A semantic perception-topological reasoning framework for robust vehicle trajectory reconstruction in distributed acoustic sensing

Distributed Acoustic Sensing (DAS) enables continuous, full-range traffic perception by transforming optical fibers into large-scale sensing arrays. However, retrieving high-fidelity trajectories remains challenging. Trajectory signatures are frequently compromised by partial observability and fragmentation stemming from physical deployment constraints. While current methods commonly employ linear priors to bridge these gaps, such rigid assumptions often fall short in variable-speed scenarios where trajectories exhibit complex non-linear behaviors. Furthermore, methods limited to local views typically lack the global context necessary to distinguish individual instances across large-scale fractures. To address these challenges, a “Semantic Perception-Topological Reasoning” framework is proposed. Synergizing local geometric perception with global topological reasoning to enable fine-grained instance-level distinction, this decoupled approach robustly repairs discontinuities in both linear and curved trajectories via physical constraints, avoiding the dependency on rigid linearity. Adopting a divide-and-conquer strategy, a lightweight CNN is first employed for unbiased pixel-level geometric perception. Subsequently, a domain-knowledge-driven graph optimization module, grounded in vehicle motion heuristics such as velocity smoothness, is introduced to refine the topology.By integrating Two-Hop Neighborhood Kinematic Constraints (ensuring kinematic smoothness) and a hierarchical restoration strategy that couples a global Linear Assignment Problem (LAP) solver with local geometric recall, this study mitigates the issue of spurious connections and limited precision often observed in shallow GNN formulations applied to DAS scenes, enabling the logical repair of signal fractures. Furthermore, to rigorously quantify robustness, a parameter-controllable simulated topology benchmark is constructed. Extensive experiments on both real-world highway DAS data and this benchmark demonstrate the system’s efficacy. Specifically, the framework achieves an F1-Score of 89.13% while maintaining low topological error (ΔNCC: 1.465) and geometric distortion (MAE: 1.0882). This work provides a physics-guided solution for precise traffic monitoring, effectively validating the robustness of integrating explicit kinematic priors with data-driven perception.