“Eyes” on the Street: Using Computer Vision to Get the Gist of How Environmental Design Shapes Crime Across Neighborhoods

Purpose: Propose a new computational approach for measuring the visual features of places emphasized by the crime prevention through environmental design (CPTED) framework, aesthetic value and natural surveillance. Inspired by lessons from cognitive psychology, this approach replaces object-based measures of environments (e.g., how many trees are present) with measures of visual 'gists' capturing broad image assessments. Methods: 4,800 respondents were surveyed to provide human ratings of five CPTED-inspired gist metrics (preference, complexity, memorability, transparency, and enclosure) for 8,249 Chicago Google Street View images. Using residual neural networks, we trained a series of computer vision models that could predict human-rated gist scores on new images. After validating out-of-sample performance, these models were applied to estimate CPTED gist scores on a larger set of 187,048 Chicago street-view images. Multi-level variance decomposition analysis was used to probe the nested geographic structure of gist metrics. Elastic net regressions were used to evaluate whether gist features were correlated with crime and prosociality (as measured via voting participation rates). Results: AI-assigned gist labels correlated highly with human-ratings, suggesting very strong out-of-sample accuracy for all gist models. Variance decomposition results suggested CPTED gists substantially vary across census tracts even when accounting for micro-spatial visual differences. Regression results suggested that gist metrics tended to be strongly correlated with crime, out-performing object-based measures in most instances. In particular, visual preference demonstrated a strong negative association with crime frequency and a positive association with prosociality. Conclusions: CPTED-inspired measures of visual gist explain variation in crime frequency and prosociality above and beyond what can be explained solely by object-based measures. Future research should seek to understand how broad visual gists and the presence of specific objects interact to shape behavior and decision-making among guardians, offenders, and targets.