Enhancing Tactile Paving Segmentation via Multidimensional Perception Networks

Abstract

Abstract

Tactile paving is essential for ensuring the safe and independent travel of visually impaired individuals in urban environments. However, existing segmentation models often fail to generalize well across diverse scenarios and rely heavily on color information, neglecting the physical structure of tactile patterns. This article introduces EGM-Unet, a novel segmentation method that incorporates an edge-aware multi-scale attentional fusion block (EMAF-Block) to enhance edge feature extraction, a recursive gated attention (RGA) mechanism to focus on critical regions, and a multidimensional cooperative aggregation attention (MCAA) module to refine texture features. Additionally, we utilize CLIPSeg with a correlative self-attention (CSA) mechanism to improve generalization. Our experiments demonstrate that EGM-Unet outperforms standard U-Net and other mainstream models, achieving a mean IoU of 93.73%, an IoU of 89.09%, and an accuracy of 98.56%. These results highlight the robustness of our method in segmenting tactile paving regions across various materials and lighting conditions, providing a solid foundation for high-precision tactile paving perception.