Day-to-Night Street View Image Generation for 24-Hour Urban Scene Auditing Using Generative AI

Area of Science:

• Urban planning and geospatial analysis
• Computer vision and artificial intelligence
• Criminology and public safety

Background:

• Nighttime urban safety is a global concern, especially in large, complex cities with higher crime rates.
• Lack of nighttime street view imagery (SVI) has limited research on perceived safety and crime prevention.
• Generative AI (GenAI) offers a potential solution for creating needed nighttime SVI from existing daytime data.

Purpose of the Study:

• To test the hypothesis that generative AI can effectively create nighttime SVIs from daytime SVIs.
• To construct a comprehensive day-and-night SVI dataset across diverse urban landscapes.
• To develop and validate a day-to-night (D2N) transformation model for urban scene perception studies.

Main Methods:

• Collected pairwise day-and-night SVIs from four cities with varied urban forms.
• Trained and validated a D2N model with brightness adjustment for transforming daytime to nighttime SVIs.
• Evaluated the performance of different GenAI models (CycleGAN, Pix2Pix, StableDiffusion) for D2N conversion.

Main Results:

• D2N transformation accuracy is influenced by urban density, building, and sky view proportions.
• CycleGAN balances accuracy, data needs, and cost effectively for D2N conversion.
• Pix2Pix and StableDiffusion show varying performance based on data availability, quality, and computational cost.

Conclusions:

• The study successfully demonstrates GenAI's capability to generate nighttime SVIs from daytime data.
• A novel D2N dataset and generator model are established, supporting future urban environmental audits.
• This work provides a foundation for utilizing SVI in urban safety and crime prevention research.