Effects of Geometric Design Features and Traditional Traffic Control Devices on Wrong-Way Driving Incident at Partial Cloverleaf Interchange Terminal: A Machine-Learning Approach
Type of DegreePhD Dissertation
Civil and Environmental Engineering
MetadataShow full item record
The partial cloverleaf (parclo) interchange is the second most popular interchange in the United States. Past studies revealed that the parclo interchange terminal is one of the most common entry points for wrong-way driving (WWD) crashes. However, to our best knowledge, few studies have focused on what interchange terminal design features caused those WWD entries at this type of interchange. This study aims to explore the effects of geometric design features and traditional traffic control devices on WWD incidents at parclo interchange terminals. In this study, a total of 75 parclo interchange terminals from 13 states were monitored in order to collect WWD incident data. Based on 5984 hours of traffic monitoring video, 410 cases of WWD incidents were captured at 28 locations. Each case of WWD incident was reviewed and analyzed to record the time of day, weather conditions, WWD distance, and reaction time to evaluate the effectiveness of traditional traffic control devices for deterring WWD. Furthermore, a total of 20 detailed design features, including geometric design, usage of traffic control devices, and AADT at each location with/without WWD incidents, were collected to analyze the cause of WWD. This study utilized three machine learning techniques, including multiple correspondence (MCA) analysis, eXtreme gradient boosting (XGboost), and least absolute shrinkage and selection operator with logistic (Lasso-logistic) regression to explore the effects of different design features on WWD incidents. The MCA analysis provided a general idea of what design combinations will cause/prevent WWD incidents at parclo interchange terminals. Based on the results, six types of parclo interchange terminals that will cause/prevent WWD incidents can be drawn. To further quantify the effects of each design feature on WWD incidents, the XGboost was applied to fit the dataset. The results of XGboost quantified the impact of each design feature on WWD incidents; it can also be utilized to predict the risk of WWD incidents by giving a specific parclo interchange terminal. In real practice, the local practitioners may need to evaluate dozens of parclo interchange terminals in order to find high-risk locations; collecting twenty design features and running a tree-based XGboost model for each location could be difficult to achieve. Therefore, lasso-logistic regression was applied in order to use fewer design features to conduct prediction and evaluation in real practice. The fitted lasso-logistic regression included eight design features instead of twenty, which made the network screening process easier and more efficient. The XGboost and lasso-logistic models were verified by using the subset of collected data. The average accuracy of the fitted XGboost model was 80%, and the average accuracy of the fitted Lasso-logistic regression model was 78%. Finally, the results produced by this study were used to propose design guidelines for deterring WWD. The fitted lasso-logistic model was used to develop a checklist for field inspection, for local practitioners to conduct network screening, and identify locations with the potential of recurring WWD incidents.