Measurement of Safety in public spaces through Street Viev Imagery

Purpose and goal: ** Denna text är maskinöversatt **

The project aimed to test and evaluate a technique to evaluate and map place safety using Machine- and Deep Learning models applied to Street View Imagery. A important objective was also to assess the potential for scaling up the technique: what generalizability do we see in the technique? What other "street qualities" might be evaluated/mapped in a similar way?

The goal achievement is considered to be good. The project has provided a lot of knowledge about the method itself, as well as the challenges and resource requirements for implementing it. Expected results and effects: ** Denna text är maskinöversatt **

The project has provided valuable experiences regarding the method studied. We hope the effect will be that more municipalities will take notice of the technology as a very cost-effective way to measure, evaluate, or monitor qualities and/or conditions in the urban space. Due to limited resources in the project and methodological challenges, the method could not be evaluated on more place qualities than "place safety."

Approach and implementation: ** Denna text är maskinöversatt **

In the study, we let a Machine Learning modeltrained to predict place safety based on a large research dataset with street-level imagesto "walk around" Tomelilla and predict place safety based on street-level images downloaded from the Google Street View Static API. The model used a large number of image attributes from each imageextracted with two pre-trained deep learning modelsas explanatory variables.

The predictions and the effort to produce them were evaluated, and an interactive map (Tomelilla´s "safety map") was created.