A Physics-Informed Neural Network for Vehicle Trajectory Reconstruction in Cut-In Scenarios with Sparse and Noisy Observations

Accurate trajectory data are fundamental to traffic modeling and autonomous vehicle development. However, reconstructing trajectories in cut-in scenarios is challenging due to complex multi-vehicle interactions and frequently sparse, noisy observations. Existing model-based methods require extensive parameter tuning, while purely data-driven methods depend on densely labeled trajectory datasets and may violate physical consistency. To address these limitations, this paper proposes CI-PINN (cut-in physics-informed neural network), a self-supervised framework for trajectory reconstruction under severe data degradation. By integrating a longitudinal interaction model that captures anticipation and relaxation behaviors, CI-PINN ensures kinematic plausibility by jointly minimizing data-fitting and physics residual losses. Experiments on the NGSIM dataset demonstrate robust performance across missing rates of 80–90%, achieving a mean absolute error of 0.91 m and a mean squared error of 2.17 m2, which are 63.2% and 78.1% lower than the best baseline method, respectively. These results demonstrate a label-efficient and physically consistent framework for trajectory reconstruction in cut-in scenarios. Beyond improving microscopic trajectory fidelity, the proposed method preserves system-level traffic metrics more reliably, facilitating more accurate safety assessments and intelligent transportation applications.