Development of a Rear-End Conflict Prediction Framework for Unsignalized Intersections along Intercity Highways Using Machine Learning

Abstract

Rear-end collisions at high-speed unsignalized intersections are one of the primary causes of crash fatalities. Time to collision (TTC) has become a popular metric for assessing traffic conflicts and proactively evaluating the risk of rear-end collisions. However, in developing countries such as India, identifying rear-end conflicts utilizing a single time-based indicator could be misleading because different vehicle classes have distinct dynamic characteristics and travel at varying speeds. Consequently, this study used three proximal indicators: TTC, deceleration rate (DR), and relative speed of leader-follower vehicles to identify rear-end conflicts along intersections’ approaches from two crash-prone unsignalized junctions located in the National Capital Region, India. Trajectory profiles of each leader-follower vehicle pair extracted using a semiautomated tool were used for estimating these indicators. Rear-end conflicts were further classified into “severe” and “nonsevere” groups based on K-means clustering and indicators’ threshold values applicable to mixed traffic conditions. The results showed that the highest percentage of severe conflict was observed at a four-legged intersection compared to a three-legged one. This study further developed a rear-end conflict prediction framework to predict rear-end conflicts based on their severity classification along high-speed unsignalized intersections’ approaches using seven machine learning (ML) algorithms. Among all, the random forest (RF) algorithm performed the best for both sites. The study outcomes suggested that despite the lack of reliable crash information, a combination of surrogate safety measures and machine learning algorithms can produce reliable models for predicting rear-end conflicts in developing countries such as India.