Comparing Machine Learning and Statistical Methods for Vehicle Safety on Horizontal Curves

This study aims to assess the probability of unsafe operations on horizontal curves resulting from speed variation, employing both statistical analysis and machine learning (ML) techniques. The statistical analysis was conducted using Minitab software to assess the probability of non-compliance through the Monte-Carlo simulation method. Additionally, the research applied three ML classification models—a novel optimized version of the Random Forest (RF) classifier, Naive Bayes (NB), and Extreme Gradient Boosting (XGBoost). Nine curves with radii ranging from 700m to 2000m were selected from two rural roads in Egypt for the study. The evaluation of non-compliance probability on each curve involved contrasting the supply (design speed, a fixed value) with the demand (actual speed, characterized by actual speed distributions). Findings revealed that using the 85th percentile speed in the analysis, the probability of non-compliance during off-peak hours exceeded 50% for all curves except two, where it reached 100%. This indicates that approximately 100% of vehicles engage in unsafe operations during off-peak hours on these specific curves. Accuracy results of the ML classifiers showed that the proposed RF classifier performed exceptionally well with a perfect score of 1.0, followed by XGB and NB classifiers for all curves. A comparative analysis between the results of statistical analysis and ML in estimating curve safety suggests that ML outperforms statistical analysis, demonstrating its potential as a more reliable tool for assessing road safety on horizontal curves.