| 總結: | Safety perception is widely considered a fundamental aspect of urban life, which significantly influences citizens' well-being and quality of life as well as having crucial impact on the nighttime economy. However, there is a scarcity of understanding of nighttime safety despite the fast-growing body of urban scene auditing research based on daytime street view imagery (SVI). To fill the gap, this study collected ∼1000 pairwise day-and-night SVIs to train a day-to-night (D2N) SVI generator to effectively predict nighttime SVI based on daytime counterpart using generative adversarial network (GAN). The accuracy of fake nighttime image was evaluated with commonly-used GAN metrics (e.g., structural similarity index, inception score) and human validation. Then, an online visual survey with 46 participants was conducted to collect their perceived safety on street scenes during daytime and nighttime (D&N), and the results become training labels for machine learning models to predict D&N safety perceptions. Our results revealed significant discrepancies in D&N safety perception. First, through correlation analysis, we found that the sky and building features matter to the prediction accuracy of generated nighttime SVIs. Second, the micro-level streetscape features (e.g., pavements, roads, and buildings) play influential roles in perceived safety. Third, higher safety perceptions are consistently found in areas with higher building density regardless of whether they are daytime or night. In contrast, untended trees and grass reduce perceived safety at night. This study provides a valuable reference for improving the accuracy of generating nighttime images from daytime SVIs. It also reveals how streetscapes affect D&N safety perceptions in high-density cities like Hong Kong, providing empirical evidence for urban design policies to facilitate nighttime attractiveness and prosperity.
|
|---|
