A child-centric analysis of community street space quality: Leveraging street view imagery and machine learning

This study investigates the correlation between the street space environment and children's subjective perceptions, a relationship which is critically important for creating child-friendly community public spaces. We first established a quality evaluation system based on children's behavior and psychology. Street view images, captured from children's eye height, were analyzed via semantic segmentation using the Cityscapes Dataset and the SegFormer algorithm to identify built environment elements. The K-means clustering and visual entropy algorithms were employed to extract street spatial features. Concurrently, the Elo image preference selection algorithm was used to quantitatively score children's subjective perceptions. These quantified objective and subjective datasets then served as the input and output variables, respectively, for a machine learning model. We applied a random forest regression algorithm to construct an evaluation model that accurately reflects street-space quality from the perspectives of children. An empirical study was conducted in several communities in Nanjing, and ArcGIS was used to visualize the evaluation results by linking them to the geographic coordinates of the measurement points. Our findings reveal significant differences in child-friendliness among communities from different construction periods, as well as notable variations in spatial quality among different street types within the same community. While street space quality results from the combined effect of multiple factors, the impact of different environmental elements on child-friendliness varies considerably.