Research on high-fidelity reconstruction algorithm for vehicle trajectory based on centralized spatiotemporal feature fusion

Abstract

Vehicle trajectory data, primarily characterized by time-series features, serves as the key information carrier for vehicle movement, which documents dynamic characteristics such as positions, velocities, and accelerations across spatiotemporal dimensions. This data underpins critical tasks including microscopic traffic flow modeling, driving state analysis, and traffic safety assessment. To address the challenge that low-sampling-rate vehicle trajectory data induces the loss of high-frequency motion details in intelligent transportation systems and that traditional methods fail to simultaneously satisfy the requirements for both high precision and motion plausibility due to disjointed spatiotemporal modeling and lack of physical constraints, this study proposes a high-fidelity vehicle trajectory reconstruction algorithm based on centralized spatiotemporal feature fusion. The algorithm, based on the SOFTS (Series-cOre Fused Time Series) framework, constructs a global-local collaborative spatiotemporal feature representation mechanism through a multidimensional enhanced STAR (Spatio-Temporal Aggregation and Representation) module. Specifically, it incorporates a spatio-temporal embedding layer to capture interdependencies between timestamps and spatial coordinates. Residual connections are employed to preserve original trajectory details, while a spatial proximity weighting mechanism optimizes core feature aggregation. A dynamic weight matrix is constructed to adaptively focus on velocity-position correlations among neighboring vehicles within a 20-meter-radius range. Kinematic constraints are integrated to ensure physical plausibility of reconstructed trajectories, including the introduction of a kinematic loss function. This function leverages acceleration smoothness regularization terms and trajectory curvature continuity regularization to guide the model toward physically feasible solutions in compliance with vehicle dynamics. To validate its effectiveness, the proposed method was extensively validated using public datasets (NGSIM, HighD, and CQSkyeyeX) and systematically compared with traditional approaches (e.g., linear interpolation) and deep learning models. Experimental results clearly show that the improved algorithm significantly outperforms baseline methods in interpolation accuracy, spatiotemporal smoothness, and computational efficiency. The research findings can be applied to high-frequency trajectory generation for autonomous driving, microscopic traffic flow simulation, and other domains, providing critical technical support for upgrading intelligent transportation systems from data collection to decision-making optimization.