Zhan Zhou, Qunyong Wu, Shiyu Yang, Jiahuan Luo
High-precision and high-sampling-rate vehicle trajectory data play a crucial role in intelligent transportation analysis. However, a large portion of existing trajectory data is sparse and requires effective techniques for recovery. Existing methods typically adopt a single form of trajectory representation, overlooking the potential of heterogeneous representations to capture spatial semantic features. Meanwhile, grid-based discretization of trajectories often results in a loss of spatial details. To address these issues, we propose a road-network-constrained, multi-representation, data-driven trajectory recovery model (MReDTrajRec). We design a fine-grained trajectory representation method based on road spatial feature-derived weighted vectors, which reduces spatial information loss while maintaining geometric fidelity. By integrating grid-based road semantic graphs with grid-discretized trajectories, we develop a convolutional operation that dynamically extracts road features surrounding the trajectory, enabling multimodal data fusion and spatial information modeling. Extensive experiments on the Porto and Rome datasets demonstrate the superiority of MReDTrajRec.