关键词: 智能交通, 路径推理, Transformer, 稀疏轨迹数据, 通行时间估计, 弱监督学习

Abstract: To address the challenges of trajectory jumps, matching ambiguities, and inaccurate travel time estimation caused by sparse and low-quality trajectory data, this paper proposes a generative path inference framework based on structure-behavior joint modeling. The framework achieves end-to-end inference through three core modules. Firstly, a variational autoencoder is employed to learn the latent style variables from raw trajectories, capturing individual preferences in route choice and travel speed. Second, a dual-head Transformer decoder is designed to simultaneously generate the complete path structures and segment-level travel times under the collaborative guidance of style variables and contextual attention mechanisms. Finally, a fixed-point theory is introduced to construct a path-time closed-loop mapping, ensuring physical consistency and structural stability of the outputs through residual constraints. The experimental results on the Porto and Chengdu datasets demonstrate that under sparse sampling intervals of 30 to 180 seconds, the proposed method significantly outperforms baseline models, such as Hidden Markov Model (HMM), Deep Map Match (DeepMM), and Learning based Map Match(L2MM) in path matching accuracy, achieving an average absolute improvement of over 10 percentage under extreme sparsity while maintaining inference latency below 0.11 seconds. Ablation studies reveal the indispensability of each component: style modeling is the central to behavioral consistency, fixed-point optimization ensures topological rationality, and the multi-attention mechanism critically determines the accuracy of travel time prediction.

Key words: intelligent transportation, path inference, Transformer, sparse trajectory data, travel time estimation, weakly supervised learning