To address the challenges of high subjectivity, difficult data acquisition, and low efficiency in current evaluation methods for child-friendly urban streets, this study proposes a deep learning-based evaluation system that integrates both concrete and abstract features. The study utilizes 1,322 street samples in Shanghai, integrating 50 quantifiable concrete features with abstract features extracted from 6,724 street view images. Perceptual survey data from children aged 7–12 were incorporated as the target output for model training. Methodologically, abstract features were first extracted from streetscapes using a ResNet18 convolutional neural network. These features were then fused with the concrete features, and a multi-layer artificial neural network was constructed to predict child-friendliness. The results demonstrate that the model achieved an average accuracy of 96.91% on the validation set and an overall accuracy of 97.35% on the test set, indicating its effectiveness in identifying street samples with low levels of child-friendliness. Further case validation demonstrates the model’s capability to rapidly identify child-unfriendly spatial characteristics at an urban scale, including poor traffic safety, inadequate pedestrian environments, and a lack of engaging elements. This study offers a novel technical pathway for the quantitative evaluation and targeted management of child-friendly streets.
Tu et al. (Mon,) studied this question.