Synthetic and real-world datasets for crosswalk segmentation under diverse weather and lighting conditions

This article presents a new dataset for crosswalk segmentation targeting assistive technologies for visually impaired individuals. The dataset combines synthetic and real-world first-person view images with corresponding binary segmentation masks. The synthetic portion contains 3000 images generated using a fine-tuned Stable Diffusion model, with 1500 images created using a standard prompt ("a crosswalk image") and 1500 additional images incorporating various environmental conditions (sunny, cloudy, rainy, and night) through specialized prompts. The real-world component comprises 300 images extracted from chest-mounted smartphone video recordings of pedestrians approaching crosswalks, carefully distributed across different environmental conditions (120 sunny, 60 cloudy, 60 rainy, and 60 night images). To ensure diversity, each physical crosswalk location appears in at most two images from different approach directions. All images in both synthetic and real-world sets were manually annotated using a custom interface where annotators defined crosswalk regions as quadrilateral polygons, creating binary masks. The dataset is organized hierarchically by image source (synthetic/real-world) and environmental condition, with consistent subfolder structures for images and their corresponding masks. This dataset addresses the scarcity of publicly available crosswalk segmentation data with environmental diversity and has potential applications in developing and benchmarking computer vision algorithms for assistive navigation systems, investigating synthetic data augmentation efficacy, and advancing pedestrian safety technologies.