Modeling Perceived Street Safety from Street View Imagery: Global and Local Perspectives on Street Space Features and Crash Incidences

Abstract. Road safety remains a global issue, with traffic crashes causing approximately 1.19 million deaths annually. While global datasets have been used to assess safety perception, limited studies examine how local perceptions vary and relate to street features and crash incidences. This study addresses that gap by modeling perceived street safety using Street View Imagery (SVI), comparing global (Place Pulse) and local (Quezon City, Philippines) perspectives. Models were developed to analyze how different populations perceive safety in relation to street space characteristics, and how these perceptions align spatially with crash incidences. A PSPNet-based semantic segmentation model extracted 28 features from SVIs which were modeled with perceptions from diverse groups defined by gender, road user role, and geographic context. Group-specific multiple linear regression models were developed to predict safety perception, and SHAPley Additive exPlanations (SHAP) interpreted feature influence. Results revealed natural and open-space elements like trees, sky, and sidewalk increased perceived safety, while vehicles, walls, and dense buildings reduced it. Local models outperformed global (R² up to 0.55 vs. ≈ 0.12), highlighting the value of localized, group-specific modeling. However, a perception–risk gap emerged: local respondents perceived crash-prone areas as safer than non-crash zones, whereas global perception aligned better with actual crash data. These findings emphasize that familiarity and environmental normalization can shape safety perception independently of real-world risk. The study highlights the limitations of relying solely on global data and advocates for equity-oriented urban design, prioritizing safe, inclusive, and context-aware public spaces that reflect lived realities of diverse communities.