Scholarlink

<p>Under adverse weather conditions, images captured by cameras were often blurred, heavily occluded, and unevenly illuminated, making it harder to distinguish targets from the background. To enhance pedestrian detection performance in such environments, the CFMU-YOLOv10n algorithm was proposed. Firstly, the CPA-Enhancer adaptive enhancement module was designed. It dynamically adjusted strategies based on weather conditions to restore pedestrian contours and textures, providing high-quality input for subsequent modules. Secondly, a lightweight C2f-Faster module was constructed by improving the Bottleneck module in C2f with FasterNet’s FasterBlock. Computational complexity and parameters were reduced, and the computation cost of the enhancement module was balanced. Additionally, the original SPPF module was replaced with a FocalModulation module, which strengthened semantic discrimination and improved accuracy for hard-to-detect targets. Finally, the regression loss function was optimized to Unified-IoU, boosting convergence efficiency and bounding box localization precision. Experimental results showed that the mAP50 of CFMU-YOLOv10n on the AdverseWeatherPed dataset reached 80.4%, representing a 6.4 percentage points improvement over the original algorithm. Its mAP50-95 increased by 4.2 percentage points and accuracy by 5.1 percentage points, validating its effectiveness in pedestrian detection tasks under adverse weather. On the ExDark and RTTS datasets, mAP50 improved by 3.9 and 0.8 percentage points respectively, demonstrating the generalization ability of the improved algorithm.</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

CFMU-YOLOv10n: Multi-Module Collaborative Optimization for Robust Pedestrian Detection in Adverse Weather

  • 期刊论文

  • 发表日期：

    2025年12月15日

查看全文
  • Ren Gao,
  • Yu Gong,
  • Longyan Xu

摘要

Under adverse weather conditions, images captured by cameras were often blurred, heavily occluded, and unevenly illuminated, making it harder to distinguish targets from the background. To enhance pedestrian detection performance in such environments, the CFMU-YOLOv10n algorithm was proposed. Firstly, the CPA-Enhancer adaptive enhancement module was designed. It dynamically adjusted strategies based on weather conditions to restore pedestrian contours and textures, providing high-quality input for subsequent modules. Secondly, a lightweight C2f-Faster module was constructed by improving the Bottleneck module in C2f with FasterNet’s FasterBlock. Computational complexity and parameters were reduced, and the computation cost of the enhancement module was balanced. Additionally, the original SPPF module was replaced with a FocalModulation module, which strengthened semantic discrimination and improved accuracy for hard-to-detect targets. Finally, the regression loss function was optimized to Unified-IoU, boosting convergence efficiency and bounding box localization precision. Experimental results showed that the mAP50 of CFMU-YOLOv10n on the AdverseWeatherPed dataset reached 80.4%, representing a 6.4 percentage points improvement over the original algorithm. Its mAP50-95 increased by 4.2 percentage points and accuracy by 5.1 percentage points, validating its effectiveness in pedestrian detection tasks under adverse weather. On the ExDark and RTTS datasets, mAP50 improved by 3.9 and 0.8 percentage points respectively, demonstrating the generalization ability of the improved algorithm.