Intelligent assessment and diagnosis of street space quality using visual transformers

Addressing the demand for street-level spatial renewal in the context of high-quality urban development, this paper establishes an intelligent assessment and diagnosis framework for street-level spatial quality that integrates visual Transformers. First, we propose the Progressive Local Feature Matching model (Transformer-PLFM), which couples the FPN feature pyramid, multi-granularity convolutional features, and self-attention mechanisms. By introducing a boundary-aware context aggregation module (CAM) and a joint semantic-boundary-intermediate feature loss, it achieves high-precision semantic segmentation of street point clouds/images. Across the Paris-Lille-3D, Semantic3D, and Toronto-3D datasets, the proposed method demonstrates competitive performance against strong benchmark baselines, reaching 77.39% mIoU and 95.96% OA on Toronto-3D. It delivers optimal or near-optimal results for categories such as pedestrians, trees, and road markings, while also reducing RMSE and MAE in high-missing pedestrian sequence completion tasks, confirming its utility for modeling complex street scenes. Second, based on a questionnaire regarding spatial preferences in old-city streets, the TrueSkill algorithm was employed to convert pairwise comparison data into continuous 0–10 perception scores. A visual environment evaluation model was then constructed using random forest regression, achieving R² = 0.713, MSE = 0.016, and MAE = 0.095, and enabling batch perception scoring for 13,024 street-view images. Finally, combining factor analysis with empirical surveys of multi-type sample streets, we quantified the nonlinear threshold effects of visual elements—sky, sidewalks, trees, roads, walls, and buildings on perceived safety and comfort. Optimal ranges were identified, such as a road coverage of 23%–30% and tree coverage of 20%–40%. A street renewal strategy was proposed that balances green visibility, color design, and activity space organization, providing data support and technical pathways for refined design and classification optimization of existing neighborhoods.