an important part of safeguarding the lives and properties of construction workers. In this paper, we propose a helmet detection method based on an improved YOLOv7-Tiny, addressing challenges related to the poor performance of small target detection and the large number of model parameters associated with traditional helmet de-tection in complex scenes. Our method incorporates a feature extraction network designed specifically for small target detection, enhancing the network’s capability to recognize small targets. Addi-tionally, we introduce a lightweight upsampling operator, CARAFE, to optimize the up-sampling process, thereby improving the quality of feature fusion. To adapt anchor frames to helmet detec-tion, we employ the K-means++ clustering algorithm to recluster the anchor frames of targets in the dataset. Furthermore, we integrate the CA attention mechanism, allowing the network to prioritize useful coordinate information and reinforce its ability to recognize small targets. Experimental results on the SHWD public dataset demonstrate that the improved algorithm enhances mAP by 14.1% and 2.8% compared to traditional algorithms like YOLOv4-Tiny and YOLOX-Tiny, respec-tively. Moreover, it improves mAP by 1.5% compared to the baseline while reducing the number of parameters by 31.7%. These results meet the requirements for safety helmet detection in real sce-narios.