Predicting hazardous events in work zones using naturalistic driving data

In the United States, someone is injured in a work zone every 14 minutes. Work zone crashes are not only a problem for the drivers, they are a serious concern for highway workers who are injured or killed by errant vehicles. Each year, over 20,000 workers are injured in work zones. A number of factors contribute to work zone crashes, including inattention, speeding, and driving under influence, all driver behavior factors. Unfortunately, driver behavior in work zones is not well understood. The naturalistic driving study (NDS) data, collected by the second Strategic Highway Research Program (SHRP 2), offers a rare opportunity for a first-hand view of crashes and near-crashes that occur around work zones and to compare them to non-work zone events. Four machine learning algorithms: Random forest, Deep Neural Network, Multilayer Feedforward Neural Network, and t-Distributed Stochastic Neighbor Embedding (t-SNE), were applied to work zone events within NDS data. The Random forest algorithm performed the best in classifying NDS data into crashes, near-crashes, and baseline using pre-event variables. The prediction accuracy for work zone events was 97.7% for three classes: crash, near-crash, and baseline and 88.7% for two classes: crash and near-crash. These accuracies were significantly higher than a Naïve predictor's accuracies of 62.6% and 74.2%, respectively. The high accuracies of Random forest models show that these models can be used to predict the occurrence of a safety critical event by only using pre-event variables.