Physics-informed neural network for cross-dynamics vehicle trajectory stitching

High-accuracy long-coverage vehicle trajectory data can benefit the investigations of various traffic phenomena. However, existing datasets frequently contain broken trajectories due to sensing limitations, which impedes a thorough understanding of traffic. To address this issue, this paper proposes a Physics-Informed Neural Network (PINN)-based method for stitching broken trajectories. The proposed PINN-based method enhances traditional neural networks by integrating physics priors, including vehicle kinematics and boundary conditions, aiming to provide information beyond training domain and regularization, thus increasing method accuracy and extrapolation ability for cross-dynamics scenarios (e.g., extrapolating from low-speed training data to reconstruct high-speed trajectories). Two publicly available vehicle trajectory datasets, NGSIM and HighSIM, were adopted to validate the proposed PINN-based method, and four biased training scenarios were designed to assess the PINN-based method's extrapolation ability. Results indicate that the PINN-based method demonstrated superior performance regarding trajectory stitching accuracy and consistency compared to benchmark models. The dataset processed using our proposed PINN-based method has been made publicly available online to support the traffic research community. Additionally, this PINN-based approach can be applied to a broader range of scenarios that include physics-based priors.