A Machine Learning Approach for Highway Intersection Risk Caused by Harmful Lane-Changing Behaviors

Highway intersection-related crashes are suspected to be associated with harmful lane-changing behaviors. To better understand the relationship between them, this study applied an innovative machine learning approach to identify crash risk factors and find solutions to reduce the intersection-related crash frequency and severity caused by harmful lane-changing behaviors. First, a vehicles approach time (VAT) model was developed to define and classify different types of harmful lane-changing behaviors. Second, the real world driving video data was collected and preprocessed to identify the potential crash risk factors of harmful lane-changing behaviors. Finally, an advanced machine learning algorithm, Lasso-LARS, was applied to analyze the relation between intersection-related crash risk factors and lane-changing behaviors. There were no significant differences in the VAT values between the VAT model and the Lasso-LARS regression model. The result shows that both the two models are suitable for the analysis of risk factors of harmful lane-changing behaviors.