Exploring Contributing Factors of Hazardous Events in Construction Zones Using Naturalistic Driving Study Data

Abstract

The naturalistic driving study (NDS) data offers a rare opportunity for a first-hand view of crashes and near-crashes (CNC) that occur around work zones. The objective of this study is twofold: 1) evaluate risk of CNC using pre-incident variables using logistical regression models and; 2) classify and predict CNC events using machine learning methods. The regression models found that driving behavior, duration of secondary task and traffic density exerted a high influence on risk of CNC in work zones. Odds of a safety critical event for different risk factors followed similar trend for work zones and non-work zones. Duration of secondary task and traffic density variables contributed to increased crash risk in work zones than non-work zones. For the second objective, four machine learning algorithms: Random forest (RF), Deep Neural Network, Multilayer Feedforward Neural Network, and t-Distributed Stochastic Neighbor Embedding (t-SNE), were applied to work zone events and non-work zone events within NDS data. The RF algorithm performed the best in classifying CNC events occurring in work zones. The prediction accuracy was 86.3% for three classes: crash, near-crash, and baseline and 91.2% for two classes: crash and near-crash. For non-work zone data, the Deep Neural Network model outperformed others in differentiating between crash and near-crash events.