Off-campus UNL users: To download campus access dissertations, please use the following link to log into our proxy server with your NU ID and password. When you are done browsing please remember to return to this page and log out.
Non-UNL users: Please talk to your librarian about requesting this dissertation through interlibrary loan.
Data-Centric Models for Critical Maneuver Prediction Using Naturalistic Driving Dataset
The last few years have seen a significant interest in driver behavior recognition. This is particularly true in the new era of advanced technologies, for instance connected and automated vehicles (CAVs) which are rapidly becoming a reality and have caught the attention of transportation researchers. During the transition period to 100 percent CAVs, CAVs and human-driven vehicles (HDVs) will have to co-exist and there will be various situations that may cause conflicts in the mixed traffic flow. It is important for CAVs to have the ability to identify and predict risky driving intention of non-CAV drivers in order to mitigate traffic collisions and enhance traffic safety. This dissertation aims at improving the current state-of-the-art (SOTA) methodology for predicting critical maneuvers. It uses a newly released naturalistic trajectory dataset recorded by unmanned aerial vehicles. In this dissertation, three sets of detection and prediction models are developed for forecasting multiclass lane changing, risky cutting-in, and critical brake events using long short-term memory (LSTM) deep neural networks. The focus of this dissertation is on the highway on/off ramp locations which generate highly dynamic interactions between adjacent vehicles which often lead to traffic conflicts. The detection and prediction models were trained for different values of prediction horizon time from 0.5 seconds to five seconds. It was found that the proposed approaches had an average accuracy of 97% with a small false alarm (<3%). Moreover, this dissertation adopted the Explainable Artificial Intelligence (XAI) concept with Shapley Additive Explanation (SHAP) feature importance method in order to address the lack of machine learning (ML) model interpretability. This approach also helped to explain the causality of critical maneuver behaviors. It is hypothesized that the methodology used in this dissertation can be used in combination with other comprehensive data fusion sources (e.g., weather conditions, highway conditions, driver socio-demographic, and physiological attributes) to improve the prediction accuracy and to lead to better safety of CAV and non-CAV in mixed traffic flow situations.
Pham, Huong, "Data-Centric Models for Critical Maneuver Prediction Using Naturalistic Driving Dataset" (2022). ETD collection for University of Nebraska - Lincoln. AAI29168483.