A study on street walkability for older adults with different mobility abilities combining street view image recognition and deep learning - The case of Chengxianjie Community in Nanjing (China)

China is on the brink of transitioning into an aged society, resulting in a growing demand for an age-friendly street-built environment. However, previous research has paid limited attention to the differentiated walking needs of older adults. To address this gap, this study investigated the relationship between street-built environments and the subjective perception of older adults with different physical capabilities, focusing on safety, comfort, and interest. The older adults were classified into three types based on their physical mobility abilities. The TrueSkill algorithm was used to develop an online image selection website to obtain perception scores for sampled pictures from these three types of older adults. Image segmentation and deep learning were combined to extract indices of street view factors, and machine learning was used to train a scoring prediction model for all streetscape pictures of the area. The study found differences in the subjective perception among all three types of older adults, namely independent elderly (A), mediated-assisted elderly (B), and dependent elderly (C). Type A older adults might be attracted to factors related to the interest of walking despite their negative impact on safety and comfort; Type B older adults were more concerned about street conditions for safety and comfort. Type C older adults were prone to the convenience of barrier-free access and visibility. This study contributes to the study of walkability by providing a research framework for the subjective walking perceptions of older adults with different physical capabilities. Additionally, the visualized walkability map can serve as a reference for architects and urban designers, further strengthening the development of age-friendly communities with the aid of human-centric computational analysis, evaluation, and design.