Scholarlink

<p>In autonomous driving navigation systems, semantic segmentation plays a crucial role in understanding complex road scenes. However, existing methods still face limitations in handling multi-scale objects, ambiguous boundaries, and modeling contextual information. The proposed DDANet, based on the DDRNet dual-resolution architecture, enhances segmentation accuracy and robustness through three key modules. Its main contributions include: (1) To address information deterioration in conventional additive fusion, we design a Bidirectional Cross-Fusion Module (BCF) using adaptive weighting techniques; (2) we introduce an Efficient Channel Attention-enhanced DAPPM (ECA-DAPPM), which adaptively assigns feature weights to improve multi-scale context awareness; (3) a Combined Boundary Module (CBM) combining Sobel edge detection with lightweight convolutions was proposed to achieve refined boundary modeling. Extensive experiments based on the Cityscapes and CamVid datasets demonstrate that DDANet achieves an mIoU reaching 78.4% and 77.3%, respectively, representing increases of 1.5% and 2.6% over baseline models.</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

DDANet: enhancing context and boundaries for semantic segmentation in autonomous driving

  • 期刊论文

  • 发表日期：

    2026年3月2日

查看全文
  • Weixi Wang,
  • Wanfang Yu,
  • Ye Liu,
  • Zhouzhou Cui,
  • Liang Shan,
  • Lanxin Xie,
  • Yutao Song

摘要

In autonomous driving navigation systems, semantic segmentation plays a crucial role in understanding complex road scenes. However, existing methods still face limitations in handling multi-scale objects, ambiguous boundaries, and modeling contextual information. The proposed DDANet, based on the DDRNet dual-resolution architecture, enhances segmentation accuracy and robustness through three key modules. Its main contributions include: (1) To address information deterioration in conventional additive fusion, we design a Bidirectional Cross-Fusion Module (BCF) using adaptive weighting techniques; (2) we introduce an Efficient Channel Attention-enhanced DAPPM (ECA-DAPPM), which adaptively assigns feature weights to improve multi-scale context awareness; (3) a Combined Boundary Module (CBM) combining Sobel edge detection with lightweight convolutions was proposed to achieve refined boundary modeling. Extensive experiments based on the Cityscapes and CamVid datasets demonstrate that DDANet achieves an mIoU reaching 78.4% and 77.3%, respectively, representing increases of 1.5% and 2.6% over baseline models.