Using street view imagery and localized crowdsourcing survey to model perceived safety of the visual built environment by gender

Abstract

Scholars have documented that perceived safety of the visual built environment (VBE) can influence human behaviors. The dual developments of street view imagery (SVI) and deep learning techniques offer a cost-effective approach to measure perceived safety. However, current SVI-based perception models often lack specific definitions of perceived safety and demographic information when collecting data for model training. Furthermore, these models are rarely validated by onsite perception evaluations, which undermines their credibility.

Given these gaps, this study builds a localized crowdsourcing survey to train crime-related and barrier-related perceived safety of the VBE captured by SVIs, and compares model-predicted perceptions with onsite perceptions. This study specifically focuses on their ability to represent onsite perceptions and examines gender differences as a test case in safety perception. This study recruits over 1,800 participants living in the Greater Toronto Area to rate SVIs in terms of crime-related and barrier-related perceived safety.

Pearson correlation coefficients reveal a positive but weak correlation between female and male safety perceptions, indicating some consistency while highlighting potential gender differences in safety perceptions. Machine-learning perception models are then trained using this localized SVI survey. Model-predicted perceptions are further validated to assess their alignments with onsite perceptions at sampling locations. The results show that model-predicted perceptions do not exactly match onsite perceptions but align better when less stringent criteria are applied (within ± 1 scale point).

In short, this study underscores the necessity of gender inclusivity and a clear definition of safety terms when using SVIs to model perceptions. While SVI-based perception models are cost-effective, the predicted perceptions cannot yet fully substitute onsite perceptions, necessitating broader research to refine the effectiveness.