Enhancing Pedestrian Route Choice with LLaVA-Generated Descriptions

Abstract Pedestrian navigation is influenced by various environmental factors such as safety, comfort, and accessibility, which have been widely discussed in the literature. While navigation applications may consider some of these aspects in their design, effectively conveying such information to users is still challenging. Presenting all potentially relevant factors can overwhelm users and reduce the efficiency and flexibility of decision-making due to interface limitations and clarity requirement. In addition, individual users may prioritize their own distinct environmental features, which highlights the need for more adaptable methods of information delivery. This study proposes pedestrian route description-LLaVA (PRDLLaVA), which enhances pedestrian navigation by providing route descriptions that convey relevant environmental information extracted from Google Street View images. With an overall accuracy of 87.60% in testing, the model can identify walkability-related factors when guided by structured prompts and fine-tuning. In addition, the generated route descriptions can depict the characteristics and changes in the pedestrian environment. The findings highlight PRD-LLaVA as a promising approach that balances efficiency and flexibility, which provides an efficient solution for pedestrian route selection.