What Makes a Place Safe? Assessing AI-Generated Safety Perception Scores Using Stockholm’s Street View Images

• In Event: WG-PLACE 6: Research on crime, place and time

Abstract

In this study, we begin by examining people’s perceptions of safety in different types of urban environments. We asked them to indicate which images of the city made them feel safe. These images are from Street Views of Stockholm, Sweden. Using people’s responses, we first create a human safety perception score based on ratings of images. Then, we train an AI model to learn human safety perception patterns, which can be leveraged to measure safety perceptions of citywide street-view imagery. Based on this, we created safety scores for each image over the whole city. Deep learning segments each pixel in the images into categories like trees, buildings and roads, helping researchers link specific objects to particular planning areas and, later, different levels of safety perceptions. We calculate the percentage of each object in the images to represent different types of urban environments in Stockholm’s six planning areas. These areas have unique architectural characteristics following specific planning guidelines, such as million-home programs and single-family homes in the Garden Cities. The study integrates AI-generated safety scores from street view images, image segmentation techniques and conventional and crowdsourced data using Geographical Information Systems (GIS) and regression models. After accounting for income, crime and other area characteristics, the models reveal that areas with lower safety scores primarily consist of areas with a relatively large percentage of roads in industrial and/or interstitial mixed residential areas. Conversely, higher safety scores are found in large but distinct combinations of buildings, vegetation and open sky, from detached single-family housing to inner city high-density built areas.

Authors

• Vania Ceccato, KTH Royal Institute of Technology

• Yuhao Kang, The University of Texas at Austin

• Jonatan Abraham, KTH Royal Institute of Technology

• Per Näsman, KTH Royal Institute of Technology

• Fábio Duarte, MIT - Massachusetts Institute of Technology

• Song Gao, University of Wisconsin-Madison, USA

• Lukas Ljungqvist, Stockholm municipality

• Fan Zhang, Peking University, Beijing

• Carlo Ratti, MIT - Massachusetts Institute of Technology