Scholarlink

<p>Accurately identifying tactile paving networks is crucial for enhancing the safety and ease of navigation for individuals with visual impairments. However, existing research primarily focuses on rapid detection of small-scale tactile paving elements using image data, with relatively fewer studies addressing the generation of geometric and attribute information for large-scale tactile paving network data. We propose SAFNet (<i>S</i>treet-view <i>A</i>nalysis and <i>F</i>usion <i>Net</i>work), a novel framework for generating vectorized tactile paving networks from street view images. SAFNet consists of two integrated components. The first is SA-DeepLabV3+. It is a spatial-channel attention-enhanced segmentation model that combines a CA-MobileNetV2 backbone with a lightweight SE-SP module to accurately detect tactile paving locations and their occupancy states (e.g., obstructions by vehicles). The second one is DF-MapGen. It is a data fusion module that maps these detected elements onto existing pedestrian vector networks, producing a topologically consistent, attribute-rich tactile paving network with geometric precision and occupancy semantics. Taking Wuhan city as a case study, experimental results show that the <i>MIOU</i>, <i>Recall</i>, and <i>Precision</i> in tactile paving recognition of SA-DeepLabv3 + achieves 85.60%, 94.69%, and 97.06%, respectively. The topological and geometric accuracies of the generated vector tactile paving network based on DF-MapGen are respectively 88.94% and 87.07%.</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

SAFNet: A Framework for Generating Tactile Paving Networks from Street View Images

  • 期刊论文

  • 发表日期：

    2026年4月13日

查看全文
  • Xue Yang,
  • Zihui Zhang,
  • Yi Zeng,
  • Zihan Kan,
  • Luliang Tang

摘要

Accurately identifying tactile paving networks is crucial for enhancing the safety and ease of navigation for individuals with visual impairments. However, existing research primarily focuses on rapid detection of small-scale tactile paving elements using image data, with relatively fewer studies addressing the generation of geometric and attribute information for large-scale tactile paving network data. We propose SAFNet (Street-view Analysis and Fusion Network), a novel framework for generating vectorized tactile paving networks from street view images. SAFNet consists of two integrated components. The first is SA-DeepLabV3+. It is a spatial-channel attention-enhanced segmentation model that combines a CA-MobileNetV2 backbone with a lightweight SE-SP module to accurately detect tactile paving locations and their occupancy states (e.g., obstructions by vehicles). The second one is DF-MapGen. It is a data fusion module that maps these detected elements onto existing pedestrian vector networks, producing a topologically consistent, attribute-rich tactile paving network with geometric precision and occupancy semantics. Taking Wuhan city as a case study, experimental results show that the MIOU, Recall, and Precision in tactile paving recognition of SA-DeepLabv3 + achieves 85.60%, 94.69%, and 97.06%, respectively. The topological and geometric accuracies of the generated vector tactile paving network based on DF-MapGen are respectively 88.94% and 87.07%.