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.

中文翻译：

---

用于交叉动力学车辆轨迹拼接的物理信息神经网络


高精度、长覆盖的车辆轨迹数据有利于各种交通现象的调查。然而，由于传感限制，现有数据集经常包含断开的轨迹，这阻碍了对交通的透彻理解。为了解决这个问题，本文提出了一种基于物理信息神经网络 （PINN） 的方法来拼接断裂的轨迹。所提出的基于 PINN 的方法通过集成物理先验（包括车辆运动学和边界条件）来增强传统神经网络，旨在提供超越训练域和正则化的信息，从而提高交叉动力学场景的方法准确性和外推能力（例如，从低速训练数据推断以重建高速轨迹）。采用两个公开可用的车辆轨迹数据集 NGSIM 和 HighSIM 来验证所提出的基于 PINN 的方法，并设计了四个有偏差的训练场景来评估基于 PINN 的方法的外推能力。结果表明，与基准模型相比，基于 PINN 的方法在轨迹拼接精度和一致性方面表现出卓越的性能。使用我们提出的基于 PINN 的方法处理的数据集已在线公开提供，以支持流量研究社区。此外，这种基于 PINN 的方法可以应用于更广泛的场景，包括基于物理的先验。