Datasets featuring tactile paving scenes play a critical role in training assistive navigation systems for visually impaired individuals. Current tactile paving datasets often contain pictures of typical urban scenes with similar paving styles. However, existing datasets commonly exhibit three major shortcomings: (1) a limited variety of tactile paving types, (2) insufficient diversity in environmental conditions such as lighting and weather, and (3) a lack of scene-level descriptions that support holistic understanding. These limitations hinder both semantic and spatial comprehension by vision-based models and restrict their real-world generalization ability. To solve these problems, we introduce TactPav, the first dataset combining semantic segmentation labels and image captions for tactile paving scenes. It features 3,601 diverse street view images with pixel-level annotations and accessibility-focused captions. To produce high-quality captions, we propose TactiCap, a structured four-stage pipeline that involves: (1) structured information extraction, (2) coherent caption synthesis, (3) human verification and correction, and (4) final refinement to ensure clarity and relevance. Experimental results demonstrate that models trained on our dataset achieve improved generalization in tactile paving segmentation across diverse environments. Moreover, evaluations of state-of-the-art image captioning models reveal notable limitations in producing accurate and context-aware descriptions.