Optimizing Urban Perception: Integrating Image Segmentation and Machine Learning to Explore Street ViewCrime Correlations in London Lewen Zhang1, Robin Song1, Wenshuo Zhang1
1 University College London, London, United Kingdom
Abstract. Urban safety is a critical concern for town planners, architects, and the general public, particularly in metropolitan areas like London. This study investigates the relationship between crime data in London and street view images (SVI) using advanced image segmentation regression models. By analyzing crime incidents alongside visual data, the research aims to provide accurate predictions of crime rates and identify key contributing factors. The project employs machine learning techniques, precisely Random Forest and XGBoost, to visualize actual versus predicted crime distributions, enhancing our understanding of urban safety dynamics. Additionally, the study explores innovative solutions for optimizing street scenes through reinforcement learning, allowing for adaptive urban planning strategies. An interactive online platform is developed, enabling users to modify street view components easily. This platform leverages Stable Diffusion to generate responsive urban scenarios, fostering community engagement and informed decision-making. By integrating geospatial analysis with visual data, this research contributes to the field of urban safety and offers practical tools for enhancing the livability and security of urban environments. Ultimately, the findings aim to inform policy decisions and improve the overall quality of life in urban settings. `Keywords: Machine Learning, Street View Images, Crime Rate
1
Background
London is often perceived on social media as a high-crime city, with urban crime data reaching 75,000 incidents in August 2024 alone, exhibiting significant geographical spatial clustering characteristics, particularly with crime rates around Oxford Street being substantially higher than in other urban areas. Existing research based on Space Syntax theory has already confirmed that urban spatial configuration has a complex structural impact on crime occurrence, with urban space integration and connectivity demonstrating a non-linear, multidimensional correlation with crime rates. This study focuses on the proportional distribution of different segmentation elements in street scenes (such as buildings, streets, greenery, etc.), and through quantitative analysis of