Towards Robust Urban Obstacle Perception for Visually Impaired Navigation: A Dataset and Baseline Study

Abstract

Urban environments pose significant mobility challenges for visually impaired individuals due to the presence of diverse obstacles and complex scenes, while many existing assistive navigation approaches rely on generic datasets or simplifying assumptions that limit their applicability to real-world urban use. This paper presents a foundational study toward urban obstacle perception for visually impaired navigation by introducing a custom vision-based dataset comprising eight urban obstacle classes relevant to assistive mobility, manually annotated in YOLO format and reflecting realistic multi-object scenes. A lightweight YOLO-based model is trained as a baseline to assess detection performance on the proposed dataset. Experimental results demonstrate the feasibility of multi-class urban obstacle detection while highlighting challenges related to scale variability and scene complexity. This work establishes an initial benchmark and dataset to support future research on robust and deployable assistive navigation systems.