Scholarlink

<p>In complex construction environments, existing algorithms exhibit significant limitations, particularly characterized by high false positive rates, frequent missed detections, and insufficient detection accuracy. To address these issues, this paper presents YOLOv10n-WDE (YOLOv10n-Wavelet Dynamic Enhancement), an enhanced helmet detection algorithm based on YOLOv10, which incorporates three key improvements. First, we developed the WFDConv (Wavelet-Frequency Dynamic Convolution) module, which integrates discrete wavelet transforms with dynamic convolution to significantly enhance the ability to capture multi-scale features, thereby improving precision and reducing false positive rates. Second, we introduced a lightweight parallel Spatial Pyramid Pooling Network (LPSPPF) that boosts feature extraction efficiency through a parallel architecture, enhancing the detection capability for small targets and consequently improving recall while minimizing missed detections. Lastly, we implemented a joint loss function mechanism that combines Focal Loss for bounding box regression with Varifocal Loss for classification optimization, thereby improving the model’s overall accuracy in complex scenarios. Experimental results show that on the SHWD dataset, YOLOv10n-WDE achieves mAP50 improvements of 5.1% and 1.8% over YOLOv8n and YOLOv10n, respectively. Its precision and recall reach 92.9% and 87.6%, both surpassing those of YOLOv8n (91.0% and 87.5%) and YOLOv10n (89.6% and 90.4%). On the SHDD dataset, compared with YOLOv10n, YOLOv10n-WDE improves precision by 5.3%, recall by 1.9%, and mAP50 by 3.7%. These enhancements fully demonstrate their effectiveness in reducing false positives and missed detections. At the same time, YOLOv10n-WDE maintains a real-time processing speed of 384 FPS, meeting the dual demands for efficiency and real-time performance in complex construction environments.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Improved YOLOv10-based real-time helmet detection algorithm for complex scenarios

  • 期刊论文

  • 发表日期：

    2025年10月6日

查看全文
  • HanTang Dong,
  • Yong Wang,
  • Duoqian Miao

摘要

In complex construction environments, existing algorithms exhibit significant limitations, particularly characterized by high false positive rates, frequent missed detections, and insufficient detection accuracy. To address these issues, this paper presents YOLOv10n-WDE (YOLOv10n-Wavelet Dynamic Enhancement), an enhanced helmet detection algorithm based on YOLOv10, which incorporates three key improvements. First, we developed the WFDConv (Wavelet-Frequency Dynamic Convolution) module, which integrates discrete wavelet transforms with dynamic convolution to significantly enhance the ability to capture multi-scale features, thereby improving precision and reducing false positive rates. Second, we introduced a lightweight parallel Spatial Pyramid Pooling Network (LPSPPF) that boosts feature extraction efficiency through a parallel architecture, enhancing the detection capability for small targets and consequently improving recall while minimizing missed detections. Lastly, we implemented a joint loss function mechanism that combines Focal Loss for bounding box regression with Varifocal Loss for classification optimization, thereby improving the model’s overall accuracy in complex scenarios. Experimental results show that on the SHWD dataset, YOLOv10n-WDE achieves mAP50 improvements of 5.1% and 1.8% over YOLOv8n and YOLOv10n, respectively. Its precision and recall reach 92.9% and 87.6%, both surpassing those of YOLOv8n (91.0% and 87.5%) and YOLOv10n (89.6% and 90.4%). On the SHDD dataset, compared with YOLOv10n, YOLOv10n-WDE improves precision by 5.3%, recall by 1.9%, and mAP50 by 3.7%. These enhancements fully demonstrate their effectiveness in reducing false positives and missed detections. At the same time, YOLOv10n-WDE maintains a real-time processing speed of 384 FPS, meeting the dual demands for efficiency and real-time performance in complex construction environments.