Observed the Unobservable: Lane Change Estimation with Following Vehicles’ Trajectories only

In intelligent transportation systems, where there is not yet 100% penetration of connected vehicles, Traffic State Estimation (TSE) is crucial for all traffic participants, including both vehicles and infrastructure. The reconstruction of occluded vehicle trajectories is vital for TSE. The presence of potential lane change (LC) behaviors in modern transportation systems adds complexity to trajectory reconstruction. LCs are common, as drivers change lanes in anticipation of potential benefits for speed and mobility gain. Estimating the start and end times for the LC process enables the possibility of reconstructing trajectories for lane-changing vehicles (LCVs), yet occlusion is a hurdle for such process. To cope with issue of LCV being occluded, in this work, we propose a method for estimating lane change times based solely on the following vehicle’s trajectory profile. We find that local optima obtained through the trajectory profile may not accurately predict lane change times, despite being the local optima that best differentiate the new segments of subtrajectories. By incorporating additional trajectory-level features and some unique points within the trajectory profile, we can enhance the accuracy of our predictions. We also investigated the effect of different feature selection methods, and we showed that with proper feature selection method, prediction accuracy can improve comparing to baseline.