Computer Vision Helps Us Get the Gist of Crime Prevention Through Environmental Design Across Neighborhoods

• In Event: 30662 - Section on Crime, Law, and Deviance Roundtables
  In Refereed Roundtable (60 minute): Table 10: Crime Prevention/ Economics, Crime, Desistance

Abstract

The crime prevention through environmental design framework argues that visual features related to aesthetic value and natural surveillance (e.g., well kept yards and lawns) impact peoples willingness to engage in prosocial activities that prevent crime. Problematically, these visual features have historically been challenging to measure at city-scale. While scholars have been able to increasingly use big data and AI to measure the visual environments of cities by counting objects that are theorized to alter crime frequency or prosociality (e.g., vegetation, fences, buildings), literature in cognitive psychology suggests behavior may be rooted in broad assessments of visual 'gists' rather than object-based perceptions. To consider this possibility, we asked 4,800 respondents to provide human ratings of five CPTED-inspired gist metrics (preference, complexity, memorability, transparency, enclosure) for 8,249 Chicago street view images collected through the Google Street View API. Using residual neural networks, we trained a series of AI models that predict gist scores on new images with a high level of accuracy according to train-test analyses. We then used our models to estimate CPTED gist scores on a larger set of 187,048 Chicago street view images. Through the use of multi-level variance decomposition analyses, for all five gist metrics we observed a substantial degree of variation in visual environments across census tracts even when controlling for micro-spatial differences across street blocks, block groups, and images. Finally, we compared aggregate tract-level gist scores to crime data from Chicago Police Department, with analyses suggesting measures of visual gist are correlated with a wide variety of crime types. Future research should continue to develop theory to identify relevent visual gists and understand the underlying mechanisms linking neighborhood visuals to crime.

Authors

• Riley Tucker, Pennsylvania State University

• Alfred Chao, University of Chicago

• Nakwon Rim, University of Chicago

• Elizabeth Gaillard, University of Chicago

• Gaby Akcelik, University of Chicago

• Marc Berman, University of Chicago