A Machine Learning Method to Decide Start Position of Mandatory Lane Changes Before Expressway Exits

Abstract: Mandatory lane change maneuvers are needed for vehicles preparing to exit an expressway section, especially for those autonomous vehicles travelling on the dedicated lane typically set on the inner side of the road. Apparently, the longitudinal position to start the whole planning and control process for lane changes is a key variable to be decided, influencing traffic safety, efficiency, and the success probability to exit finally. We hope that the lane changing will not start too early to expand the weaving area, nor too late to increase safety risks and the probability to even miss the exit.
Previous studies on lane change decision algorithms or warning signs fixed roadside for exits mostly only assume a start position value or simply analyze the effects of several discretized values. Recent there are some studies on optimization methods of the start position in continuous spaces but focus mainly on urban intersections or two-lane highways using precise mathematical models. As for multi-lane expressways' exits, there is still a lack of data-based and efficient methods which can quickly adapt to different flow states. Besides, few studies analyze different requirements of exit lane changes for different autonomous vehicle types, behavior models and even platoons.
This paper presents a method to determine the starting position for lane changes prior to a vehicle exiting an expressway, specifically addressing one of the challenges faced by autonomous vehicles transitioning from the innermost dedicated lane to the exit ramp. To achieve this, we employ traffic simulation with SUMO for data collection and utilize machine learning techniques to develop probabilistic models which represent the distribution of end positions for lane changes and assess the success probability of entering exit ramps under various road conditions and traffic states.
Specifically, we first collect simulation data by running multiple trials under predefined start positions and traffic conditions, recording both end positions on the outermost lane and corresponding bool outcomes of success. Second, we establish classification boundaries to differentiate successful and unsuccessful trial samples, enabling the estimation of success probability using Probability Density Estimation (PDE). Third, by repeating this process across different values of start positions and traffic conditions, a comprehensive model of success probability assessments using Gaussian Process Regression (GPR). The start position as the input of the model can then be decided in reverse based on the requirements of exiting success probability as the output of the model.
Within the vehicle-infrastructure integration system, this method can be adaptive to diverse traffic conditions, providing fast real-time guidance for connected vehicles which inclues automated cars with ACC driving characteristics, human-driven coaches with IDM driving characteristics and truck platoons with CACC driving characteristics. We find the start positions for them vary a lot from each other. Compared with traditional, experience-based approaches, our approach enhances safety and improves traffic flow efficiency upstream of expressway exits for these three vehicle types.