Abstract: To address the issue of missed and incorrect detections of small objects in helmet-wearing detection within construction scenarios, caused by crowding, occlusion, and complex backgrounds, a cross-layer multi-scale helmet detection algorithm with a dual attention mechanism based on YOLOv8n was proposed. First, small object detection head was designed to enhance the model's ability to detect small objects. Second, Double Attention mechanism was embedded in feature extraction network to enhance capture target features in complex scenarios. Third, the feature fusion network was replaced with the S-GFPN (Selective Layer Generalized Feature Pyramid Network), improved with the RepGFPN (Re-parameterized Generalized Feature Pyramid Network), to enable multi-scale fusion of small object feature layers with other layers and establish long-term dependencies, thus reducing background interference. Finally, the MPDIOU (Intersection over Union with Minimum Points Distance) loss function was employed to address sensitivity issues related to scale changes. Experimental results on the GDUT-HWD public dataset show that the improved model increases the mAP@0.5 by 3.4 percentage points compared to the baseline. Detection accuracy for various helmet colors improves by 2.0, 1.1, 4.6, and 9.1 points, respectively. The model also outperforms the baseline in five complex scenarios: density, occlusion, small targets, glare, and darkness. This provides an effective method for helmet-wearing detection in real-world construction scenarios.

Key words: complex scenarios, object detection, small objects, multi-scale feature fusion, You Only Look Once v8 &#40

摘要： 为解决施工场景下安全帽佩戴检测时，由于人员密集、遮挡和复杂背景等原因造成小目标出现漏检和错检的问题，提出了一种基于YOLOv8n（You Only Look Once v8n）的双重注意力机制的跨层多尺度安全帽佩戴检测算法。首先，设计增加微小目标检测头，提高模型对小目标的检测能力；其次，在特征提取网络中嵌入Double Attention注意力机制，更关注复杂场景下目标信息的特征捕获；再者，将特征融合网络替换成基于重参数化泛化特征金字塔网络（RepGFPN）改进后的跨层多尺度特征融合（Selective Layer Generalized Feature Pyramid Network, S-GFPN）结构，实现小目标特征层信息和其他特征层的多尺度融合，建立长期的依赖关系，从而抑制背景信息的干扰；最后，采用MPDIOU（Intersection over Union with Minimum Points Distance）损失函数解决尺度变化不敏感的问题。在公开数据集GDUT-HWD上实验结果表明，改进后的模型比基线模型mAP@0.5的值提升了3.4个百分点，对各色安全帽的检测精度值分别提升了2.0、1.1、4.6、9.1个百分点，在密集、遮挡、小目标、反光和黑暗五类复杂场景下可视化检测效果也优于基线模型，为实际施工场景中安全帽佩戴检测提供了一种有效方法。

关键词: 复杂场景, 目标检测, 小目标, 多尺度特征融合, YOLOv8