A multivariate-based conflict prediction model for a Brazilian freeway.

Area of Science:

• Traffic Engineering
• Transportation Safety
• Data Science in Transportation

Background:

• Real-time collision risk prediction models are vital for dynamic traffic management systems aiming to enhance road safety.
• Proactive safety improvements necessitate accurate crash occurrence and collision risk prediction models.

Purpose of the Study:

• To present a multivariate framework for selecting variables for a traffic conflict prediction model on the Brazilian BR-290/RS freeway.
• To develop and evaluate a data-driven model for estimating traffic conflict occurrence using freeway surveillance data.

Main Methods:

• Application of Bhattacharyya Distance (BD) and Principal Component Analysis (PCA) for variable importance assessment.
• Utilizing Linear Discriminant Analysis (LDA) with selected variables to estimate conflict occurrence.
• Employing a matched control-case technique with traffic data from surveillance cameras.

Main Results:

• Key variables impacting conflict prediction include total flow, lane occupancy standard deviation differences, and speed's coefficient of variation.
• The developed LDA-PCA model identified typical Brazilian freeway traffic heterogeneity leading to aggressive maneuvers.
• The LDA-PCA model demonstrated superior performance over the LDA-BD model.

Conclusions:

• The LDA-PCA model offers a robust approach for traffic conflict prediction, achieving 76% classification accuracy and 87% sensitivity.
• Understanding traffic pattern heterogeneity is essential for mitigating aggressive driving behaviors and improving freeway safety.
• This framework provides valuable insights for developing effective traffic safety management strategies on Brazilian freeways.