Identifying high-risk roadways using crowdsourced speed variation features from big mobility data

Effectively identifying high-risk roadways is crucial for enhancing road safety; however, traditional crash-based safety assessments are reactive and may underestimate risks. Leveraging emerging big mobility data collected from connected vehicles (CV), this study presents a new approach to proactively identifying high-risk roadways using crowdsourced speed features. Ten speed variation features were generated from one month of CV data to depict speed dynamics on single carriageway ‘A’ roads in Oxfordshire, UK. To capture the temporal dynamics of speed variations, temporally disaggregated features were generated from peak hours, non-peak hours, and weekends. This study employed various statistical and machine learning-based regression models, including Classification and Regression Tree (CART), Random Forest, and eXtreme Gradient Boosting (XGBoost), to analyze the relationship between speed-related features and three safety performance measures based on five years of crash data. The SHapley Additive exPlanation (SHAP) approach was employed to interpret the modelling outputs. Results demonstrated that crowdsourced speed variation features, especially from different temporal windows, are valuable surrogate safety measures for quantifying crash risks. Features related to acceleration (acceleration noise) and speed variance (skewness index) significantly impacted the modelling results. These findings could help transport practitioners better understand how speed variations relate to crash risks and unlock the potential for conducting proactive safety assessments using crowdsourced driving behaviour data. • The relationships between different speed variations and crash risks are examined. • Temporal speed features from peak, off-peak, and weekend are generated. • CV-based speed variation features are effective SSMs for quantifying crash risks. • Acceleration noise and speed variance significantly impact crash risk.