Crime and Visually Perceived Safety of the Built Environment: A Deep Learning Approach

Although the influence of the built environment on both crime and people’s safety perceptions is well documented in the international literature, less evidence is found regarding the link between urban safety perceptions and crime occurrence. In this article, we investigate the potential relationship between crime and visual perceived safety (VPS), using Stockholm, Sweden as a case. Central to the study is the VPS score, a detailed measure of VPS and situational fear, created by combining a deep learning model with a data set of local street view images and citizen impressions. We examine this measure together with traditional crime records to compare the city’s distribution of safety and crime. First, geographical patterns and spatial clusters of high and low levels of crime and VPS were detected. Then, drawing from principles of environmental criminology, a spatial regression was used to examine the relationship between the VPS score and crime, controlling for sociodemographics and land-use factors. Findings show that crime rates of different types are significant predictors of poor VPS, but mismatching geographies of perceived safety and crime are common. The article discusses the findings and finishes by highlighting the impact of these results for research and practice.