Random forest models for identifying motorway Rear-End Crash Risks using disaggregate data

This paper presents an approach to develop motorway Rear-End Crash Risk Identification Models (RECRIM) using disaggregate traffic data, meteorological data and crash database for a study site at a two-lane-per-direction section on Swiss (right-hand driving) motorway A1. Traffic data collected from inductive double loop detectors provide instant vehicle information such as speed, time headway, etc. We define traffic situations (TS) characterized by 22 variables representing traffic status for 5-minute intervals. Our goal is to develop models that can separate TS under non-crash conditions and TS under pre-crash conditions using Random Forest - an ensemble learning method. Non-crash TS were clustered into groups that we call traffic regimes (TR). Precrash TS are classified into TR so that a RECRIM for each TR is developed. Interpreting results of the models suggests that speed variance on the right lane and speed difference between two lanes are the two main causes of the rear-end crashes. The applicability of RECRIM in a real-time framework is also discussed.