Scholarlink

<p>The accurate detection of traffic signs by autonomous driving systems is critical for road safety. Although this task faces challenges such as a wide variety of categories, significant size variations, and strong environmental interference, traffic signs possess distinct characteristics, including bright colors, prominent corner features, and a spatial distribution where targets are typically located in the upper-right region of the image. To address these characteristics, this paper proposes the HLG-FPN feature pyramid network, which enhances multi-scale representation and improves small object detection by fusing high- and low-level semantic information with fine-grained features across scales. The HLG-FPN incorporates two key modules: HF Swin-T and C2f_EVC, forming an architectural design that combines intra-layer feature enhancement with multi-scale fusion. Specifically, HF Swin-T employs a window-based attention mechanism to focus on the upper-right region of the image, strengthening spatial-context modeling. Meanwhile, C2f_EVC captures corner regions, which are prevalent in traffic signs, thereby improving local structural sensitivity. Experimental results show that the HLG-FPN model achieved Precision, Recall, mAP@0.5, and mAP@.5:.95 of 84.4%, 73.6%, 83.0%, and 62.9% respectively on the TT100k dataset, representing improvements of 5.7%, 7.1%, 8.1%, and 6.2% over the Yolov8n model, demonstrating significant performance enhancement.</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

Traffic sign detection algorithm with multi-layer fusion and intra-layer emphasis

  • 期刊论文

  • 发表日期：

    2025年7月25日

查看全文
  • Rui Li,
  • Hu He

摘要

The accurate detection of traffic signs by autonomous driving systems is critical for road safety. Although this task faces challenges such as a wide variety of categories, significant size variations, and strong environmental interference, traffic signs possess distinct characteristics, including bright colors, prominent corner features, and a spatial distribution where targets are typically located in the upper-right region of the image. To address these characteristics, this paper proposes the HLG-FPN feature pyramid network, which enhances multi-scale representation and improves small object detection by fusing high- and low-level semantic information with fine-grained features across scales. The HLG-FPN incorporates two key modules: HF Swin-T and C2f_EVC, forming an architectural design that combines intra-layer feature enhancement with multi-scale fusion. Specifically, HF Swin-T employs a window-based attention mechanism to focus on the upper-right region of the image, strengthening spatial-context modeling. Meanwhile, C2f_EVC captures corner regions, which are prevalent in traffic signs, thereby improving local structural sensitivity. Experimental results show that the HLG-FPN model achieved Precision, Recall, mAP@0.5, and mAP@.5:.95 of 84.4%, 73.6%, 83.0%, and 62.9% respectively on the TT100k dataset, representing improvements of 5.7%, 7.1%, 8.1%, and 6.2% over the Yolov8n model, demonstrating significant performance enhancement.