FPVCrosswalk2025: A dataset for first-person view crosswalk segmentation in adverse weather and lighting conditions

This dataset supports crosswalk segmentation for use in assistive navigation technologies for the visually impaired. It includes 3000 synthetic and 300 real-world first-person view (FPV) images, each paired with a binary segmentation mask. The synthetic set was generated using a fine-tuned Stable Diffusion model, with prompts covering general and varied environmental conditions (sunny, cloudy, rainy, and night). The real-world images are distributed across the same environmental conditions and extracted from video recordings of pedestrians approaching crosswalks. Each crosswalk appears in no more than two images from different directions to ensure diversity. All images were manually annotated using a custom tool that defines crosswalks as quadrilateral regions. The dataset is organized by source (synthetic/real) and environmental conditions, with matching folder structures for images and masks. It can be used for training and evaluating computer vision models, exploring synthetic data augmentation, and improving assistive systems for visually impaired pedestrians.本数据集可支撑视障人士辅助导航技术中的人行横道分割任务。其包含3000张合成图像与300张真实世界第一视角（first-person view, FPV）图像，每幅图像均配有一幅二值分割掩码。合成数据集通过微调后的Stable Diffusion模型生成，所用提示词覆盖通用且多样的环境场景，包括晴天、阴天、雨天与夜间。真实世界图像同样涵盖上述四类环境条件，提取自行人趋近人行横道的视频录制素材。为保障数据多样性，每个人行横道仅在最多两幅来自不同拍摄方向的图像中出现。所有图像均通过定制化标注工具完成人工标注，标注规则将人行横道界定为四边形区域。数据集按来源（合成/真实）与环境条件进行分类组织，图像与对应掩码的文件夹结构保持统一。本数据集可用于计算机视觉模型的训练与评估、合成数据增强技术的探索，以及面向视障行人的辅助系统优化工作。