Predicting Street Crime Through Environmental Design and Human Perception

Link to the paper

This project demonstrates how to:

• Extract micro built environment (MBE) features from street view imagery using deep learning
• Quantify human perceptions of urban spaces (safety, beauty, depression) at scale
• Apply machine learning models to predict crime likelihood at street intersections
• Analyze the relationship between environmental design and criminal activity

Features

• Deep Learning Segmentation: 19 MBE types extracted using PSPNet trained on Cityscapes dataset
• Perception Modeling: 6 human perception variables (safety, beauty, liveliness, etc.) derived from Place Pulse 2.0
• Crime Classification: XGBoost model achieving 86% accuracy in predicting crime vs non-crime intersections
• City-wide Analysis: Comprehensive study of Toronto’s street intersections and crime patterns
• CPTED Integration: Findings aligned with Crime Prevention Through Environmental Design principles

Technical Stack

• Deep Learning

  • PSPNet (Pyramid Scene Parsing Network): Semantic segmentation of street environments
  • Cityscapes Dataset: 25,000 annotated urban street scenes for model training
  • GluonCV: Pre-trained computer vision models implementation

• Machine Learning

  • XGBoost: Primary classification model for crime prediction
  • Logistic Regression: Statistical analysis of feature associations
  • Feature Importance Analysis: Data impurity-based metrics for variable ranking

• Perception Analysis

  • Place Pulse 2.0 Dataset: 1.17M pairwise comparisons of street imagery
  • Strength-of-Schedule Algorithm: Converting pairwise comparisons to 0-1 perception scores
  • Support Vector Classifier: 79.44% accuracy in safety perception classification

• Geospatial Data

  • Toronto Police Open Data: 242,879 crime records (2014-2022)
  • Toronto Census Data 2016: Demographic control variables
  • Google Street View API: High-resolution street imagery collection

• Statistical Analysis

  • Variance Inflation Factor (VIF): Multicollinearity detection
  • Akaike Information Criterion (AIC): Model comparison and selection
  • Five-fold Cross-validation: Model validation and performance assessment

Project Story

Traditional crime analysis relies on aggregated census data that fails to capture fine-grained environmental details around crime locations. This study addresses the gap by:

• Micro-scale Analysis: Moving beyond census blocks to examine specific street intersection characteristics
• Perception Integration: Incorporating human perceptions beyond just safety (beauty, depression, liveliness)
• Environmental Design Focus: Testing Crime Prevention Through Environmental Design (CPTED) principles with data

The research collected street view imagery for crime and non-crime intersections across Toronto, extracting detailed environmental features and human perception scores to understand what makes certain locations more susceptible to criminal activity.

Key Findings

• Environmental Factors: Mobility-related elements (roads, vehicles, sidewalks) showed strong positive association with crime
• Vegetation Effect: Natural elements like vegetation significantly reduced crime likelihood
• Perception Insights: Beauty and depression perceptions were better predictors than traditional safety measures
• Model Performance: MBE variables explained crime events more effectively than perception variables alone
• Urban Planning Impact: Findings support targeted interventions in street design and environmental aesthetics

Results

• Best Classification Model: XGBoost with all variables
  • Accuracy: 86% in five-fold cross-validation
  • Top predictors: Traffic lights, buildings, vegetation, beautiful perception
• Statistical Significance: All perception variables except “lively” showed significant associations with crime
• Feature Importance: Beautiful perception ranked highest among perception variables
• Practical Applications: Results inform urban planning strategies for crime prevention

Implications

This research provides evidence-based insights for:

• Urban Planners: Balancing accessibility with safety in public space design
• Policymakers: Understanding how environmental aesthetics influence crime patterns
• Community Safety: Implementing targeted interventions based on environmental risk factors
• CPTED Practitioners: Validating environmental design principles with large-scale data analysis

Conclusion

The study demonstrates that street-level environmental characteristics and human perceptions are significantly associated with crime events. By integrating computer vision, machine learning, and urban theory, this framework offers a scalable approach to understanding and predicting crime patterns for safer, more sustainable urban development.

The methodology can be adapted to other cities with street view imagery coverage, providing a valuable tool for evidence-based urban planning and crime prevention strategies.