UROP Fellowship: Biomedical and Life Sciences
Research Mentor(s): Brian Lin, PhD
Research Mentor Institution/Department: University of Michigan Transportation Research Institute
Presentation Date: Wednesday, August 4th
Session: Session 3 (5pm-6:20pm EDT)
Breakout Room: Room 1
“Although there have been many recent strides in automated vehicle features (e.g., lane-keeping assist or automatic emergency braking) and autonomous vehicles (e.g., Cruise), there is still a significant need for research on human drivers’ behavior and decision-making. One larger goal of this research is to increase users’ acceptance and comfort with these driving technologies by helping align the technology’s actions / warnings more closely with human behavior and expectations. This research project, in particular, applies data mining (including statistical and machine learning methods, primarily via MATLAB code) to naturalistic driving data in order to model drivers’ behavior in response to two specific scenarios: cut-ins and cut-outs of the lead vehicle on US freeways.
The PI’s previous research team had already extracted relevant events from larger datasets, applied statistical tests, and created predictive models. They ultimately defined a cut-in scenario as another vehicle changing lanes into the subject vehicle’s lane within 30 m in front of the subject vehicle. Similarly, a cut-out scenario was defined as a lead vehicle within 30 m in front of the subject vehicle changing lanes from the subject vehicle’s lane to an adjacent lane. The previous team selected events through a combination of automatic, coded restrictions and manual video verification, resulting in 799 cut-ins and 684 cut-outs. There are different variables associated with each event, such as speed, acceleration, range, and range rate between the lead and subject vehicles, transversal (lateral distance) between vehicles, and yaw rate. Support vector machine, decision tree, bagged decision tree, and logistic regression models were built to predict whether the average acceleration during the lane-changing process (lasting, on average, 2.39 seconds) would be either below -0.1 m/s2 for a cut-in, or above 0.2 m/s2 for a cut-out. Prediction made use of the values of significant variables at the time of the Mobileye’s detection of a new lead vehicle. The highest accuracy was 80.9% for cut-in and 72.4% for cut-out.
Despite these fairly high accuracies, there were still many additional questions and possible variables remaining to be investigated. Other variables that could be derived from the original variables were identified through literature review, including Deceleration Rate to Avoid a Crash (DRAC), Time Headway (THW), Time to Collision (TTC), the inverse of TTC, Enhanced Time to Collision (ETTC), and jerk (the derivative of acceleration). Additional possible methods identified through literature review include k-nearest neighbors (KNN), hierarchical cluster analysis (HCA), artificial neural network (ANN), XGBoost, and multilevel mixed-effects linear modeling. Our research expands on the prior research team’s work by acquiring additional visualizations and statistics of variables’ distributions and possible intercorrelations, testing new variables found in literature review for significance, exploring the use of new methods, and investigating other possible dependent variables (e.g., speed change, brake / accelerator pedal application). In modeling driver behavior, this project will contribute to work done on autonomous vehicles that respond to the other road users’ maneuvers in a human-centered manner. ”
Authors: Laura Bartz; Beverly Liu; Brian T. W. Lin
Research Method: Computer Programming