A Combined Car-Following and Overtaking Approach to Reconstructing Vehicle Trajectories: A Case Study in Shenzhen

Abstract

Vehicle trajectories display several problems such as multi-source heterogeneity, space–time-limited fragmentation, and low-penetration-rate sparsity, failing to support traffic state estimation and traffic collaborative control. Fully sampled vehicle trajectory reconstruction has received increasing attention. However, existing approaches to trajectory reconstruction focus on simple scenarios with single lanes, short distances, and non-interference, ignoring driving dynamics such as overtaking, resulting in limited applicability. To improve the utilization of sparse trajectories, a novel trajectory reconstruction framework is tailored by incorporating car-following and overtaking behaviors. First, an improved method is proposed to recognize driving behavior patterns, and the optimal number of the cluster-based patterns are determined. Then, an improved piecewise cubic Hermite interpolating polynomial algorithm is designed to reconstruct pattern-recognized probe vehicle (PV) trajectories, by integrating local weighted regression. Second, a piecewise order-changing model is proposed to capture overtaking dynamics and arrival stochasticity. Third, by incorporating the K-dimensional tree algorithm, a novel algorithm is developed to draw a speed contour map, providing regional speed baselines for estimating unknown non-probe vehicle (NPV) trajectories. Finally, an integrated approach combining the car-following (CF) and inverse car-following (ICF) models is tailored to avoid the accumulation of errors and address overlapping problems in trajectories reconstructed by the CF model. A weighted fusion model is tailored to fuse the two NPV trajectory candidates estimated by the CF and ICF models. A model comparison proves the superiority of the combined approach over the CF model (41.39%–51.33%) and reveals that the higher the penetration rate of PVs, the better the reconstruction outcomes.