Decoding Criminogenic Visual Cues: Saliency Detection & Urban Imagery for Understanding Crime Perception Across Neighborhoods

• In Event: Communities and Place: Community Perceptions, Education, Violence, and a Public Safety Collaborative

Abstract

A number of crime and place theories have highlighted the role that visual cues in the environment play in shaping behavior among guardians, targets, and offenders. An emerging literature has begun to capitalize on computer vision AI and ‘big’ datasets of urban imagery to test these theories in new ways at unprecedented scales. To improve our understandings of how environmental design features drive crime and human perception of place, we apply interpretable machine learning techniques to analyze how crime-oriented vision models process visual information. Using Google Streetview images and crime data, we train image-based models to predict crime risk and extract pixel saliency information to determine which visual features drive prediction. Additionally, we conduct pixel saliency analysis on models trained using human assessments of environmental features (visual preference, environmental transparency) to provide new insights about how objective urban features are subjectively processed and interpreted to construct psychological meanings of place. By aggregating pixel salience measures to the neighborhood-level, we investigate whether the visual cues associated with crime and place perceptions vary with sociodemographic community features. As a whole, these analyses demonstrate how this novel methodological approach can be used to push image-based crime research forward.

Authors

• Riley Tucker, Pennsylvania State University

• Alfred Chao, University of Chicago